MAGNETIC MULTILAYER NANOPILLARS FOR THE …...torque and damping can be afiected by the addition of...
Transcript of MAGNETIC MULTILAYER NANOPILLARS FOR THE …...torque and damping can be afiected by the addition of...
MAGNETIC MULTILAYER NANOPILLARS FOR THE
STUDY OF CURRENT-INDUCED REVERSAL OF A
THIN MAGNETIC LAYER
A Dissertation
Presented to the Faculty of the Graduate School
of Cornell University
in Partial Fulfillment of the Requirements for the Degree of
Doctor of Philosophy
by
Nathan C. Emley
December 2005
c© 2005 Nathan C. Emley
ALL RIGHTS RESERVED
MAGNETIC MULTILAYER NANOPILLARS FOR THE STUDY OF
CURRENT-INDUCED REVERSAL OF A THIN MAGNETIC LAYER
Nathan C. Emley, Ph.D.
Cornell University 2005
This dissertation presents a detailed summary of the nanofabrication of, and mea-
surements performed on, magnetic multilayer nanopillars for the study of current-
induced reversal of a thin ferromagnetic (FM) layer. These samples consist of a
FM(thick)/Cu/FM(thin) trilayer that has been patterned to lateral dimensions
on the order of 100 nm into pillar-shaped structures through an intricate fabri-
cation process. Current is confined to flow perpendicularly to the plane of the
layers becoming spin-polarized by the natural spin-filtering effects of the FM lay-
ers. When impingent upon the FM(thin) layer, the spin polarization can excite a
dynamic response from the magnetic moment through an exchange of spin angular
momentum, known as a spin torque.
Spin torque-induced reversal of the FM(thin) moment depends upon the strength
of the spin torque per unit current as well as the magnitude of the dissipation of
magnetic energy or damping. I present two studies of spin torque-induced rever-
sal in magnetic multilayer nanopillars demonstrating that the strength of the spin
torque and damping can be affected by the addition of magnetic material in prox-
imity to the FM(thin) layer. In the first study, I add a third, FM(bottom) layer
nearly adjacent to and aligned anti-parallel with the FM(thick) layer. The net
strength of the spin torque acting on the FM(thin) layer is reduced due to oppos-
ing spin-filtering effects from the anti-parallel-aligned FM(bottom) and FM(thick)
layers.
In the second study, I perform time-resolved measurements of spin torque-
induced reversal in FM(thick)/Cu/FM(thin) trilayer nanopillars at bath temper-
atures 4.2 K to 160 K. Comparison of the data to detailed simulations of the
switching yields numerical values for the spin torque and damping as functions of
temperature. The damping is strongly temperature-dependent, which is attributed
to the pinning behavior of a thin antiferromagnetic oxide layer around the perime-
ter of the FM(thin) layer that forms naturally during sample fabrication. This
adventitious pinning layer can have a major impact on spin torque phenomena.
BIOGRAPHICAL SKETCH
Born in December, 1975 to parents Warren and Sheri Emley, Nathan was the
youngest of three children. He grew up in rural NH and at a very early age showed
a natural inclination towards athletics and proved to be a student of slightly-above-
average scholastic ability. He attended the Derryfield School in Manchester, NH
for high school and mustered admission into the Honors Program at the University
of Massachusetts, Amherst. Although his initial academic focus was in Astronomy,
an uninspiring summer internship at an Astrophysical think tank caused him to
change his major to Physics. Nathan received his B.S. summa cum laude in Physics
from UMass in 1998, whereupon he stayed on for one year as a graduate researcher.
He then decided to pursue a graduate career at Cornell University in Applied
Physics, where he received his M.S. in 2003 and his Ph.D. in 2005. After Cornell,
Nathan will be a post-doctoral researcher at UC Berkeley, where he will further
his interest in studying nanostructures for potential technological impact.
For as long as he can remember, Nathan has been incredibly fortunate to have
been inspired at auspicious times by the right role models — teachers, friends, and
family members — something for which he is eternally grateful.
iii
To my family:
Warren,
Sheri,
Shelley,
and
Tyler.
iv
ACKNOWLEDGEMENTS
I wish to acknowledge my advisor Bob Buhrman who taught me so much about
doing quality science. I thank him for not giving up on me when he had plenty
of opportunities to do so. His hands-off management style allowed me to explore
scientific or engineering problems to whatever end I sought. It also forced me
to rely on, and subsequently develop, my own strengths and skills. It was an
educational experience that has left me with a stronger foundation. His success in
getting grant money is what made much of this work possible.
I also wish to acknowledge our post-doc Ilya Krivorotov whose profound, en-
cyclopedic knowledge of Physics, coupled with his kind, gentle, and unassuming
demeanor, made him such a wonderful reference. This thesis would definitely not
be what it is without his help.
During my time in the Buhrman group, I was fortunate to have met so many
gifted individuals. I would like to acknowledge my good friend Andrew Perrella,
whose leadership and inexhaustible positive energy are true gifts and whose friend-
ship is a rare treasure. Ozhan Ozatay, a great friend and gifted scientist, has shared
many dinners and coffees with me discussing the mysteries of the universe.
Other students in the Robert Buhrman and Dan Ralph groups have contributed
to the successful completion of my graduate career each in their own unique way,
be it by providing friendly chats or by helping with sickly vacuum systems: Phil
Mather, Pat Braganca, John Read, Andrei Garcia, Eric Ryan, Jack Sankey, Sergey
Kiselev, Kiran Thadani, Juting Zhai, Greg Fuchs, Vlad Pribiag, and Preeti Chal-
sani.
I would also like to acknowledge Andre Mkhoyan for his diligent effort in
preparing and imaging specimens in the scanning transmission electron micro-
v
scope (STEM). Likewise, I must also thank Jordan Katine and his research staff at
Hitachi Global Storage Technologies for providing 3-terminal nanopillar samples.
I very much wish to acknowledge my family, to whom I dedicate this thesis,
for they, above all others, have shown me unconditional support: My father War-
ren Emley, mother Sheri Poftak, sister Shelley Jones, and brother Tyler Emley.
Others deserving acknowledgement for their friendship and support both before
and during these difficult graduate years are Ben Pagnini, Aimee Crump, Ben
Russell, Chris DeCicco, Marcel Blais, J. Robert Lennon, Theo Wong, Mike San-
tarlas, Stacey Emley, Ben Jones, Matt Poftak, Julia Emley, Mark Tuominen, Don
Candela, Robert Krotkov, Dudley Cotton, and Dennis Holland.
Of course, the long hours were not possible without periodic caffeine “refuels”.
Thanks goes to Stella’s, Collegetown Bagels, Libe Cafe, and Gimme for the quality
coffee.
vi
TABLE OF CONTENTS
1 Introduction 1References for Chapter 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2 Spin transfer-based magnetoelectronics: potential future tech-nologies 62.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62.2 Spin transfer introduction . . . . . . . . . . . . . . . . . . . . . . . 7
2.2.1 Spin transfer by spin filtering . . . . . . . . . . . . . . . . . 102.2.2 Spin transfer by averaging . . . . . . . . . . . . . . . . . . . 122.2.3 Spin transfer in an asymmetric FM/NM/FM trilayer . . . . 12
2.3 Magnetization excitations induced by a spin torque . . . . . . . . . 162.3.1 Spin torque in the Landau-Lifshitz-Gilbert equation . . . . . 162.3.2 Magnetization reversal . . . . . . . . . . . . . . . . . . . . . 202.3.3 Persistent dynamics . . . . . . . . . . . . . . . . . . . . . . . 25
2.4 Potential future spin transfer-based technologies . . . . . . . . . . . 282.4.1 Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 282.4.2 Microwave oscillators . . . . . . . . . . . . . . . . . . . . . . 312.4.3 Global storage . . . . . . . . . . . . . . . . . . . . . . . . . . 322.4.4 Programmable logic . . . . . . . . . . . . . . . . . . . . . . . 35
References for Chapter 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3 Nanopillar fabrication at Cornell: present and future 433.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 433.2 C mask nanopillar fabrication process . . . . . . . . . . . . . . . . . 44
3.2.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . 443.2.2 Guidelines as of late 2005 . . . . . . . . . . . . . . . . . . . 70
3.3 The ion beam deposition system . . . . . . . . . . . . . . . . . . . . 963.3.1 Deposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . 973.3.2 Ion milling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
3.4 Aligned electron-beam lithography . . . . . . . . . . . . . . . . . . 1023.4.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1023.4.2 Alignment marks . . . . . . . . . . . . . . . . . . . . . . . . 1033.4.3 Global alignment . . . . . . . . . . . . . . . . . . . . . . . . 1093.4.4 Fine alignment . . . . . . . . . . . . . . . . . . . . . . . . . 1153.4.5 Successful demonstration of fine e-beam alignment . . . . . . 127
3.5 Chemical mechanical polishing . . . . . . . . . . . . . . . . . . . . . 1433.5.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1433.5.2 Adhesion layer . . . . . . . . . . . . . . . . . . . . . . . . . 1453.5.3 Polishing dielectrics . . . . . . . . . . . . . . . . . . . . . . . 1453.5.4 Effects of changing CMP process parameters . . . . . . . . . 1483.5.5 Comparison with ion mill planarization . . . . . . . . . . . . 150
vii
3.6 Radio frequency (RF) backsputtering . . . . . . . . . . . . . . . . . 1583.7 Hydrogen silsesquioxane (HSQ) . . . . . . . . . . . . . . . . . . . . 160
3.7.1 History . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1603.7.2 HSQ as an e-beam resist . . . . . . . . . . . . . . . . . . . . 1623.7.3 Handling and processing . . . . . . . . . . . . . . . . . . . . 164
3.8 Examining ion milled nanostructures with STEM . . . . . . . . . . 1653.8.1 Fabrication of HSQ lines . . . . . . . . . . . . . . . . . . . . 1673.8.2 STEM images of patterned lines . . . . . . . . . . . . . . . . 1673.8.3 Shadowing effects: a simple model . . . . . . . . . . . . . . . 173
3.9 Proposed future HSQ-based nanopillar fabrication . . . . . . . . . . 1923.9.1 Purpose . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1923.9.2 Fabrication process . . . . . . . . . . . . . . . . . . . . . . . 1943.9.3 HSQ nanopillar images . . . . . . . . . . . . . . . . . . . . . 2143.9.4 Unknowns . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
References for Chapter 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
4 Synthetic antiferromagnetic layers in magnetic nanopillars 2244.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2244.2 Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2254.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2284.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232References for Chapter 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
5 Performing time-resolved spin torque-driven switching measure-ments at variable T 2355.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2355.2 Hardware setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
5.2.1 Construction of the high frequency dipper . . . . . . . . . . 2365.2.2 Custom high frequency holders . . . . . . . . . . . . . . . . 2435.2.3 Procedure for Au ribbon bonding . . . . . . . . . . . . . . . 2585.2.4 Connecting to the dipper . . . . . . . . . . . . . . . . . . . . 2695.2.5 High frequency setup . . . . . . . . . . . . . . . . . . . . . . 2735.2.6 Voltage calibration at the sample input . . . . . . . . . . . . 2745.2.7 Sample alignment . . . . . . . . . . . . . . . . . . . . . . . . 276
5.3 Performing time-resolved pulsed current measurements . . . . . . . 2765.3.1 Preparation . . . . . . . . . . . . . . . . . . . . . . . . . . . 2765.3.2 Software control . . . . . . . . . . . . . . . . . . . . . . . . . 2785.3.3 Data acquisition . . . . . . . . . . . . . . . . . . . . . . . . . 2815.3.4 Data analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 284
References for Chapter 5 . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
viii
6 Time-resolved spin torque switching and enhanced damping inPy/Cu/Py nanopillars 2946.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2946.2 Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2956.3 Macrospin simulation . . . . . . . . . . . . . . . . . . . . . . . . . . 2986.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3096.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3146.6 Exchange biasing in other samples . . . . . . . . . . . . . . . . . . . 3146.7 Further details of simulation . . . . . . . . . . . . . . . . . . . . . . 321
6.7.1 Stochastic distribution of initial angle . . . . . . . . . . . . . 3216.7.2 Simulation parameters A, B, and α0 . . . . . . . . . . . . . 3216.7.3 Simulation constants . . . . . . . . . . . . . . . . . . . . . . 3276.7.4 Simulating a real pulse shape . . . . . . . . . . . . . . . . . 336
References for Chapter 6 . . . . . . . . . . . . . . . . . . . . . . . . . . . 338
7 Novel device structure: 3-terminal nanopillar 3407.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3407.2 Proposed fabrication process . . . . . . . . . . . . . . . . . . . . . . 3447.3 Measured devices from Hitachi Global Storage Technologies . . . . 386
7.3.1 Fabrication differences . . . . . . . . . . . . . . . . . . . . . 3867.3.2 3-terminal devices . . . . . . . . . . . . . . . . . . . . . . . . 3867.3.3 Experimental results . . . . . . . . . . . . . . . . . . . . . . 3907.3.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . 400
References for Chapter 7 . . . . . . . . . . . . . . . . . . . . . . . . . . . 405
8 Conclusion 406References for Chapter 8 . . . . . . . . . . . . . . . . . . . . . . . . . . . 409
ix
LIST OF TABLES
3.1 Sputter wafer: basic guidelines. . . . . . . . . . . . . . . . . . . . . 713.2 C evaporation: wafer pre-clean. . . . . . . . . . . . . . . . . . . . . 723.3 C evaporation: parameters. . . . . . . . . . . . . . . . . . . . . . . 723.4 PMMA bilayer spin-coating. . . . . . . . . . . . . . . . . . . . . . . 753.5 Develop PMMA. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 763.6 Cr lift off. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 773.7 Etch C in O2 plasma (RIEx). . . . . . . . . . . . . . . . . . . . . . 783.8 Photolithography: resists used for different photolithography steps. 793.9 Photolithography: spin recipes. . . . . . . . . . . . . . . . . . . . . 803.10 Photolithography: exposure, development, and descum numbers. . 813.11 Ion milling: Veeco and IBD. . . . . . . . . . . . . . . . . . . . . . . 843.12 IBD ion milling: beam parameters. . . . . . . . . . . . . . . . . . . 843.13 IBD ion milling: gas parameters. . . . . . . . . . . . . . . . . . . . 853.14 Veeco ion milling: beam parameters. . . . . . . . . . . . . . . . . . 873.15 Veeco ion milling: PECVD oxide etch numbers. . . . . . . . . . . . 883.16 Strip resist. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 923.17 IBD deposition: beam parameters. . . . . . . . . . . . . . . . . . . 943.18 IBD deposition: gas parameters. . . . . . . . . . . . . . . . . . . . 953.19 Dielectric deposition thickness uniformity. . . . . . . . . . . . . . . 1463.20 Dielectric CMP etch rate and uniformity: “Oxide” recipe. . . . . . 1473.21 PECVD nitride CMP etch rate and uniformity with different recipes.1473.22 RF backsputter parameters. . . . . . . . . . . . . . . . . . . . . . . 1603.23 HSQ process parameters. . . . . . . . . . . . . . . . . . . . . . . . 1663.24 Critical point drying. . . . . . . . . . . . . . . . . . . . . . . . . . 1663.25 Benefits of the HSQ process. . . . . . . . . . . . . . . . . . . . . . 1943.26 HSQ e-beam resist parameters. . . . . . . . . . . . . . . . . . . . . 194
5.1 High frequency dipper: parts. . . . . . . . . . . . . . . . . . . . . . 2385.2 High frequency holder: parts. . . . . . . . . . . . . . . . . . . . . . 2565.3 Front panel input for Time Machine 2.8.vi. . . . . . . . . . . . . . 280
6.1 Effective spin polarization as a function of Hext. . . . . . . . . . . 307
x
LIST OF FIGURES
2.1 Two-channel model of GMR. . . . . . . . . . . . . . . . . . . . . . 82.2 Spin transfer illustration. . . . . . . . . . . . . . . . . . . . . . . . 92.3 Spin transfer by spin filtering. . . . . . . . . . . . . . . . . . . . . . 112.4 Spin transfer by averaging. . . . . . . . . . . . . . . . . . . . . . . 132.5 Asymmetric FM/NM/FM trilayers: AP to P spin torque. . . . . . 142.6 Asymmetric FM/NM/FM trilayers: P to AP spin torque. . . . . . 152.7 Torques on a magnetic moment. . . . . . . . . . . . . . . . . . . . 182.8 Spin torque dynamics. . . . . . . . . . . . . . . . . . . . . . . . . . 192.9 Spin torque acting on an elliptical thin-film nanomagnet. . . . . . . 212.10 Spin torque switching: H⊥ = 0, α = 0. . . . . . . . . . . . . . . . . 232.11 Spin torque switching: H⊥ = 0, α 6= 0. . . . . . . . . . . . . . . . . 242.12 Spin torque switching: H⊥ 6= 0, α 6= 0. . . . . . . . . . . . . . . . . 262.13 Spin torque persistent dynamics. . . . . . . . . . . . . . . . . . . . 272.14 MRAM architecture. . . . . . . . . . . . . . . . . . . . . . . . . . . 352.15 Spin torque-actuated recording (STAR). . . . . . . . . . . . . . . . 36
3.1 C mask process: sputter deposition of metallic layers. . . . . . . . . 453.2 C mask process: evaporate C. . . . . . . . . . . . . . . . . . . . . . 453.3 C mask process: spin PMMA bilayer. . . . . . . . . . . . . . . . . 463.4 C mask process: e-beam lithography to define nanohole in PMMA. 473.5 C mask process: develop and evaporate Cr. . . . . . . . . . . . . . 483.6 C mask process: liftoff Cr. . . . . . . . . . . . . . . . . . . . . . . . 493.7 C mask process: oxygen plasma C. . . . . . . . . . . . . . . . . . . 493.8 C mask process: photolith 1 (PL(1)): “Define Leads”. . . . . . . . 503.9 C mask process: ion mill 1 (IM(1)): “Isolate Devices”. . . . . . . . 513.10 C mask process: photolith 2 (PL(2)): “Define Pillar”. . . . . . . . 523.11 C mask process: ion mill 2 (IM(2)): “Define Pillar” (1/2). . . . . . 533.12 C mask process: ion mill 2 (IM(2)): “Define Pillar” (2/2). . . . . . 543.13 C mask process: evaporate 50 A of SiOx. . . . . . . . . . . . . . . 543.14 C mask process: PECVD oxide to insulate nanopillar. . . . . . . . 553.15 C mask process: ion mill 3 (IM(3)): “Planarize SiO2” (1/2). . . . . 563.16 C mask process: ion mill 3 (IM(3)): “Planarize SiO2” (1/2). . . . . 573.17 C mask process: photolith 3 (PL(3)): “HF Etch”. . . . . . . . . . . 583.18 C mask process: profilometry to measure oxide height above pillar. 593.19 C mask process: photolith 4 (PL(4)): “Protect Shorts”. . . . . . . 603.20 C mask process: evaporate oxide. . . . . . . . . . . . . . . . . . . . 613.21 C mask process: photolith 5 (PL(5)): “Oxide Window”. . . . . . . 623.22 C mask process: ion mill 4 (IM(4)): “Thin Oxide Above Pillar”. . . 633.23 C mask process: photolith 6 (PL(6)): “Top Leads”. . . . . . . . . . 643.24 C mask process: ion mill 5 (IM(5)): “Open Pillar”. . . . . . . . . . 653.25 C mask process: oxygen plasma to remove C cap. . . . . . . . . . . 663.26 C mask process: ion mill 6 (IM(6)): “Clean Top Contact”. . . . . . 67
xi
3.27 C mask process: deposit top leads in situ. . . . . . . . . . . . . . . 683.28 C mask process: completed device. . . . . . . . . . . . . . . . . . . 693.29 RBS spectra on IBD-deposited AlOx film. . . . . . . . . . . . . . . 993.30 AJA and IBD deposited film resistance versus temperature. . . . . 1003.31 IBD ion milling: uniformity. . . . . . . . . . . . . . . . . . . . . . . 1013.32 Lift off profile with standard resist: fencing. . . . . . . . . . . . . . 1053.33 Lift off profile with undercut resist: no fencing. . . . . . . . . . . . 1063.34 Pattern transfer with interlayer (1/2). . . . . . . . . . . . . . . . . 1073.35 Pattern transfer with interlayer (2/2). . . . . . . . . . . . . . . . . 1083.36 Finding circle1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1103.37 Record mark coordinates for e-beam alignment. . . . . . . . . . . . 1113.38 Alignment marks for multiple aligned exposures in positive tone
resist. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1173.39 Shared origin of the different sets of alignment marks. . . . . . . . 1183.40 Exposure pattern for positive resist multiple alignments. . . . . . . 1193.41 Alignment marks for multiple aligned exposures in negative tone
resist. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1213.42 Exposure pattern for negative resist multiple alignments. . . . . . . 1223.43 Fine alignment search dimensions. . . . . . . . . . . . . . . . . . . 1253.44 SEM images of VB6 fine alignment. . . . . . . . . . . . . . . . . . 1293.45 Fine alignment demonstration: define holes for alignment marks. . 1313.46 Fine alignment demonstration: multilayer liftoff. . . . . . . . . . . 1323.47 Fine alignment demonstration: deposit C. . . . . . . . . . . . . . . 1333.48 Fine alignment demonstration: clear C from Au mark holes. . . . . 1343.49 Fine alignment demonstration: define Au mark base. . . . . . . . . 1353.50 Fine alignment demonstration: first level e-beam exposure (1/2). . 1363.51 Fine alignment demonstration: first level e-beam exposure (2/2). . 1373.52 Fine alignment demonstration: deposit Cr, etch C. . . . . . . . . . 1383.53 Fine alignment demonstration: define alignment marks. . . . . . . 1393.54 Fine alignment demonstration: define nanopillar. . . . . . . . . . . 1403.55 Fine alignment demonstration: e-beam alignment to nanopillar. . . 1413.56 Fine alignment demonstration: aligned nanopillar. . . . . . . . . . 1423.57 CMP schematic. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1443.58 CMP run to run variation. . . . . . . . . . . . . . . . . . . . . . . 1493.59 CMP rotation factor variation. . . . . . . . . . . . . . . . . . . . . 1513.60 CMP chuck pressure variation. . . . . . . . . . . . . . . . . . . . . 1523.61 AFM of PECVD oxide, as deposited. . . . . . . . . . . . . . . . . . 1543.62 AFM of PECVD oxide, after ion mill planarization. . . . . . . . . . 1553.63 AFM of PECVD oxide after CMP-smoothing oxide. . . . . . . . . 1563.64 RMS roughness of a CMP-smoothed PECVD oxide layer after ion
milling. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1573.65 HSQ lines: deposit layers. . . . . . . . . . . . . . . . . . . . . . . . 1683.66 HSQ lines: spin e-beam resist bilayer. . . . . . . . . . . . . . . . . 1693.67 HSQ lines: e-beam expose & develop. . . . . . . . . . . . . . . . . 170
xii
3.68 Top down schematic of HSQ lines. . . . . . . . . . . . . . . . . . . 1713.69 HSQ lines: SEM after development (1/3). . . . . . . . . . . . . . . 1723.70 HSQ lines: SEM after development (2/3). . . . . . . . . . . . . . . 1733.71 HSQ lines: SEM after development (3/3). . . . . . . . . . . . . . . 1743.72 HSQ lines: AFM after development. . . . . . . . . . . . . . . . . . 1743.73 HSQ lines: define line. . . . . . . . . . . . . . . . . . . . . . . . . . 1753.74 HSQ lines: SEM after ion mill (1/3). . . . . . . . . . . . . . . . . . 1763.75 HSQ lines: SEM after ion mill (2/3). . . . . . . . . . . . . . . . . . 1773.76 HSQ lines: SEM after ion mill (3/3). . . . . . . . . . . . . . . . . . 1783.77 HSQ lines: AFM after ion mill. . . . . . . . . . . . . . . . . . . . . 1783.78 HSQ lines: AlOx deposition. . . . . . . . . . . . . . . . . . . . . . 1793.79 HSQ lines: CMP planarization & photolithography. . . . . . . . . . 1803.80 HSQ lines: photolith to define top leads. . . . . . . . . . . . . . . . 1813.81 HSQ lines: deposit top leads. . . . . . . . . . . . . . . . . . . . . . 1823.82 Cross-sectional STEM image of a line. . . . . . . . . . . . . . . . . 1833.83 STEM images of lines: layers identified. . . . . . . . . . . . . . . . 1843.84 Ion mill model: line schematic. . . . . . . . . . . . . . . . . . . . . 1863.85 Ion mill model: linear mill mask. . . . . . . . . . . . . . . . . . . . 1873.86 Ion mill model: linear mask shadowing as a function of position. . . 1883.87 Ion mill model: cylinder schematic. . . . . . . . . . . . . . . . . . . 1893.88 Ion mill model: cylindrical mill mask. . . . . . . . . . . . . . . . . 1903.89 Ion mill model: cylindrical mask shadowing as a function of position.1913.90 HSQ process, step 1: Sputter multilayer stack. . . . . . . . . . . . 1953.91 Step 2: spin e-beam resist bilayer. . . . . . . . . . . . . . . . . . . 1963.92 Step 3: e-beam exposure and develop. . . . . . . . . . . . . . . . . 1973.93 Step 4: define nanopillar. . . . . . . . . . . . . . . . . . . . . . . . 1983.94 Step 5: insulate nanopillar. . . . . . . . . . . . . . . . . . . . . . . 1993.95 Step 6: photolith 1 – “Isolate Devices” (1/2). . . . . . . . . . . . . 2003.96 Step 6: photolith 1 – “Isolate Devices” (2/2). . . . . . . . . . . . . 2013.97 Step 7: HF etch AlOx. . . . . . . . . . . . . . . . . . . . . . . . . . 2023.98 Step 8: isolate devices. . . . . . . . . . . . . . . . . . . . . . . . . . 2033.99 Step 9: photolith 2 – “AlOx Refill” (1/2). . . . . . . . . . . . . . . 2043.100 Step 9: photolith 2 – “AlOx Refill” (2/2). . . . . . . . . . . . . . . 2053.101 Step 10: AlOx refill. . . . . . . . . . . . . . . . . . . . . . . . . . . 2063.102 Step 11: CMP planarization & HSQ mask removal. . . . . . . . . . 2073.103 Step 12: photolith 3 – “Define Bonding Pads” (1/2). . . . . . . . . 2073.104 Step 12: photolith 3 – “Define Bonding Pads” (2/2). . . . . . . . . 2083.105 Step 13: clear bonding pads. . . . . . . . . . . . . . . . . . . . . . 2093.106 Step 13: clear bonding pads – HF corrosion. . . . . . . . . . . . . . 2093.107 Step 14: deposit bonding pads. . . . . . . . . . . . . . . . . . . . . 2103.108 Step 15: photolith 4 – ”Define Top Leads” (1/2). . . . . . . . . . . 2103.109 Step 15: photolith 4 – ”Define Top Leads” (2/2). . . . . . . . . . . 2113.110 Step 16: clean off Ta cap. . . . . . . . . . . . . . . . . . . . . . . . 2123.111 Step 17: deposit top leads. . . . . . . . . . . . . . . . . . . . . . . 213
xiii
3.112 FOx-12 nanopillars: different sizes. . . . . . . . . . . . . . . . . . . 2153.113 FOx-12 nanopillars: variation in size. . . . . . . . . . . . . . . . . . 2163.114 FOx-16 nanopillars: circular pillars. . . . . . . . . . . . . . . . . . 2183.115 FOx-16 nanopillars: elliptical pillars. . . . . . . . . . . . . . . . . . 219
4.1 GMR of SAF nanopillar. . . . . . . . . . . . . . . . . . . . . . . . . 2274.2 Spin transfer in SAF nanopillar. . . . . . . . . . . . . . . . . . . . 2294.3 Field-current phase diagram SAF nanopillars. . . . . . . . . . . . . 231
5.1 HF dipper construction: dipper schematic. . . . . . . . . . . . . . . 2375.2 HF dipper construction: adaptor. . . . . . . . . . . . . . . . . . . . 2395.3 HF dipper construction: teflon spacer. . . . . . . . . . . . . . . . . 2405.4 HF dipper construction: 180 bend in input line. . . . . . . . . . . 2415.5 HF dipper construction: sample holder (1/4). . . . . . . . . . . . . 2455.6 HF dipper construction: sample holder (2/4). . . . . . . . . . . . . 2465.7 HF dipper construction: sample holder (3/4). . . . . . . . . . . . . 2475.8 HF dipper construction: sample holder (4/4). . . . . . . . . . . . . 2485.9 HF dipper construction: sample cover (1/3). . . . . . . . . . . . . . 2495.10 HF dipper construction: sample cover (2/3). . . . . . . . . . . . . . 2505.11 HF dipper construction: sample cover (3/3). . . . . . . . . . . . . . 2515.12 HF dipper construction: CPW (1/2). . . . . . . . . . . . . . . . . . 2525.13 HF dipper construction: CPW (2/2). . . . . . . . . . . . . . . . . . 2535.14 HF dipper construction: chain of connections. . . . . . . . . . . . . 2555.15 Pulse distortion from probe contacts. . . . . . . . . . . . . . . . . . 2605.16 Ribbon bonding: bonder tool. . . . . . . . . . . . . . . . . . . . . . 2615.17 Ribbon bonding: Al sonicating head. . . . . . . . . . . . . . . . . . 2635.18 Ribbon bonding: Au sonicating head. . . . . . . . . . . . . . . . . 2645.19 Bonded sample: in holder. . . . . . . . . . . . . . . . . . . . . . . . 2705.20 Bonded sample: bonds between CPW and sample. . . . . . . . . . 2715.21 Bonded sample: bonds on leads. . . . . . . . . . . . . . . . . . . . 2725.22 High frequency setup schematic. . . . . . . . . . . . . . . . . . . . 2755.23 Calibration of the voltage pulse. . . . . . . . . . . . . . . . . . . . 2775.24 Pulsed current acquisition sequence. . . . . . . . . . . . . . . . . . 2835.25 Example of pulsed current data. . . . . . . . . . . . . . . . . . . . 2855.26 Directory structure necessary to execute the analysis software. . . . 2875.27 Example of time-shifting data. . . . . . . . . . . . . . . . . . . . . 288
6.1 Spin torque switching of a nanomagnet: dV/dI versus ItextrmDC . . . 2966.2 Spin torque switching of a nanomagnet: time-resolved reversal. . . 2976.3 Switching speed versus I for AP to P switching. . . . . . . . . . . . 3026.4 Switching speed versus I for P to AP switching. . . . . . . . . . . . 3036.5 Critical currents normalized to M2
s (T ). . . . . . . . . . . . . . . . . 3046.6 Critical currents normalized to spin torque parameters. . . . . . . . 3056.7 Best fit macrospin simulation parameters A, B, and α0. . . . . . . 307
xiv
6.8 Effects of changing simulation constant Hext. . . . . . . . . . . . . 3086.9 Direct evidence of exchange biasing of the free layer. . . . . . . . . 3126.10 Progression of random exchange biasing of the free layer. . . . . . . 313
6.11 Exchange biasing after a second cool down, in Hcoolapp = -3.2 kOe. . 315
6.12 Minor loop progression after a second cool down, in Hcoolapp = -3.2
kOe. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3166.13 Exchange biasing after cooling in Hcool
app = +3.2 kOe. . . . . . . . . 318
6.14 Minor loop progression after cooling in Hcoolapp = +3.2 kOe. . . . . 319
6.15 A sample with no stochastic field switching. . . . . . . . . . . . . . 3206.16 Averaging over many simulated switches. . . . . . . . . . . . . . . 3226.17 Effects of changing simulation parameter A. . . . . . . . . . . . . . 3236.18 Effects of changing simulation parameter B. . . . . . . . . . . . . . 3246.19 Effects of changing simulation parameter α0. . . . . . . . . . . . . 3256.20 Asymmetric spin torque function g(θ). . . . . . . . . . . . . . . . . 3266.21 Effects of changing simulation constant E0. . . . . . . . . . . . . . 3286.22 Effects of changing simulation constant HK. . . . . . . . . . . . . . 3296.23 Effects of changing simulation constant θmis (1/3). . . . . . . . . . 3306.24 Effects of changing simulation constant θmis (2/3). . . . . . . . . . 3316.25 Effects of changing simulation constant θmis (3/3). . . . . . . . . . 3326.26 Effects of changing simulation constant area. . . . . . . . . . . . . 3336.27 Effects of turning off Ohmic heating. . . . . . . . . . . . . . . . . . 3346.28 Effects of turning off the angular dependence of damping. . . . . . 3356.29 Effects of simulating with a real pulse shape. . . . . . . . . . . . . 337
7.1 3-terminal nanopillar: basic geometry. . . . . . . . . . . . . . . . . 3437.2 Step 1: photolith 1, define global alignment marks (1/6). . . . . . . 3457.3 Step 1: photolith 1, define global alignment marks (2/6). . . . . . . 3467.4 Step 1: photolith 1, define global alignment marks (3/6). . . . . . . 3477.5 Step 1: photolith 1, define global alignment marks (4/6). . . . . . . 3487.6 Step 1: photolith 1, define global alignment marks (5/6). . . . . . . 3497.7 Step 1: photolith 1, define global alignment marks (6/6). . . . . . . 3497.8 Step 2: multilayer deposition. . . . . . . . . . . . . . . . . . . . . . 3507.9 Step 3: spin e-beam resist. . . . . . . . . . . . . . . . . . . . . . . . 3517.10 Step 4: e-beam expose alignment marks. . . . . . . . . . . . . . . . 3517.11 Step 5: define fine alignment marks. . . . . . . . . . . . . . . . . . 3527.12 Top view. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3527.13 Step 6: spin e-beam resist. . . . . . . . . . . . . . . . . . . . . . . . 3537.14 Step 7: e-beam define bottom pillar mask. . . . . . . . . . . . . . . 3547.15 Bottom exposure e-beam pattern. . . . . . . . . . . . . . . . . . . . 3547.16 Step 8: oxygen plasma PMMA underlayer. . . . . . . . . . . . . . . 3557.17 Side view of alignment marks. . . . . . . . . . . . . . . . . . . . . . 3557.18 Top view. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3567.19 Step 9: ion mill define bottom pillar. . . . . . . . . . . . . . . . . . 356
xv
7.20 Step 10: insulate nanopillar with AlOx. . . . . . . . . . . . . . . . 3577.21 Step 11: photolith 2, isolate devices (1/2). . . . . . . . . . . . . . . 3577.22 Step 11: photolith 2, isolate devices (2/2). . . . . . . . . . . . . . . 3587.23 Step 12: HF etch AlOx. . . . . . . . . . . . . . . . . . . . . . . . . 3597.24 Step 13: ion mill to isolate devices. . . . . . . . . . . . . . . . . . . 3597.25 Step 14: measure AlOx height. . . . . . . . . . . . . . . . . . . . . 3607.26 Step 15: photolith 3, AlOx refill (1/2). . . . . . . . . . . . . . . . . 3607.27 Step 15: photolith 3, AlOx refill (2/2). . . . . . . . . . . . . . . . . 3617.28 Step 16: AlOx refill. . . . . . . . . . . . . . . . . . . . . . . . . . . 3627.29 Step 17: CMP planarization. . . . . . . . . . . . . . . . . . . . . . 3627.30 Step 18: photolith 4, sputter top layers (1/2). . . . . . . . . . . . . 3627.31 Step 18: photolith 4, sputter top layers (2/2). . . . . . . . . . . . . 3637.32 Step 19: clean off Ta cap. . . . . . . . . . . . . . . . . . . . . . . . 3647.33 Step 20: deposit top layers. . . . . . . . . . . . . . . . . . . . . . . 3647.34 Step 21: photolith 5, open alignment marks (1/2). . . . . . . . . . 3657.35 Step 21: photolith 5, open alignment marks (2/2). . . . . . . . . . 3667.36 Step 22: open alignment marks. . . . . . . . . . . . . . . . . . . . . 3677.37 Step 23: spin e-beam resist. . . . . . . . . . . . . . . . . . . . . . . 3677.38 Step 24: e-beam define top pillar mask. . . . . . . . . . . . . . . . 3687.39 Top exposure e-beam pattern. . . . . . . . . . . . . . . . . . . . . . 3697.40 Step 25: oxygen plasma PMMA underlayer. . . . . . . . . . . . . . 3707.41 Top view. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3707.42 Step 26: ion mill define top pillar. . . . . . . . . . . . . . . . . . . 3717.43 Step 27: insulate nanopillar with AlOx. . . . . . . . . . . . . . . . 3717.44 Step 28: measure AlOx height. . . . . . . . . . . . . . . . . . . . . 3727.45 Step 29: photolith 6, second AlOx refill (1/2). . . . . . . . . . . . . 3727.46 Step 29: photolith 6, second AlOx refill (2/2). . . . . . . . . . . . . 3737.47 Step 30: second AlOx refill. . . . . . . . . . . . . . . . . . . . . . . 3747.48 Step 31: second CMP planarization. . . . . . . . . . . . . . . . . . 3747.49 Step 32: photolith 7, open leads (1/2). . . . . . . . . . . . . . . . . 3747.50 Step 32: photolith 7, open leads (2/2). . . . . . . . . . . . . . . . . 3757.51 Step 33: HF etch leads. . . . . . . . . . . . . . . . . . . . . . . . . 3767.52 Step 34: make Au contacts. . . . . . . . . . . . . . . . . . . . . . . 3767.53 Top view. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3777.54 Step 35: photolith 8, define middle lead (1/2). . . . . . . . . . . . . 3777.55 Step 35: photolith 8, define middle lead (2/2). . . . . . . . . . . . . 3787.56 Step 36: deposit Cu middle lead. . . . . . . . . . . . . . . . . . . . 3797.57 Step 37: photolith 9, protect shorts (1/2). . . . . . . . . . . . . . . 3807.58 Step 37: photolith 9, protect shorts (2/2). . . . . . . . . . . . . . . 3817.59 Step 38: protect shorts. . . . . . . . . . . . . . . . . . . . . . . . . 3827.60 Step 39: photolith 10, define top lead. . . . . . . . . . . . . . . . . 3837.61 Step 40: clean off Ta cap. . . . . . . . . . . . . . . . . . . . . . . . 3847.62 Step 41: deposit Cu top lead. . . . . . . . . . . . . . . . . . . . . . 3857.63 Shorting issues (1/3). . . . . . . . . . . . . . . . . . . . . . . . . . 387
xvi
7.64 Shorting issues (2/3). . . . . . . . . . . . . . . . . . . . . . . . . . 3887.65 Shorting issues (3/3). . . . . . . . . . . . . . . . . . . . . . . . . . 3897.66 Schematic of Hitachi 3-terminal nanopillar. . . . . . . . . . . . . . 3917.67 SEM images of 3-terminal nanopillar. . . . . . . . . . . . . . . . . 3927.68 SEM image of a completed 3-terminal device. . . . . . . . . . . . . 3937.69 3-terminal metallic nanopillar: top-to-middle-lead GMR. . . . . . . 3967.70 3-terminal metallic nanopillar: top and bottom pillar GMR. . . . . 3977.71 3-terminal MTJ nanopillar: pinned SAF bottom pillar resistance
versus area. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3997.72 3-terminal MTJ nanopillar: GMR. . . . . . . . . . . . . . . . . . . 4017.73 3-terminal MTJ nanopillar: uncompensated voltage offset. . . . . . 4027.74 3-terminal MTJ nanopillar: compensated voltage offset. . . . . . . 403
xvii
CHAPTER 1
INTRODUCTION
Spin-based electronics is a burgeoning field. Also known as spintronics, this is
an area of study focusing on devices whose functionality is based primarily on the
spin degree of freedom of the carriers. This is in contrast to conventional electron-
ics, which exploits only the charge of the carriers. Using either the spin in tan-
dem with the charge or alone, spintronics has some advantages over conventional,
charge-based semiconductor electronics, including higher integration density, non-
volatility, decreased power dissipation, and faster processing speeds [1].
The incorporation of magnets is a natural path for spin-based electronics to
follow since ferromagnetic materials, by nature of the Zeeman splitting of the
majority and minority energy states, have a net spin imbalance at the Fermi level.
This spin imbalance is useful for creating spin polarized currents [2, 3]. One such
use is to impart some of the spin angular momentum of this current into another
magnet. This exchange of spin angular momentum from the spin polarized current
to the magnetic moments can result in a mutual torque which can consequently
incite a dynamical response from the moment. This phenomenon is known as the
spin transfer effect [4]. The magnetically-based device which exhibits the spin
transfer effect most clearly is the magnetic trilayer, which consists of two magnetic
layers separated by a non-magnetic spacer. One layer is made deliberately thinner
than the other and, due to its over smaller moment, is labelled as the “free” layer
because it will be excited by the spin transfer torque at lower current than the
thicker, “fixed” layer.
In Chapter 2, I introduce many of the central ideas of the spin transfer effect,
including some of the specific mechanisms as to how the spin torque is applied
1
2
to the magnetic moment. The dynamical response of the magnetic moment from
application of the spin torque can be quite intricate and so to illustrate some
of these magnetization trajectories, I present the spin torque in the context of
the Landau-Lifshitz-Gilbert equation, the preeminent equation for magnetization
dynamics. I also provide my own outlook for potential future spin transfer-based
technologies, discussing from a business perspective how spin transfer-actuated
magnetoelectronic devices may be incorporated into existing mainstream or niche
markets, and list a few possible technologies where spin transfer could effectively
be exploited.
To date, devices exhibiting the most compelling and quantitatively significant
spin transfer data are magnetic multilayer nanopillars [5], where the multilayer
films are lithographically patterned to nanoscale dimensions, and nanoconstriction
point contacts to a continuous multilayer film [6]. These devices are sandwiched
between two electrodes for electronic measurement. Nano-scale dimensions of the
conduction channel are necessary to prevent self-field effects from dominating over
spin transfer effects [7].
This dissertation deals exclusively with the patterned nanopillar, and Chapter
3 summarizes the fabrication of magnetic multilayer nanopillars at Cornell Uni-
versity, expanding on the process first outlined in a previous thesis [7]. I have also
included an up-to-date and complete set of guidelines documenting much of the
process lore involved with fabricating nanopillars. I also discuss some new process-
ing methods that may be potentially useful for future fabrication processes.
In Chapter 4, I describe an experiment successfully demonstrating that all of
the layers within the nanopillar can affect the magnitude of the spin transfer effect.
Additional magnetic layers have been included in the normal trilayer composition
3
and, due to these new layers being within the spin-relaxation length of the free
magnetic layer, become active contributors to the overall spin transfer effect.
Chapter 5 describes the technical details of performing time-resolved measure-
ments of spin transfer-induced switching of the free layer at variable temperature.
The merger of high-speed electronics and cryogenic hardware is non-trivial and I
give present information in a very practical format, including detailed instructions
on how to construct specific and integral parts of the experimental setup, how to
use the controlling software already written to run the experiments, and how to
understand the data analysis code.
Building upon the technical details of the experimental setup described in
Chapter 5, Chapter 6 describes an experiment measuring time-resolved magne-
tization reversal dynamics of spin transfer excitations as a function of temper-
ature. The experimental results are compared with simulated switching events
using the macrospin approximation, which assumes a completely uniform magne-
tization. The results indicate that an antiferromagnetic oxide, formed naturally
on the free layer sidewalls, can have a major impact on spin transfer phenomena,
implying that finer control of nanopillar sidewalls is necessary in order to optimize
spin transfer behavior.
In Chapter 7, I introduce the 3-terminal nanopillar, a novel spin transfer device
geometry where a third electrode contacts the central, non-magnetic spacer of the
normal trilayer. Realization of such a device would be very important from a tech-
nological perspective and I present some initial results of such devices fabricated at
Hitachi Global Storage Technologies. Although such a device was not successfully
fabricated at Cornell, I describe a process a future student may follow to construct
such a device.
4
Finally, I conclude this dissertation with a recapitulation of the most significant
conclusions of my experiments that were summarized in Chapters 4 and 6.
5
References for Chapter 1
[1] Wolf S.A., Awschalom D.D., Buhrman R.A., Daughton J.M., von Molnar S.,Roukes M.L., Chtchelkanova A.Y., & Treger D.M., Spintronics: A spin-basedelectronics vision for the future, Science 294, 1488 (2001).
[2] Meservey R. & Tedrow P.M., Spin-polarized electron tunneling, Phys. Rep.238, 173 (1994).
[3] Stiles M.D. & Zangwill A., Noncollinear spin transfer in Co/Cu/Co multilayers,J. Appl. Phys. 91, 6812–6817 (2002).
[4] Slonczewski J.C., Current-driven excitation of magnetic multilayers, J. Magn.Magn. Mater. 159, L1 (1996).
[5] Katine J.A., Albert F.J., Buhrman R.A., Myers E.B., & Ralph D.C., Current-driven magnetization reversal and spin-wave excitations in Co/Cu/Co pillars,Phys. Rev. Lett. 84, 3149 (2000).
[6] Rippard W.H., Pufall M.R., Kaka S., Russek S.E., & Silva T.J., Direct-currentinduced dynamics in Co90Fe10/Ni80Fe20 point contacts, Phys. Rev. Lett. 92,027201 (2004).
[7] Albert F.J., The Fabrication and Measurement of Current Perpendicular tothe Plane Magnetic Nanostructures for the Study of the Spin Transfer Effect,Ph.d., Cornell University (2003).
CHAPTER 2
SPIN TRANSFER-BASED MAGNETOELECTRONICS:
POTENTIAL FUTURE TECHNOLOGIES
2.1 Background
The rich and complex properties of magnetic materials have stimulated a vibrant
field of research for the better part of the past century. Recently, experiments
involving multilayer stacks of n non-magnet (NM)/ferromagnet (FM) metallic bi-
layers have shown that the spin component of an electron can play a significant
role in the properties of electron transport. In both current-in-plane (CIP) [1, 2]
and current-perpendicular-to-plane (CPP) [3, 4] geometries, electrical resistance
of the multilayer RML was found to change as the angle of orientation between
the magnetic layers was swept from alternately anti-aligned (↑↓) to all aligned
(↑↑) by application of an in-plane magnetic field H. The change in resistance
∆RML ≡ RML(↑↓) − RML(↑↑) is known as giant magnetoresistance (GMR), and
was attributed to spin-dependent conduction through the magnetic multilayer
stack.
Initial models of CPP-GMR [5] were based on what is known as the two-
channel assumption for spin conduction. This assumption states that the net cur-
rent flowing through a (NM/FM)n multilayer is made up of two distinct electron
spin-populations, majority (+) and minority (–), that flow parallel to one another
but do not share their constituent electrons via some spin-flip scattering process.
Consequently, conduction is divided into two separate channels that can be de-
scribed by two parallel chains of resistors (one +, the other –), shown pictorially
in Figure 2.1. Using classical Boltzmann drift diffusion, initial models derived the
6
7
overall resistance of the multilayer for two orientations of the magnetic moments
of the FM layers, (↑↑) or (↑↓) (Figures 2.1(a) and (b), respectively), where spin-
dependent resistances were calculated independently for both channels. The (↑↑)orientation represents minimum (maximum) scattering for the +(–) spin chan-
nel and a smaller RML than for the (↑↓) orientation, which scatters equally for
both channels. The transition between (↑↓) and (↑↑) orientation (which yields the
∆RML) can be achieved by application of an external magnetic field H in tandem
with an H = 0 (↑↓) orientation from indirect exchange coupling through a proper
choice of NM material and thickness [6].
The spin-current plays the role of the “indicator” in the GMR effect, where
the orientation of the magnetic layers is detectable only because there is a spin-
dependence to the electron scattering. Any mechanism which preferentially selects
spin angular momenta, such as the spin-dependent scattering of the different +
and – channels must, by definition, be imparting a torque on the applied current.
This important fact was described in the seminal paper by Slonczewski [7] where
he points out that the torque applied to the spin-polarized electron current by
the magnetization, like any torque, is mutual. He then went on to predict that,
if the current was large enough, the mutual torque between the electrons and
magnetization could actually push the magnetization. Such a circumstance, where
the spin-current plays the additional role of the “actuator”, is the inverse of the
GMR effect and is known as spin transfer.
2.2 Spin transfer introduction
This spin transfer effect comes from the mutual torque between the spin-polarization
of the current and the magnetic moment. In the Slonczewski picture, it is specif-
8
(a)
(b)
multilayer
NM FM
resistor network
majority (+)
majority (+)
minority (-)
minority (-)
Figure 2.1: The two-channel (majority (+) and minority (–)) parallel resis-tor networks are drawn above the multilayer, where the grey rectangles are thebulk, spin-dependent resistances of the FM (blue) layers, and the orange rectan-gles are the spin-independent resistances of the NM (orange) layers. Interfacial(spin-dependent) resistances are not shown for simplicity. Relative sizes of rectan-gles are proportional to the amount of scattering present in that layer for a givenchannel. Electrons scatter more for oppositely aligned magnetizations, as illus-trated by their trajectories (solid black lines) trough the multilayer stack. (a) (↑↓)alignment of the multilayer, where both (+) and (–) electrons see equal resistances,RML(↑↓). (b) (↑↑) alignment of the multilayer, where (+) electrons are shunted bylower resistances while (–) electrons see higher resistances RML(↑↑) < RML(↑↓).
9
yx
z
ϕ
ρ FM
electron flow
m = m z
torque
Figure 2.2: Illustration of the spin torque. Electrons (red arrows) propagatetowards a magnetic layer with some finite angle ρ to the magnetization direction~m (= mx in this cartoon). As the electrons flow into the magnet, there is amutual torque between the magnetization (yellow arrow) and the spin polarization,represented here by the green arrows. The spin torque acts to make the spinpolarization and the magnetization parallel. If the initial current was unpolarized,the magnetization would polarize the electrons, but experience no net torque sincetorques from individual electrons are randomized with respect to one other andsum to zero. For the incoming electrons, the polar angle is ρ and the azimuthalangle is ϕ. The latter defines the transverse component of the spin polarization.
ically the absorption of the transverse, or perpendicular component of the spin
polarization by the magnetization that gives rise to this mutual torque [7]. As
electrons impinge upon the NM/FM interface at some finite angle to the magne-
tization, an exchange interaction occurs which acts to make the spin polarization
axis and the magnetization parallel [8]. This interaction is known as the spin
torque and is illustrated in Figure 2.2.
Only for a spin-polarized current will there be a spin torque, since an un-
polarized current will apply spin torques in random directions, over the entire
distribution of electrons, that sum to zero. Let us assume that a spin polarizing
source is close to the magnetization in question for the following treatment. It is
10
sufficient to have another magnetic layer in proximity to spin polarize the current
(close enough so that the spins do not relax). The spin torque is predominantly a
band structure effect that is generated mostly by two different parts.
2.2.1 Spin transfer by spin filtering
Spin transfer by spin filtering is due to interfacial band structure where, due to
Zeeman splitting in the ferromagnet, there is an unequal alignment between the
NM conduction band and the two spin sub-bands (majority or minority) in the FM.
Calculations of interfacial band structure between different NM/FM combinations
have demonstrated that transmission across the interface is spin dependent and, for
Cu/Ni and Cu/Co, favors majority electrons (i.e. electrons with spin polarization
parallel to the magnetization [9]. This preferential reflection of electrons of a given
spin population is know as spin filtering and has been studied by Andreev reflection
in point contacts to continuous magnetic films [10]. For an electron current with
arbitrary spin polarization (ρ, ϕ) defined relative to a coordinate system like that
in Figure 2.2, it is always possible to define the wave function of the electrons in
terms of the basis vectors of the magnetization (| ↑〉, | ↓〉),
|ψ〉 = [a(ρ, ϕ)| ↑〉+ b(ρ, ϕ)| ↓〉]ei~k·~r (2.1)
where ρ and ϕ are the azimuthal and polar angles, respectively (as shown in Fig-
ure 2.2), a(ρ, ϕ) and b(ρ, ϕ) are direction-dependent probability amplitudes, ~k is
the wave vector (which satisfies hk2/2m = EF , where EF is the Fermi energy), and
~r = (x, y, z) is the spatial variable. The probability of transmission or reflection
from the NM/FM interface is dependent upon a(ρ, ϕ), b(ρ, ϕ), and ~k [11]. The re-
11
yx
zFM
Reflection
Transmission
Figure 2.3: Spin transfer by spin filtering. Electrons propagate towards the FMlayer from the left hand side at a finite angle to the magnetization. Preferential spinscattering at the interface transmits spin up electrons while reflecting spin downelectrons. Even though the incident electrons come in at an arbitrary angle, theirwave function |ψ〉 can be defined in terms of the basis vectors of the magnetization| ↑〉 and | ↓〉.
flected and transmitted electrons are spin polarized collinear to the magnetization,
meaning all transverse spin angular momentum has been absorbed, or transferred
to the magnet. Which population of electron (majority versus minority) is pref-
erentially scattered at the interface depends upon the choice of materials for NM
and FM, although Cu/Co, Cu/CoFe, and Cu/Py (Py = Ni81Fe19) all exhibit fil-
tering that favors transmission of majority electrons. Spin filtering is illustrated
in Figure 2.3.
12
2.2.2 Spin transfer by averaging
A second method of spin transfer comes from the band structure of the FM material
itself. An electron not collinear with the magnetization propagates towards the
NM/FM with states that are a superposition of the majority and minority basis
states of the FM material, as shown in equation 2.1. The two wave vectors for
these basis states are equal in the NM, k↑x = k↓x. However, once the electron
enters the FM material, the majority and minority basis state wave vectors are
no longer the same, k↑x 6= k↓x [12]. This difference in wave vector leads to a rapid
spatial precession of the electron spin, having a spatial frequency ∆k = k↓x 6= k↑x.
The precession frequency varies dramatically over the entire Fermi surface, and so
when averaged over all current-carrying electrons, the transverse spin polarization
of the total current averages to zero. Spin transfer by averaging is illustrated in
Figure 2.4. The spatial precession which averages out the transverse component
of the net spin polarization take place on the order of a few monolayers in the FM
material [12]. As such, both mechanisms (spin filtering and averaging) effectively
occur at the NM/FM interface.
2.2.3 Spin transfer in an asymmetric FM/NM/FM trilayer
Nearly all spin transfer experiments are performed in a FM/NM/FM trilayer struc-
tures that have either been lithographically defined into a pillar structure, or have
been contacted to by a nanocontact. The typical structure is to make one of the
FM layers thicker than the other. Doing so allows the thinner (or “free”) layer to be
excited by the spin torque while the thicker (or “fixed”) layer provides the spin po-
larization. A schematic of spin transfer in an asymmetric FM(fixed)/NM/FM(free)
trilayer is shown in Figures 2.5 and 2.6. For electrons flowing from the fixed towards
13
yx
z FM
Spatial precession
Figure 2.4: Spin averaging occurs because the majority and minority basis statesdo not share the same wave vectors once the electron enters the FM. This non-zero ∆k = k↓x 6= k↑x leads to rapid spacial precession of the electron spin. Whenaveraged over all current-carrying electrons, the transverse spin polarization of thetotal current averages to zero.
the free layer, the spin polarization is parallel to FM(fixed) and the spin torque
acts to make the two magnetic layers parallel (Figure 2.5). If the electrons are
instead flowing from the free towards the fixed layer, the electrons (spin-polarized
by the free layer) are spin-filtered by the fixed layer. There is a torque imparted
on the fixed layer, but it is too small to excite any response from the thicker fixed
layer. The reflected electrons now strike the free layer with a polarization opposite
to the previous case, and the resulting spin torque acts to make the two layers
antiparallel (Figure 2.6).
14
Fixed Free
Parallel
(a)
(b)
electron flow (-I)
Figure 2.5: Spin torque in an asymmetric FM/NM/FM trilayer. (a) The initialorientation of the two magnetic layers, thick (“fixed”) and thin (“free”), is antipar-allel. Electrons emanating from the fixed layer are spin polarized parallel to thefixed layer, traverse left-to-right across the (invisible) spacer, and impart a torqueon the free layer that acts to make the two layers parallel (b).
15
Fixed Free
Antiparallel
reflected
by spin-filtering
(a)
(b)
electron flow (+I)
Figure 2.6: Spin torque in an asymmetric FM/NM/FM trilayer. (a) The initialorientation of the two magnetic layers, thick (“fixed”) and thin (“free”), is parallel.Electrons now traverse right-to-left across the (invisible) spacer emanating fromthe free layer. The electrons are spin filtered at the fixed layer so that the reflectedelectrons now impinging the free layer are spin-polarized antiparallel to the fixedlayer. The torque imparted on the free layer now acts to make the two layersantiparallel (b).
16
2.3 Magnetization excitations induced by a spin torque
2.3.1 Spin torque in the Landau-Lifshitz-Gilbert equation
Here I introduce some formalism to the spin transfer effect. The singular dynamic
equation in magnetism is the Landau-Lifshitz-Gilbert (LLG) equation.
dm
dt= γ · m× ~Heff − α · m× (m× ~Heff)− (
γh
e)η · I|~m| · m× p× m (2.2)
Equation 2.2 is a summation of different torques acting on a magnetic moment
~m, such that there is a dynamical response dmdt
and, as written here, includes
right hand side terms (from left to right) precessional torque (~τH), damping torque
(~τd), and spin torque (~τst). The first term, the precessional torque, is due to the
effective magnetic field acting on the moment (~τH). The effective field ~Heff is
the net magnetic field acting on ~m, including contributions from demagnetization
fields, applied fields, and any field coming from magnetostatic edge charges, called
dipole fields, on other magnets in proximity to ~m. γ is the gyromagnetic ratio.
This torque acts to precess ~m about ~Heff. The second term in equation 2.2 is the
damping torque (~τd), a dissipative term which acts to reduce the angle of precession.
This known as Gilbert damping with α called the Gilbert damping parameter. The
debate as to the dominant damping mechanisms involved for patterned magnetic
layers is ongoing [13, 14], although a prevalent damping mechanism in magnetic
thin films is spin orbit coupling [15].
The third term in equation 2.2 is the spin torque (~τst) which is imparted on
~m by the spin polarization η from the DC applied current I. p is the polarization
direction. As written in equation 2.2, the angular dependence of the spin torque
17
is simply sin θ, where θ is the angle between p and m. In reality, this spin torque
function likely has more asymmetry about θ = π2
than does a simple sine curve,
as indicated in the original paper by Slonczewski [7] as well as by recent theoreti-
cal [16, 17] and experimental [18] work. Nevertheless, using a simple sine curve is
fine if only for illustrative purposes.
The interaction of all of these torque terms is shown in Figure 2.7. The direction
of ~τH is determined by the direction of the field Heff, the direction of ~τd is always
to reduce the precessional angle towards the precession axis, and the direction of
~τst is defined by the spin polarization direction ~p so that ~m is drawn more parallel
to ~p. In the absence of any spin torque, the moment merely precesses around
Heff, shedding its magnetic energy through damping such that it spirals inwards.
However, a spin torque can, depending on the current polarization, work against
the damping and when it does is sometimes referred to as an “anti-damping”.
Because of the spin torque can do work against the damping, more interesting
dynamics can occur. Under the simplifying assumption that the magnetization be-
haves as a macrospin, namely that all the constituent moments move in lock-step
with each other, these excitations can be subdivided into two types: persistent and
transient dynamics. In the case of persistent dynamics, the moment undergoes pre-
cessional, oscillatory motion essentially indefinitely until the spin torque is turned
off. Second is the case of transient dynamics, where the moment is excited between
two states via some dynamical trajectory. Precessional turn-on (from a static to a
dynamical state) and magnetization reversal (from a static to another static state)
are two examples of transient dynamics. These different types of dynamics are
accessible through the size of the magnetic field applied to the sample [19] and are
shown schematically in Figure 2.8.
18
Figure 2.7: Orientation of the three torques on the magnetic moment ~m repre-sented in the LLG equation of equation 2.2: precession (~τH), damping (~τd), and
spin transfer torque (~τst). ~Heff is the total effective field on the nanomagnet and~p is the direction of spin polarization of the current.
19
(a)
(b)
electron flow
appH
Figure 2.8: Spin torque dynamics of a macrospin magnetic moment. Electrons(red arrows) propagate in from the left and interact with the magnetization m(blue arrow). Depending on the magnitude of the applied field, the dynamicalresponse from the spin torque can be (a) persistent or (b) transient depending onthe magnitude of Happ.
20
2.3.2 Magnetization reversal
Here I illustrate the role of different parameters in the spin torque-driven switching
of a nanomagnet. Specifically, I describe how out-of-plane anisotropy, or demag-
netization field, ~H⊥ (which is directed perpendicular to the plane of the film,
~H⊥ = H⊥x, and is proportional to the out-of-plane component of the magnetiza-
tion mx) and the Gilbert damping α affect the trajectory of a thin-film nanomagnet
by selectively removing both terms from the LLG equation. Seeing the resulting
trajectories gives considerable insight into how spin-torque switching works in the
context of the LLG equation in the macrospin approximation.
The simulations in this section are numerical integrations of the LLG equation
identical to that shown elsewhere [20]. As in equation 2.2, the torque function for
these simulations is a simple sine curve, which is sufficient to illustrate the essential
physics in the spin torque switching dynamics. The nanomagnet in question is a
thin-film magnet patterned into an elliptical shape. The spin torque is orientated so
as to incite a magnetization reversal dynamic, as shown schematically in Figure 2.9,
where the coordinate system is identical to those in Figures 2.2, 2.3, and 2.4,
although the perspective is different. The applied magnetic field Happ is set along
the easy, or long axis of the ellipse (z), parallel to the in-plane anisotropy field HK.
A necessary condition for magnetization reversal is that Happ < HK. Both Happ
and HK contribute to the precessional and damping torques on the magnet (as
per equation 2.2), but do nothing to distort the circular shape of the trajectory.
Since the spin torque (∝ sin θ) is identically zero for θ = 0 or 180 between the
spin polarization ~p and ~m, a small angular deviation (∼1) is introduced in the
simulation to produce a finite spin torque. Field and current values are chosen so
that the resulting trajectories are pedagogically helpful.
21
y
x
z
φ
θ
appH
Figure 2.9: This cartoon shows the orientation of the spin polarization of theincoming current and the magnetization of the elliptical thin-film nanomagnet inquestion. Angular coordinates for the magnetization are the precessional angle (θ)and the polar angle (φ).
H⊥ = 0; α = 0
By setting the out-of-plane demagnetization field H⊥ = 0, we are essentially alter-
ing this elliptical thin-film pancake shaped nanomagnet into a rotationally sym-
metric ellipsoid, where demagnetizing fields are now uniform at all precessional
angles. Removing the damping is akin to rendering this into a “frictionless” mag-
netic system. The simulated spin torque switching is shown in Figure 2.10. The
two important things to notice are the following. First, is that the trajectory is
completely circular (Figure 2.10(a)), which is expected since the spin torque only
affects the precessional angle θ and not the polar angle φ. Secondly, the precession
angle is swept out uniformly over the course of the trajectory (Figure 2.10(b)).
This is because the spin torque is not working against any damping, and as such,
the upper and lower hemispheres of the switching trajectory are mirror symmet-
22
ric. The reversal of precession direction at mz = 0 is a natural occurrence of the
precessional torque term (∝ m× ~Heff) changing sign.
H⊥ = 0; α 6= 0
Now that the damping term is restored, it is expected that the upper and lower
hemispheres of the switching trajectory will no longer be mirror symmetric, al-
though the trajectory still ought to be circular since H⊥ = 0. The spin torque acts
to increase θ, while the damping acts to reduce θ relative to the precession axis.
This means that for θ < π2, the ~τst and ~τd oppose one another, but they both point
in the same direction for θ > π2. This behavior is reflected in the simulated switch,
shown in Figure 2.11.
There is a sizeable asymmetry between upper and lower hemispheres (Fig-
ure 2.11(a)) due to the changing directionality of the damping at θ = π2. The
“ring-up” in the upper hemisphere is where ~τst is doing work against ~τd, whereas
the more rapid decay, or “ring-down”, of the precession in the lower hemisphere
shows that the two are working together. The z-component of the magnetization
mz (Figure 2.11(b)) reflects this confrontation between ~τst and ~τd more subtly by
a slightly more gradual change in mz in the upper hemisphere than the more rapid
change of mz in the precession ring-down in the lower hemisphere.
H⊥ 6= 0; α 6= 0
As a third example, I now include both damping and out-of-plane demagnetization
field, which is much more in-line with a real sample. The effect of H⊥ 6= 0 is clear
from the simulation shown in Figure 2.12, where the trajectory has been flattened
along the out-of-plane direction. H⊥ then acts to keep the magnetization in plane.
23
m x
m y
1
1
1
0
0
0
-1-1
-1
m z
(a)
(b)
mz
Figure 2.10: (a) A simulated switching event for a nanomagnet with H⊥ = 0and α = 0. (b) The z-component of the magnetization, representing the azimuthalprojection of ~m from up to down.
24
(a)
(b)
m x
m y
1
1
1
0
0
0
-1-1
-1
m z
mz
Figure 2.11: (a) A simulated switching event for a nanomagnet with H⊥ = 0and α 6= 0. (b) The z-component of the magnetization, representing the azimuthalprojection of ~m from up to down.
25
The reason for this flattened, more elliptical precession (Figure 2.12(a)) is because
there is now a second precessional axis, coming from H⊥, that points out of plane.
As the moment ~m begins to sweep out of plane from precession due to Happ and
HK (about z), the H⊥ grows in response and induces a rapid precession (about x)
across the plane of the nanomagnet. This traversing motion ends when the moment
re-enters the plane of the nanomagnet at the half-cycle of the normal precession
about Happ and HK. The mz-component (Figure 2.12(b)) now shows oscillations
associated with this elliptical precession. These sorts of oscillations have been
observed in time-resolved measurements of spin torque-driven switching [21].
2.3.3 Persistent dynamics
As shown in the cartoon of Figure 2.8(a), the spin torque can also excite a persistent
oscillatory response from the nanomagnet. In the single domain approximation,
the determining factor for persistent dynamics versus a transient switching event
is that the in-plane applied field Happ must be larger than the in-plane anisotropy
field HK. The range of behavior for spin torque-actuated persistent dynamics is
quite diverse since there is a large expanse of phase space (in current and Happ).
Experimental and theoretical descriptions of persistent dynamics are discussed
elsewhere [7, 19, 22–26]. Using the same simulation as described above, I show a
simulated persistent dynamic turn-on, where Happ > HK, in Figure 2.13. The
flattened, elliptical shape is reminiscent of the switching event of Figure 2.12,
except that now the dynamics wind up in a continuing oscillation in an elliptical
orbit. Note that the oscillations, after turn on, range between -0.35 < mz < 0.95.
26
(a)
(b)
m x
m y
1
1
1
0
0
0
-1-1
-1
m z
mz
Figure 2.12: (a) A simulated switching event for a nanomagnet with H⊥ 6= 0and α 6= 0. (b) The z-component of the magnetization, representing the azimuthalprojection of ~m from up to down.
27
(a)
(b)
m x
m y
1
1
1
0
0
0
-1-1
-1
m z
mz
Figure 2.13: (a) A simulated persistent dynamic turn-on (Happ > HK). (b)The z-component of the magnetization, representing the azimuthal projection of~m from up to down. Notice that mz oscillates between -0.35 and 0.95.
28
2.4 Potential future spin transfer-based technologies
2.4.1 Outlook
The ability to switch a magnetic layer by application of a spin polarized cur-
rent is very appealing from a technological application standpoint. The two most
important reasons for this are first because of the inherent non-volatility of a fer-
romagnetic element, making it a viable candidate for memory applications, and
second because of the low power requirements for spin transfer actuation, making
it appealing for mobile applications that require operation on battery power alone.
However, in spite of recent laboratory results showing tremendous promise for spin
transfer-based technologies involving magnetization reversal [18, 27] and persistent
dynamics [28–30], it must be appreciated how deeply entrenched current technolo-
gies are and how much inertia high-volume production companies have. Often,
millions of dollars have been invested in fabrication tools which makes shifting to
a completely new technology economically unviable. Because of the large capi-
tal investment, high technology companies typically plan their technology market
strategies nearly a decade in advance. So, if a new technology comes along, such as
spin transfer, it is not likely to be adopted as a mainstream technology merely be-
cause “it works”. However, video cassettes are rapidly being replaced with DVDs,
portable CD players are nearly obsolete due to portable MP3 players, and FLASH
memory has removed the need for floppy drives. So clearly, technological shifts
do occur, and in fact, are inevitable, but it is important to recognize how they
happen. Prospects for any new technology, including spin transfer-based devices,
to be successfully adopted into high-volume production weigh on four separate
possibilities.
29
Disruptive technology
First is the possibility that spin transfer devices will render all competing tech-
nologies obsolete and would consequently demand industry-wide adoption. This is
known as “revolutionary” or “disruptive” technology which may require and even
justify a complete replacement of supporting technologies or technological architec-
tures. One of the most extreme examples of this type was the invention of the solid
state transistor. Its superiority over its primary competition, the vacuum tube,
made its ubiquity in industrial application inevitable, even though it required new
supporting hardware architectures (power supplies, wiring, controls, etc.).
Substitutive technology
A second possibility is for spin transfer devices to replace an incumbent technology
once it has become untenable or too costly to continue production of subsequent
generations of the same device technology but without generating an entirely new
supporting technology architecture. This is known as “substitutive” technology
because the supporting architecture from the now obsolete devices would only see
ancillary changes (at most) for the replacement devices. This demands a different
set of requirements for successful adoption because now the spin transfer device
technology must conform to the pre-existing architecture (voltage levels, signal-to-
noise (SNR) requirements, operating temperature, etc.). There may also be other
competing incremental technologies and it is not always clear which is the best
alternative. A good example of this, although not at the device level, was the
move to replace the 3.5” floppy disk drive. Before CD burning and FLASH drives
became available, there were several options, such as magneto-optical drives, ZIP
disks, and tape drives.
30
Paradigm-shifting technology
A third possibility for spin transfer, which is similar to the first two in that an in-
cumbent technology is replaced with the new technology, is different in that it does
not rely on the individual device of the new technology being superior. Instead,
it is the system of many inferior devices working together that produce a superior
technology. This sort of technological change relies on a significant, possibly even
complete redesign of the architecture on top of which the devices operate and may
represent the “last option” a company has it if disruptive of substitutive options
are not available. This technology change is labelled as “paradigm-shifting” be-
cause it usually calls for a re-evaluation of accepted methods and designs that have
worked previously, but are somehow failing to meet current needs. This is some-
times associated with a sense that the current technology has grown to be overly
complicated, and is beginning to fail out of sheer complexity. Such a phenomenon
has been discussed in operating system software platforms [31]. Paradigm-shifting
technology and substitutive technology could both be applied to fulfill the same
technological problem. Namely, that the current technology will soon fail and a
new technology must be implemented in order for a company to survive.
Niche technology
The fourth possibility is for some highly specific, or “niche”, application that
could be filled by spin transfer. Unlike the previous three cases, where I have
looked at spin transfer as somehow replacing a pre-existing device technology, the
niche allows for spin transfer to select its market based on its strengths. For spin
transfer-actuated devices, the strengths are non-volatility, endurance, low power,
31
and speed1. This kind of technology will usually be successful in more specialized
markets, rather than in mainstream markets.
2.4.2 Microwave oscillators
Spin transfer devices, which can exhibit microwave signals (see e.g. ref. [19]), are
desirable for such oscillator applications because they are local (i.e. on-chip) and
actuated by a DC current. This gets around the cumbersome problem of interfac-
ing with an off-chip oscillator crystal. Oscillators are of paramount importance to
communications technologies as information will be encoded by amplitude (AM)
or frequency (FM) modulation on the carrier waves they produce that can then be
broadcast. Recent demonstration of frequency modulation of spin transfer oscilla-
tors [28] confirms that such devices are technically compatible with the requisite
behavior for communications technologies. The Figure of merit for an oscillator
source is its Q factor, the ratio of the oscillation frequency (f0) to the full-width
half-max (FWHM) of the peak in frequency space. The FWHM is an estimate of
how much variation there is in f0. Qs on the order of 104 have been observed in
nanoconstriction point contacts to continuous magnetic multilayer films [28], an
important step in furthering the spin transfer-based technology.
Another technical consideration is the output power of these oscillators. Two
recent papers have shown that, by placing two nanoconstriction contacts in proxim-
ity to one another on a magnetic multilayer film, current through both constrictions
can excite “phase-locked” oscillations in the magnetic material underneath each
point contact, meaning that both magnetic regions naturally become dynamically
1The claim of high density, such as in MRAM, has not yet been confirmed forCPP spin transfer devices.
32
synchronous [29, 30]. This has the combined effect of increasing output power,
as phase variation between the two dynamically excited magnetic sections can
decrease overall signal2, and also reducing the FWHM since the active magnetic
volume has doubled in size and is hence less susceptible to thermal fluctuations
which are a main source of linewidth broadening. Modulation and phase-locking of
microwave signals are positive developments for spin transfer-actuated microwave
sources as a real technology. If adopted as an on-chip microwave oscillator, this
would be viewed as a “disruptive” technology. This is because the outgoing res-
onator or oscillator technology still continues to work, but has been surpassed by
the superior spin transfer technology.
2.4.3 Global storage
The limits of future hard drive storage
Progress in hard drive storage technology over the past several years has achieved
areal densities of 100 Gb/in2, and are rapidly approaching 1 Tb/in2. In fact, the
rate of areal density growth is faster than “Moore’s Law”, which is the equivalent
scaling trend for computer processers, measured in transistors/in2 (see e.g. [32]).
This poses a tremendous challenge to hard drive companies because the technology
of hard drives is approaching fundamental limits. In a hard drive, there are four
major technologies working in unison, each of which has its own problems and
challenges. The first two, which I mention here but do not expound on, are the
mechanical components of the hard drive, namely the boom which holds the head
at a fly height of ∼10 nm above the platter to a precision of the track width as the
2In fact, the increase in power is non-linear and, during phase-locking, shouldbe larger than the net sum of the two constituent output powers [29].
33
platter spins at speeds upwards of ∼7200 rpm (see e.g. [33]), and the write head
which encodes the information to the magnetic media one bit at a time without
erroneously addressing neighboring bits.
The other two, the platter media, the magnetic films which store that actual bits
of data, and the read head that collects the information from the media are both
approaching limitations, mainly in SNR. In the hard drive industry, success thrives
on increasing the areal density while maintaining a sufficient SNR level. As areal
densities increase, length scales for the physically relevant parameters of media
(track width and bit length) and read heads (sensor width) shrinks proportionately.
For the media material (with grain volume V and grain anisotropy Ku), there are
three intertwined challenges, known as a “trilemma”, where the storage bits must
simultaneously be writable (small Ku), thermally stable (large Ku · V ), and retain
high SNR (small V ,3). Future short term goals are to employ perpendicular (i.e.
flux out of plane) media as opposed to longitudinal (in-plane) media. In the longer
term though, other media materials ideas include patterned media, where the bits
are lithographically defined, or heat-assisted writing, where the bits are engineered
with high anisotropy and can only be switched with the application of a short heat
pulse from a laser adjacent to the write head.
Future read heads must have the sensitivity to detect the small amount of
magnetic flux coming from reduced bit size. And as such, read head design has
changed to meet this added sensitivity initially changing from inductive read heads
to anisotropic magnetoresistance (AMR), to current in-plane giant magnetoresis-
tance (CIP-GMR), to current perpendicular-to-plane (CPP-GMR), with future
plans involving tunnelling magnetoresistance (TMR) [32]. In addition to read
3Small grain volume means more grains per bit N , and SNR ∝ √N [34].
34
head physics, increased sensitivity can be gained through nano-scale dimensions
on the order of 10 nm [35]4 With future developments in media and read heads
field experts have claimed that the hard drive, as we know it today, will exist for
10 to 15 more years.
Magnetically-based storage beyond the hard drive
Beyond this time-line, what will take the place of hard drives for future storage
technology? In addition to this un-named replacement technology having at least
the same capabilities as the current hard drive, it must also be inexpensive to
make. Over the next 10 to 15 years, fabrication technology will definitely change,
so it is not realistically possible to know exactly what will be cheap to make in
the future. All the same, the need for fast access times (i.e. read/write time),
non-volatility, and density—the same advantages which have made hard drives so
successful in the past and present—make magnetically-based storage technologies
appealing.
There is a possible future for spin transfer-actuated global storage, which could
be classified as a “paradigm-shifting” technology because it would be a complete
re-design of the storage hardware and would almost unquestionably see the removal
of a spinning platter scanned by a read/write head at the end of a robotic arm.
A possible geometry of spin torque-actuated recording (STAR) is an architecture
similar to MRAM, which employs a two-dimensional array of magnetic tunnel
junctions (MTJ), each of which is independently accessible through two sets of
perpendicular write lines, shown in Figure 2.14 [36–38]. The hypothetical STAR
storage would be a three-dimensional extension of this, as shown in Figure 2.15.
4Spin transfer itself is considered a significant noise source in CPP-GMR readheads.
35
Bit Lines
Wor
d Lin
es
Figure 2.14: Magnetoresistive random access memory (MRAM) two-dimensionaladdressing architecture, where all MTJs lie between two mutually perpendicularsets of lines, a ‘Bit’ and ‘Word’ line. Each MTJ is written by the sum of Oerstedfields from the two lines.
In this storage architecture, individual MTJs are the bits that can be accessed
by combination of ‘Bit’ and ‘Word’ lines, just as in the two-dimensional MRAM
architecture, but is essentially infinitely repeatable in higher levels.
2.4.4 Programmable logic
Miniaturization limitations on current logic elements
One of the best examples of a “substitutive” technology has been Si transistor
scaling. Miniaturization of Si devices is an ongoing practice that has propagated
the computer industry for nearly four decades. However, the foreseeable end to
this scaling trend is coming into focus (see e.g. ref. [39]). What will replace Si
as the next great technology? At present, no one knows what the right answer
is. However, there are several candidates vying for this position, including spin,
molecular, and carbon-based electrons as well as “computation-on-demand” [31].
36
Bit Lines
Wor
d Lin
es
Lev
els
Iwrite
Figure 2.15: Spin torque-actuated recording (STAR), a hypothetical storage con-cept where a three-dimensional network of MTJs, each accessible through indepen-dent ‘Bit’ and ‘Word’ lines, similar to the MRAM writing scheme (Figure 2.14),except now the devices are written by CPP-spin transfer torque. There is no limitto how many levels can be added in this architecture.
37
New computer architecture for the “end of CMOS scaling”
Logic operations, such as AND, OR, etc. comprise basic operations of a com-
puter processor upon which more complex mathematical computations may be
performed. The current, rigid transistor-based logic elements have their function-
ality fixed by wiring. Once executed, the information must be put into memory to
prevent it from being lost since the elements are volatile. Imagine, instead, if these
basic logic operations were reconfigurable in real-time. For example, being able to
controllably change the logic operation from AND to NAND. Such reconfigurable,
or programmable logic, allows a means of achieving faster computational speed
(without having to resort to further scaling of devices) as a processor would be
able to reconfigure itself in real-time to achieve maximum computational efficiency.
Several papers have discussed the efficacy of employing MTJs as the individ-
ual logic elements [40–42] due to their inherent non-volatility, which would save
time in the overall computation by not having to store the data off-chip. Like
MRAM (Figure 2.14), the MTJs are actuated by Oersted field writing. One
such publication [40] describes how a single MTJ can be configured to AND, OR,
NAND, and NOR logic operations with a coordinated programming sequence be-
tween the two input lines A and B and an additional programming line C. The
FM(switch)/barrier/FM(reference) MTJ logic element is magnetically engineered
such that the Oersted field from a single input/programming line (A, B, or C) is
insufficient to switch any layer, while the combination of fields from the input lines
(A + B) can switch FM(switch) and the sum of all three (A + B + C) switches
both FM layers. Logic functionality is pre-selected by setting different antipar-
allel or parallel orientations of the MTJ by addressing various combinations or
sequences of A, B, and C. Negation operations (NAND and NOR) are achieved
38
simply by reversing the direction of FM(reference) layer between addressing A, B,
and C beforehand.
Spin transfer-actuation would be slightly more complicated than employing
Oersted fields for programmable logic applications. This is because, for Oersted
field-actuation, the polarity of switching from the input lines A and B is inde-
pendent of the orientation of the FM(reference) layer, while this is not the case
for spin transfer switching, where polarity of switching is defined by the relative
orientation of FM(switch) and FM(reference) (as shown in Figures 2.5 and 2.6).
Consequently, negation of the AND and OR operations (NAND and NOR), which
was achieved by Oersted field-actuation through reorientation of FM(reference),
must be achieved in other ways. One possible method is clever engineering of
nanopillar layering where, by incorporation of spin-valve and tunnel junctions, one
is able to tune different thresholds for switching currents. This is a practical goal
since, in general, spin torque per unit current is dissimilar between tunnelling and
electron transport spin transfer phenomena [7, 27, 43].
Comparing switching speeds between spin transfer-actuated logic elements and
Si-based transistor logic, spin transfer switching has been measured in spin-valves
with the conventional magnetization-in-plane geometry on the timescale of nanosec-
onds [14, 18, 21, 44], nearly an order of magnitude slower than Si-based transistors.
However, faster switching times (∼100 psec, more in line with what is achieved with
Si) are predicted to come from spin transfer switching with out-of-plane spin polar-
ization [45], akin to precessional switching from field-driven reversal [46]. Regard-
less, the efficiency and non-volatility of a magnetically-based programmable logic
architecture may increase the overall computational speed, making programmable
logic a viable “paradigm-shifting” alternative to CMOS.
39
References for Chapter 2
[1] Baibich M.N., Broto J.M., Fert A., NguyenVanDau F., Petroff F., EitenneP., Creuzet G., Friederich A., & Chazelas J., Giant magnetoresistance of(001)Fe/(001)Cr magnetic superlattices, Phys. Rev. Lett. 61, 2472–2475(1988).
[2] Binasch G., Grunberg P., Saurenbach F., & Zinn W., Enhanced magnetore-sistance in layered magnetic structures with antiferromagnetic inerlayer ex-change, Phys. Rev. B 39, 4828–4830 (1989).
[3] W. P. Pratt J., Lee S.F., Slaughter J.M., Loloee R., Schroeder P.A., & BassJ., Perpendicular giant magnetoresistance of Ag/Co multilayers, Phys. Rev.Lett. 66, 3060–3063 (1991).
[4] Gijs M.A.M., Lenczowski S.K.J., & Giesbers J.B., Perpendicular giant mag-netoresistance of microstructured Fe/Cr magnetic multilayers from 4.2 to 300K, Phys. Rev. Lett. 70, 3343–3346 (1993).
[5] Valet T. & Fert A., Theory of perpendicular magnetoresistance in magneticmultilayers, Phys. Rev. B 48, 7099–7113 (1993).
[6] Parkin S.S.P., More N., & Roche K.P., Oscillations in exchange coupling angmagnetoresistance in metallic superlattice structures Co/Ru, Co/Cr, andFe/Cr, Phys. Rev. Lett. 64, 2304–2307 (1990).
[7] Slonczewski J.C., Current-driven excitation of magnetic multilayers, J. Magn.Magn. Mater. 159, L1–L7 (1996).
[8] Waintal X., Myers E.B., Brouwer P.W., & Ralph D.C., Role of spin-dependentinterface scattering in generating current-induced torques in magnetic multi-layers, Phys. Rev. B 62, 12317–12327 (2000).
[9] Stiles M.D., Spin-dependent interfacial transmission and reflection in magneticmultilayers (invited), J. Appl. Phys. 79, 5805–5810 (1996).
[10] Upadhyay S., Louie R.N., & Buhrman R.A., Spin filtering by ultrathin ferro-magnetic films, Appl. Phys. Lett. 74, 3881–3883 (1999).
[11] Stiles M.D. & Zangwill A., Anatomy of spin-transfer torque, Phys. Rev. B 66,014407 (2002).
[12] Stiles M.D. & Zangwill A., Noncollinear spin transfer in Co/Cu/Co multilay-ers, J. Appl. Phys. 91, 6812–6817 (2002).
[13] Tserkovnyak Y. & Brataas A., Enhanced Gilbert damping in thin ferromag-netic films, Phys. Rev. Lett. 88, 117601 (2002).
40
[14] Emley N.C., Krivorotov I.N., Garcia A.G.F., Ozatay O., Sankey J.C., RalphD.C., & Buhrman R.A., Time-resolved spin torque switching and enhanceddamping in Py/Cu/Py spin-valve nanopillars (2005), submitted to PRL.
[15] Schreiber F., Pflaum J., Frait Z., Muhge T., & Pelzl J., Gilbert damping andg-factor in Fe(x)Co(1-x) alloy films, Solid State Comm. 93, 965–968 (1995).
[16] Slonczewski J.C., Currents and torques in metallic magnetic multilayers, J.Magn. Magn. Mater. 247, 324 (2002).
[17] Xiao J., Zangwill A., & Stiles M.D., Boltzmann test of Slonczewskis theoryof spin-transfer torque, Phys. Rev. B 70, 172405 (2004).
[18] Braganca P.M., Krivorotov I.N., Ozatay O., Garcia A.G.F., Emley N.C.,Sankey J.C., Ralph D.C., & Buhrman R.A., Reducing the critical currentfor short-pulse spin-transfer switching of nanomagnets, Appl. Phys. Lett. 87,112507 (2005).
[19] Kiselev S.I., Sankey J.C., Krivorotov I.N., Emley N.C., Schoelkopf R.J.,Buhrman R.A., & Ralph D.C., Microwave oscillations of a nanomagnet drivenby a spin-polarized current, Nature (London) 425, 380–383 (2003).
[20] Sun J.Z., Spin-current interaction with a monodomain magnetic body: Amodel study, Phys. Rev. B 62, 570–578 (2000).
[21] Krivorotov I.N., Emley N.C., Sankey J.C., Kiselev S.I., Ralph D.C., &Buhrman R.A., Time-domain measurements of nanomagnet dynamics drivenby spin-transfer torques, Science 307, 228–231 (2005).
[22] Kiselev S.I., Sankey J.C., Krivorotov I.N., Emley N.C., Rinkoski M., PerezC., Buhrman R.A., & Ralph D.C., Current-induced nanomagnet dynamicsfor magnetic fields perpendicular to the sample plane, Phys. Rev. Lett. 93,036601 (2004).
[23] Kiselev S.I., Sankey J.C., Krivorotov I.N., Emley N.C., Garcia A.G.F.,Buhrman R.A., & Ralph D.C., Spin-transfer excitations of permalloy nanopil-lars for large applied currents, Phys. Rev. B. 72, 064430 (2005).
[24] Rippard W.H., Pufall M.R., Kaka S., Russek S.E., & Silva T.J., Direct-currentinduced dynamics in Co90Fe10/Ni80Fe20 point contacts, Phys. Rev. Lett. 92,027201 (2004).
[25] Berger L., Emission of spin waves by a magnetic multilayer traversed by acurrent, Phys. Rev. B. 54, 9353–9358 (1996).
[26] Bazaliy Y.B., Jones B.A., & Zhang S.C., Modification of the Landau-Lifshitzequation in the presense of a spin-polarized current and colossal-and giant-magnetoresistive materials, Phys. Rev. B. 57, R3213–R3216 (1998).
41
[27] Fuchs G.D., Krivorotov I.N., Braganca P.M., Emley N.C., Garcia A.G.F.,Ralph D.C., & Buhrman R.A., Adjustable spin torque in magnetic tunneljunctions with two fixed layers, Appl. Phys. Lett. 86, 152509 (2005).
[28] Pufall M.R., Rippard W.H., Kaka S., Silva T.J., & Russek S.E., Frequencymodulation of spin-transfer oscillators, Appl. Phys. Lett. 86, 082506 (2005).
[29] Kaka S., Pufall M.R., Rippard W.H., Silva T.J., Russek S.E., & Katine J.A.,Mutual phase-locking of microwave spin torque nano-oscillators, Nature (Lon-don) 437, 389–392 (2005).
[30] Mancoff F.B., Rizzo N.D., Engel B.N., & Tehrani S., Phase-locking in double-point-contact spin-transfer devices, Nature (London) 437, 393–395 (2005).
[31] Nair R., Role of new technology in computing systems ofthe future, Cornell Center for Nanoscale Systems (2005),http://www.cns.cornell.edu/CNSAnnualSymposium2005.html.
[32] Covington M., Spin transfer and other challenges in datastorage, Cornell Center for Nanoscale Systems (2004),http://www.cns.cornell.edu/documents/MarkCovingtonCNSSymp504.pdf.
[33] Hall M., Seagate powers world’s most popular PC hard drive with half-terabyte and faster 3 Gbit/s serial ATA interface; introduces faster, highercapacity external and portable drives, Technical report, Seagate Technol-ogy (2005), http://www.seagate.com/cda/newsinfo/newsroom/releases/ arti-cle/0,,2733,00.html.
[34] O’Handley R.C., Modern Magnetic Materials: Principles and Applications,chapter 17, p. 696, John Wiley & Sons, Inc. (2000).
[35] Driskill-Smith A.A.G., Katine J.A., Druist D.P., Lee K.Y., Tiberio R.C., &Chiu A., Electron-beam lithography for the magnetic recording industry: fab-rication of nanoscale (10 nm) thin-film read heads, Microelectronic Engineer-ing 73-74, 547–552 (2004).
[36] Slaughter J.M., Dave R.W., DeHerrera M., Durlam M., Engel B.N., JaneskyJ., Rizzo N.D., & Tehrani S., Fundamentals of MRAM technology, J. Super-cond. 15, 19–25 (2002).
[37] Engel B.N., Rizzo N.D., Janesky J., Slaughter J.M., Dave R., DeHerrera M.,Durlam M., & Tehrani S., The science and technology of magnetoresistivetunneling memory, IEEE Trans. Nanotech. 1, 32–38 (2002).
[38] DeBrosse J., Gogl D., Bette A., Hoenigschmid H., Robertazzi R., Arndt C.,Braun D., Casarotto D., Havreluk R., Lammers S., Obermaier W., Reohr W.,Viehmann H., Gallagher W.J., & Muller G., A high-speed 128-kb MRAM
42
core for future universal memory applications, IEEE J. Sol.-State. Circ. 39,678–683 (2004).
[39] Solomon P., Is there life beyond CMOS?, Cornell Center for Nanoscale Sys-tems (2005), http://www.cns.cornell.edu/documents/cnf05.pdf.
[40] Ney A., Pampuch C., Koch R., & Ploog K.H., Programmable computing witha single magnetoresistive element, Nature (London) 425, 485–487 (2003).
[41] W. C. Black J. & Das B., Programmable logic using giant-magnetoresistanceand spin-dependent tunneling devices (invited), J. Appl. Phys. 87, 6674–6679(2000).
[42] Richter R., Boeve H., Bar L., Bangert J., Klostermann U.K., Wecker J., &Reiss G., Field programmable spin-logic based on magnetic tunnelling ele-ments, J. Magn. Magn. Mater. 240, 127–129 (2002).
[43] Slonczewski J.C., Currents, torques, and polarization factors in magnetic tun-nel junctions, Phys. Rev. B 71, 024411 (2005).
[44] Kaka S., Pufall M.R., Rippard W.H., Silva T.J., Russek S.E., Katine J.A., &Carey M., Spin transfer switching of spin valve nanopillars using nanosecondpulsed currents, J. Magn. Magn. Mater. 286, 375–380 (2005).
[45] Kent A.D., Ozyilmaz B., & del Barco E., Spin-transfer-induced precessionalmagnetization reversal, Appl. Phys. Lett. 84, 3897–3899 (2004).
[46] Kaka S. & Russek S.E., Precessional switching of submicrometer spin valves,Appl. Phys. Lett. 80, 2958–2960 (2002).
CHAPTER 3
NANOPILLAR FABRICATION AT CORNELL: PRESENT AND
FUTURE
3.1 Introduction
This chapter summarizes much of the sizeable and diverse nanofabrication expertise
within the Cornell nanomagnetics group. The highly successful C (carbon) mask
nanopillar fabrication process, first outlined by Frank Albert in his Ph.D. thesis [1],
is now the standard recipe for all spin-transfer nanopillar devices fabricated at
Cornell. As more nanopillar wafers have been successfully processed, fine tuning
of the fabrication steps has increased device yield significantly. The first section
summarizes the most current set of process guidelines for nanopillar fabrication.
The rest of the chapter discusses further details on current techniques but also
introduces some new nanofabrication techniques available at the Cornell NanoScale
Science and Technology Facility (CNF) that are as of yet unused in nanopillar
processing but may be of use for future nanopillar fabrication processes. Such
techniques may prove useful for more efficient nanopillar fabrication recipes with
still better device yield (e.g. section 3.9) or for fabrication of more complex devices
(e.g. Ch.6). This chapter serves as a bridge linking the current level of fabrication
expertise with explored but unutilized techniques that may open new possibilities
for future research projects.
43
44
3.2 C mask nanopillar fabrication process
This section comprises the most up to date guidelines for nanopillar fabrication
using the C mask technique. It may serve as a guide to all those new to the process.
3.2.1 Overview
This section is a step by step visual guide to the C mask fabrication process
originally outlined in a previous thesis [1].
45
Figure 3.1: Sputter deposition of metallic layers.
Figure 3.2: Evaporate C (egun).
46
Figure 3.3: Spin PMMA bilayer.
47
Figure 3.4: E-beam lithography to define nanohole in PMMA.
48
Figure 3.5: Develop and evaporate Cr.
49
Figure 3.6: Liftoff Cr.
Figure 3.7: O2 plasma C.
50
Figure 3.8: Photolith 1 [PL(1)]: “Define Leads” mask. Black is Cr, white isglass.
51
Figure 3.9: Ion Mill 1 [IM(1)]: “Isolate Devices”.
52
Figure 3.10: Photolith 2 [PL(2)]: “Define Pillar” mask. Black is Cr, white isglass.
53
Figure 3.11: Ion Mill 2 [IM(2)]: “Define Pillar”.
54
Figure 3.12: Ion Mill 2 [IM(2)]: “Define Pillar”.
Figure 3.13: Evaporate 50 A of SiOx to protect C mask from etching during thePECVD oxide deposition.
55
Figure 3.14: PECVD oxide to insulate nanopillar.
56
Figure 3.15: Ion Mill 3 [IM(3)]: “Planarize SiO2”.
57
Figure 3.16: Ion Mill 3 [IM(3)]: “Planarize SiO2”.
58
Figure 3.17: Photolith 3 [PL(3)]: “HF Etch” mask. Black is Cr, white is glass.
59
Figure 3.18: Profilometry to measure oxide height above pillar.
60
Figure 3.19: Photolith 4 [PL(4)]: “Protect Shorts” mask. Black is Cr, white isglass.
61
Figure 3.20: Evaporate oxide.
62
Figure 3.21: Photolith 5 [PL(5)]: “Oxide Window” mask. Black is Cr, white isglass.
63
Figure 3.22: Ion Mill 4 [IM(4)]: “Thin Oxide Above Pillar”.
64
Figure 3.23: Photolith 6 [PL(6)]: “Top Leads” mask. Black is Cr, white is glass.
65
Figure 3.24: Ion Mill 5 [IM(5)]: “Open Pillar”.
66
Figure 3.25: Oxygen plasma to remove C cap.
67
Figure 3.26: Ion Mill 6 [IM(6)]: “Clean Top Contact”.
68
Figure 3.27: Deposit top leads in situ.
69
Figure 3.28: Completed device.
70
3.2.2 Guidelines as of late 2005
This section is my attempt to document a reliable, thorough, and up-to-date set of
fabrication guidelines so that fabrication lore is not lost as other students familiar
with the process graduate. Subsequent changes will be made as unforseen issues
arise, so treat this set of guidelines as an evolving document. This section does not
give a step by step run through of the process, but gives important details about
all of the steps in the C mask process and should be used in conjunction with the
process overview in the previous section.
Sputter wafer (AJA in D12)
There should be at least two managers for the AJA sputter system. The tool sees
heavy use and so by upholding the present scheduling system, potential conflict
between users will be avoided. All target changes must be cleared with the man-
agers, and they are free to choose who will be trained on target changing. Tool
maintenance must be headed by the managers and all adjustments, changes, or
upgrades to the system must both be documented in the logbook and be made
aware of to the AJA user community.
Sputter calibrations should be done for each gun that is disassembled during a
vent. Any new sputter condition (e.g. different gas, pressure, power, etc.) requires
a new calibration. Whenever the chamber sees oxygen (i.e. after pump down or
after oxidation of layers), all oxidizable targets must be conditioned (plasma on,
shutter closed) for 1 – 1.5 min. Any more than this and you are just wasting target
material. The N2 gun next to the AJA must have a filter on it as film damage may
result if used without one. See Table 3.1 for general sputter guidelines.
71
Table 3.1: Sputter guidelines for the AJA. There is now so much flexibility inthee capabilities of the tool, it is almost misleading to include power levels, sputterpressures, and deposition rates. All these should be recorded in the logbook.
Base pressure ∼10−8 Torr
Stage height is indicated on the ruler for both magnetic
and normal stages
Stage rotation = 100%
Gun tilt = 0.4”
C mask layer
This section discusses the C layer, which serves as the ion mill mask which defines
the top-down size and shape of the nanopillar. The C film is electron-gun (e-gun)
deposited, in either of the SC4500 CNF evaporators, after the sputter deposition
of the multilayer in the AJA. Despite the fact that the C evaporation is the first
step after sputter deposition, the top capping layer (typically Pt) can become
contaminated with particulates and adsorbed water from the air. Both of these
contaminants can lead to fracture and lift off if the wafer surface is not properly
treated before deposition. Lift off of the C, if it does occur, usually does so during
sonication, which is used for electron-beam (e-beam) lithography development and
stripping. Removal of these contaminants before C evaporation is successfully
accomplished with a brief hotplate bake and N2 spray (see Table 3.2).
The evaporation source is a crucible filled with a small clumpy pile of 1/4”
diameter × 1/4” long cylindrical C rods. Consequently, the surface seen by the
beam continually changes as the beam is swept out over the source. This gives rise
to a very erratic deposition rate which persists throughout the entire deposition
72
Table 3.2: Pre-clean the wafer before C evaporation, getting rid of surface mois-ture and other particulates. The bake is more important to do during the morehumid times of the year. Always use the N2.
(1) Hotplate bake at 170 for 5 min
(2) N2 spray before loading into evaporator
Table 3.3: Parameters for C e-gun evaporation.
Parameter Value
Density n 2.25 g/cm3
Z ratio 3.26
Power ramp rate 1%/min
Evap rate < 25 A/min
Threshold 10 - 11%
since the C does not melt or smooth itself out. Although the instantaneous rate
fluctuates quite dramatically, when averaged out over a minute the maximum rate
should be < 25 A/min (see Table 3.3).
Evaporated C films are fragile because of the low energies (∼1 eV) with which
the C atoms impinge the substrate surface. These low energies lead to sp2 bonding
between C atoms which is considerably weaker than sp3 bonds. Formation of
the stronger sp3 bond is a physical process [2] and occurs for incoming C atoms
more energetic than what is achievable with evaporation. The strength of sp3
bonding typifies diamond like carbon (DLC) films, which can be deposited by such
high energy processes (> 100 eV) as ion beam deposition (IBD), plasma-enhanced
chemical vapor deposition (PECVD), or pulsed laser deposition (PLD) [2].
73
Interestingly, the main reason why C has been such a successful material for
nanopillar fabrication at Cornell is because of its fragile nature. Removing the C
from the wafer is done with an O2 plasma, typically in the RIEx tool. Etch rates
are on the order of 40 nm/min in a 90W plasma with 30 mTorr O2 flowed at 30
sccm. An IBD-deposited C film had an almost negligible etch rate under the same
plasma conditions, possibly indicating a more diamond like film requiring either
higher plasma powers or a more reactive etch gas to remove.
Tests of a 1000 A C film e-gun evaporated onto a Cu-coated wafer which had
been treated only with the N2 spray (Table 3.2) yielded the following results when
exposed to common fabrication chemical treatments. The wafer was scribed into
chips before these tests were performed. All tests were done on a single chip except
for the n-methyl-2-pyrrolidone (NMP) bath.
Acetone ultrasound for 2 min:
Almost no peeling aside from some C film fracture where the glass slide was
put down for scribing.
Spin & bake S1813:
Normal processing speeds, temperatures, etc. (Table 3.9, except N2 was used
to clean off chip while spinning (i.e. not Acetone/IPA). No visible change in C
film quality. Also, no discoloration of Cu.
Strip S1813:
Acetone bath ∼1 min, spray in transit to new Acetone bath. Sonicate for 1
min, spray in transit to IPA bath, N2 blow dry. No visible change in C film quality.
Spin & bake PMMA:
Normal processing speeds, bakes, times, etc. (Table 3.4). No visible change in
C film quality, although any exposed Cu shows discoloration and roughness (likely
74
reaction with PMMA or its solvent Anisole).
O2 etch of C:
Stripped PMMA and etched C film in an O2 plasma (RIEx) 90 sec (standard
plasma params: 90W, 30 sccm, 30 mTorr). No problems with C etching after
photoresist and e-beam resist processing.
Soak in heated NMP bath for ∼6 hours:
This is on a fresh C film not put through previous processing steps. No visible
change in C film quality.
VB6 exposure
It has become clear that the specific wafer layerings, causing fine differences in the
flux of electrons backscattered from the layers themselves, affect the resulting pillar
dimensions. As such it is recommended that, for each set of different layerings, one
should perform a dose test to ascertain the actual sizes of the nanopillar exposures.
The C mask must be included for the dose test.
Once the C is down, the next step is to spin on the PMMA bilayer (Table 3.4).
Empirically, it has been found that during the bake of the bottom PMMA layer,
pits can sometimes form in the C layer, rendering it useless as a mask layer. A
correlation was made between the formation of the pits and whether or not the
wafer was cleaned with the typical Acetone/IPA spray (while spinning) just before
coating with e-beam resist. Wafers not sprayed with Acetone/IPA never showed
this pitting problem. Although the precise reason for the pitting is unknown, it
may be related to Acetone drying on the wafer surface, since an Acetone soak
before PMMA spin-coating yielded no pitting (see previous section). It is best to
clean the C-coated wafer with N2 only to avoid this problem. If pitting does occur,
75
Table 3.4: PMMA bilayer spin-coating.
(1) N2 spray clean wafer while spinning
(2) Spin resist (4% 495k PMMA in Anisole) at 3200 rpm for 60 sec
(bottom layer, 161 nm)
(3) Solvent bake on 170C hotplate for 15 min
(4) Chill block for 10 sec
(5) N2 spray clean wafer while spinning
(6) Spin resist (2% 950k PMMA in MIBK) at 2000 rpm for 60 sec
(top layer, 106 nm)
(7) Solvent bake on 170C hotplate for 15 min
(8) Chill block for 10 sec
strip the resist and O2 plasma off the C layer. Re-evaporate the C and proceed.
To properly dispense the PMMA, use a fresh clean pipette for both resists.
Prep the pipette by spraying off the outside of the pipette with N2 and then by
siphoning some PMMA from the top of the fluid to rinse out the pipette interior.
Dispense into the spinner bowl and then siphon fresh PMMA from the top of the
fluid for spinning. Expel four drops into the spinner bowl before applying to the
wafer. Press and hold the start pedal immediately after dispensing as the PMMA
can begin to coagulate in a few seconds. When spinning the PMMA, if, for some
reason, either of the PMMA layers is not satisfactory (e.g poor coverage, PMMA
coagulation, or dust) it is best to strip the resist, remove the C in an O2 plasma,
re-evaporate the C, and then re-spin the PMMA.
Just before doing the VB6 exposure, at the end of the jobcal execution, an
error saying “WARNING: NO CONJUGATE BLANKING” may appear. The
76
Table 3.5: Develop the PMMA bilayer.
(1) Sonicate in IPA:DI Water (7:3) mixture for 60 sec
(2) Transfer to IPA bath
(3) Flush with IPA
(4) N2 blow dry
information sheet provided by CNF staff says to ignore this warning, but it is
best if the user re-runs the jobcal. Conjugate blanking prevents the beam from
exposing regions (≤ 100 nm) just outside the nanopillar exposure (see for example
Figure 3.114(b)).
PMMA development
The developer used is IPA:DI Water (7:3) with sonication [3]. The developing
process is shown in Table 3.5. The term “Flush” means to spray continuously for
3 – 5 sec over entire wafer surface. Even if the wafer was properly cleaned before
the C evaporation, some C lift off may still occur but typically < 5% of the total
wafer area, and predominantly at the wafer edge.
Evaporation and lift off of Cr pattern-transfer layer
The Cr serves to transfer the e-beam lithography pattern into the underlying C.
The bilayer recipe exposes so that there is an undercut profile. This allows for
easy liftoff of the Cr evaporation. Deposition can be done in either of the SC4500
evaporators at the CNF. Use Cr rods for thermal evaporation, although e-gun Cr
also works. The Cr rods can break easily, so it is better to place two rods in parallel
in the source. Evaporate at 3 A/sec (other parameters are posted on the SC4500s).
77
Table 3.6: Cr lift off.
(1) Acetone soak until Cr disintegrates (∼30 sec)
(2) Flush with Acetone
(3) Transfer to new Acetone bath
spraying with Acetone in transit
(4) Sonicate for 45 sec
(5) Flush with Acetone
(6) Transfer to IPA bath
spraying with Acetone in transit
(7) Flush with IPA
(8) N2 blow dry
For lift off, follow the steps in Table 3.6. Note that for all Acetone bath soaks,
it is important never to let the Acetone dry on the wafer. Acetone will leave a
residue which may affect the yield or even the vacuum compatibility of the wafer.
To remedy this, always provide fresh liquid Acetone to the wafer surface (via a
spray bottle) when transferring to a new bath. IPA is a solvent for the Acetone,
so getting some Acetone in the IPA bath will not contaminate the IPA, especially
since it will only be in trace amounts.
C etch in O2 plasma (RIEx)
This step etches off the C everywhere from the wafer except where the e-beam
defined Cr dots are. These C dots will serve as the ion mill mask that defines the
pillar. It is believed that this is an anisotropic etch (recipe in Table 3.7). Even
when we attempted to isotropically etch C in a C/Si bilayer to achieve undercuts
78
Table 3.7: Etch C in an O2 plasma on the RIEx tool.
(1) Clean chamber with O2 plasma for 10 min beyond a
plasma color change to greenish-yellow
(Power = 100W, pressure = 30 mTorr, flow = 30 sccm O2)
(2) Etch the wafer in an O2 plasma until 10 sec beyond a
plasma color change to greenish-yellow
(Power = 90W, pressure = 30 mTorr, flow = 30 sccm O2)
in the C, there was never any evidence that this was ever achieved.
Photolithography
The six different photolithography (PL) steps and their respective resist layers are
shown in Table 3.8. Since the C mask is still exposed for the PL(1) and PL(2)
steps, do not do the Acetone/IPA clean. When doing the Acetone/IPA spray
clean for PL(3) – PL(6), spray first with Acetone for ∼5 sec then start the IPA
spray with the Acetone still spraying, then after ∼3 sec stop the Acetone spray.
Continue to spray IPA for ∼7 sec longer. The spinning will dry the wafer surface.
For PL(3) – P(6), the adhesion promoter P20 is used because Shipley resists do
not adhere well to oxide. P20 spins on only a few monolayers thick and does not
affect the exposure, development, or descum times. The resist spinning recipes for
are summarized in Table 3.9.
All photolith development should be done in a glass or plastic beaker filled to∼1
cm with 300MIF, a tetramethylammonium hydroxide (TMAH)-based developer.
Although the CNF staff will claim that the Hamatech automatic developers work
just as well, we have found on several occasions that the Hamatech development
79
Table 3.8: Resists for the different photolithography steps. PL(1) and PL(2)are both ion milling steps and so it is best to use the thicker S1827 resist whichstrips easier after cross-linking in the ion mill. All photolithography done on topof PECVD oxide should use the P20 adhesion promotor spun before the actualresist. PL(5) and PL(6) ion mill the PECVD oxide using very narrow photolithfeatures, the smallest being a 5 µm line. Use the thinner S1813 resist to reduceshadowing of the ion beam.
Photolithography Step Resist
(1) “Define Leads” S1827
(2) “Define Pillar” S1827
(3) “HF Etch” P20/(S1813 or S1827)
(4) “Protect Shorts” P20/S1827
(5) “Oxide Window” P20/S1813
(6) “Top Leads” P20/S1813
was incomplete. This lack of reliable development is a serious problem and so it is
best to use a 300MIF-filled beaker. Exposure, development and descum recipes are
summarized in Table 3.10. For the descum, first clean the chamber as described
in Table 3.7.
Shipley 1813: limitations
In the nanopillar process, photolithography is used to pattern large scale fea-
tures such as leads and bonding pads. Shipley 1813 (S1813) photoresist has been
the resist of choice because of its ease of use and handling and also because of
its versatility as a pattern-transfer layer in both etching and lift off deposition
processes. However, limitations with S1813 become significant for certain mask
layer procedures. For example, photoresist used as etch masks during ion milling
steps can become cross-linked from exposure to the ion beam, making it almost
80
Table 3.9: Photoresist spinning. All pipettes should be prepped as they are fore-beam resist spinning.
Clean Wafer
(1) PL(1) & PL(2): N2 spray clean wafer while spinning.
PL(3) – PL(6): Acetone/IPA spray clean wafer while spinning.
Spin Adhesion Promoter (PL(3) - PL(6) only)
(2) Dispense P20 with pipette onto wafer so that it coats entire surface.
Let it sit for 20 sec.
(3) Spin P20 with 3000 rpm program.
Spin Photoresist
(4) Siphon nearly a full pipette worth of photoresist from the top of fluid.
(5) Spin resist (both S1827 and S1813) with 2000 rpm program:
S1813 ∼2 µm thick; S1827 ∼4 µm thick.
Solvent Bake
(6) 115C hotplate: S1813 60 sec; S1827 90 sec.
(7) Chill block for 10 sec.
81
Table 3.10: Exposure, development and descum numbers for both S1813 andS1827. The S1813 is developed for longer only because the photolith features forPL(5) and PL(6) are very small.
S1813 PL(5) & PL(6)
Exposure time 1 sec (5x-stepper)
Development 120 sec with agitation. Flush with DI water. N2 blow dry.
Descum 15 sec in O2 plasma
S1827 PL(1) – PL(4)
Exposure time 1.5 sec (5x-stepper)
Development 90 sec with agitation. Flush with DI water. N2 blow dry.
Descum 20 sec in O2 plasma
82
impervious to the typical Acetone solvent. Extended sonication time in a solvent
bath does not completely remove all cross-linked photoresist from the patterned
features, and the unwanted resist can reduce device yield.
Shipley 1827
In order to overcome the problems of ion beam-induced resist cross-linking, a
simple approach would be to increase the thickness of the resist layer. Having
a thicker resist layer would increase the resistance to cross-linking from the ion
bombardment simply because there are more polymeric chains available to absorb
the ion energy. In effect, a thicker resist layer would transform into a bilayer, a top
crust of cross-linked resist atop a normal resist layer that is susceptible to normal
solvation, after ion milling.
This reasoning is somewhat naive at the outset since the entire resist feature,
including its sides, will be bombarded by ions, since milling is typically done at
a finite tile angle φ. Instead of a simple bilayer, one is left with a shell of cross-
linked resist surrounding the normal resist material. It turns out, however, that
sonication is much more effective in stripping this thicker resist mill mask. S1813
is typically spun on at 2000 rpm, yielding a resist layer ∼2 µm thick. A higher
concentration, but chemically identical resist, Shipley 1827 (S1827) spins on at ∼4
µm thick. As such, S1827 is a better choice as an ion mill mask than S1813.
Although using a thicker photoresist layer does, in principle, reduce the res-
olution of the photolithography, the smallest resist features for nanopillar wafer
processing are 5µm (lateral dimension) lines, which is far above the resolution
limit for the GCA-6300 5× (g-line) projection lithography tool at the CNF. The
process parameters in this section have been found to produce successful features
when photolithography is performed with S1827 spun onto metals, thermal SiO2,
83
and PECVD oxide, and are good even when the adhesion promotor P20 is pre-spun
before the S1827.
Storage, pipette dispensing, and stripping for S1827 are the same as for S1813.
The process parameters for this section are summarized in tables 3.9 and 3.10.
As the photolithography used in the nanopillar fabrication recipe does not push
the resolution limits of CNF photolithography tools, the Table 3.10 represents a
working set of process numbers that lie within a range of acceptable values.
Ion milling
There are two ion milling tools available for nanopillar fabrication: the Veeco (at
the CNF) and the IBD (in Clark Hall, room D12). Proper choice of ion milling
tool for the different fabrication steps is shown in Table 3.11. The IBD is a much
more versatile tool with stage cooling down to Tstage = -20.0C, etching in either
an Ar or Ar + O2 ambient, and in situ depositing metals and/or oxides. The most
recent etch calibrations along with all the relevant parameters should be recorded
in the IBD logbook.
Ion Milling (IM) steps IM(1) “Isolate Devices” and IM(2) “Define Pillar” are
done in the IBD because of the stage cooling. The wafers at these two ion milling
steps have 10 µm× 200 µm resist lines and 80 µm resist squares which are harder to
remove in an Acetone bath upon cross-linking from exposure to the ion beam. The
effects of cross-linking can be mitigated somewhat by etching on a cooled stage.
Etch rates of nearly all of the materials in the nanopillar process are kept up-to-
date for the IBD. IM(6) “Clean Top Contact” is also done in the IBD because top
leads will be deposited in situ directly after the etch. The etch source parameters
are shown in Table 3.12 and gas parameters in Table 3.13 for standard ion milling.
84
Table 3.11: Different ion mill steps should be done in different tools. The Veecotends to have higher etching rates, while the IBD has the advantage of stage coolingand has most of the etch calibrations for metallic layers.
Ion Mill Step Tool to Use
(1) “Isolate Devices” IBD
(2) “Define Pillar” IBD
(3) “Planarize SiO2” Veeco
(4) “Thin Oxide Above Pillar” Veeco
(5) “Open Pillar” Veeco
(6) “Clean Top Contact” IBD
Table 3.12: Use these parameters for standard ion milling in the IBD. Applyheat sink compound to the wafer back when mounting on the stage.
Parameter Symbol Value
Beam current Ibeam 70 mA
Beam voltage Vbeam 200 V
Accelerator voltage Vaccel 200 V
Discharge voltage Vdis 40.0 V
Neutralizer current Ineut Ibeam (with shutter open)
Base pressure Pbase 1 – 8 × 10−7 Torr
Stage temperature Tstage -10C
Stage tilt angle φ 135 (= 45 from beam)
Stage rotation ROT ON
Shutter duty cycle duty 30 sec OPEN/30 sec CLOSED
85
Table 3.13: Use these parameters for standard ion milling in the IBD. PCM isthe pressure read by the capacitance manometer showing ambient pressure duringthe mill.
Ar only
FLO2 (Ar) 0.64 sccm
SET2 0.96 sccm
PCM 0.16 – 0.19 mTorr
Ar + O2
FLO1 (Ar + O2) 1.92 sccm
SET1 5.96 – 7.12 sccm
FLO2 (Ar) 0.64 sccm
SET2 0.96 sccm
PCM 0.27 – 0.29 mTorr
86
In the IBD, there are two known instabilities with the etching power supply
(Advanced Energy) that cause jumps in Ibeam, leading to erratic etching rates and
unpredictable etch depths. The first has been attributed to the power supply taking
a finite time to warm up once turned on, while the second and more significant
one has been attributed to a problem occurring only when the source is turned on.
To avoid the first problem, turn on the power supply the same time the gas is
turned on because it takes several minutes for the pressure to stabilize at the PCM
listed above (∼2 min) giving the power supply enough time to warm sufficiently.
For the second problem, once the pressure has stabilized and the source is on, run
the source for 8 – 10 min with the shutter closed. At some point, Ibeam will jump
up to ∼85 mA. At this time, adjust the cathode current knob to reestablish Ibeam
= 70 mA and the source will be stable for the duration of the etch. Ibeam will be
differen if the shutter is closed or open. For 70 mA etching currents should be
approximately Ibeam = 72 mA and Ineut = 60 mA when the shutter is closed. They
will both stabilize very rapidly at 70 mA once the shutter is opened.
IM(3) “Planarize SiO2”, IM(4) “Thin Oxide Above Pillar”, and IM(5) “Open
Pillar” should be done in the Veeco. Parameters for Veeco etching are in Table 3.14.
For these steps, cross-linked photoresist is not a significant concern either because
it is not present (as in IM(3)) or there are no small resist features, just sheets of
resist with holes exposed in them (as in IM(4) and IM(5)). Planarization takes
much less time in the Veeco because of the higher etch rates. IM(4) and IM(5)
both etch PECVD oxide under identical conditions, and the etch rates for IM(5)
should be calibrated from the etching in IM(4) for each wafer since stoichiometric
differences in the PECVD oxide, which can drift from wafer to wafer, will result
in slight variations in the mill rate. A summary of different nominal etching rates
87
Table 3.14: Use these parameters for standard ion milling in the Veeco. Use heatsink compound as done in the IBD.
Parameter Symbol Value
Beam voltage Vbeam 500 V
Accelerator voltage Vaccel 200 V
Discharge voltage Vdis 40.0 V
Neutralizer current Ineut Ibeam (with shutter open)
Base pressure Pbase 1 – 5 × 10−6 Torr
Ar pressure PCM 0.12 mTorr
Stage rotation ROT ON
Shutter duty cycle duty 20 sec OPEN/40 sec CLOSED
for PECVD oxide etching in the Veeco is given in Table 3.15.
When doing the planarization step, use the 85 bevelled edge and a level to get
correct stage tilt φ. Secure the wafer to the stage only with heat sink compound.
Do not use clips, as they will significantly shadow the wafer at such large φ. Even
when levelled, the stage tilt will change slightly when pressing the wafer onto the
stage with the heat sink compound. Although this can be corrected for somewhat
by re-levelling, the wafer will shift once the chamber is pumped out as gas trapped
between the heat sink compound and the wafer escapes and pushes the wafer
slightly. This is generally not correctable, so it is best to put only a small amount
of heat sink compound near the center of the wafer, and not near the edges, which
may prevent gas traps. It is because of these fine shifts in the tilt angle, coupled
with the rapidly varying etch rate at angles close to 90, that wafers etched for equal
amounts of time can show wide variation in the oxide height after planarization.
88
Table 3.15: Approximate PECVD oxide etch rates. φ = stage tilt.
Step (φ, Ibeam) Planar Rate [A/sec]
IM(3) “Planarize SiO2” (85, 60 mA) 0.5 - 0.8
IM(4) “Thin Oxide Above Pillar” (20, 60 mA) 4.5
IM(5) “Open Pillar” same as IM(4) —
other (0, 40 mA) 2.9
“ ” (25, 40 mA) 4.2
“ ” (25, 60 mA) 4.5
“ ” (45, 40 mA) 5.4
“ ” (85, 40 mA) 0.4
Thick PECVD oxide (4× the pillar height as measured through profilometry of
the cap-to-bottom-lead etch height) is deposited to account for this variation.
Ion mill planarization works because at such a high incidence angle (φ ≈ 85)
the inwards etch rate of the oxide pillar sidewalls is larger than the planar etch
rate of the oxide. Planarization with this technique is limited to smaller features
(∼200 nm lateral width) because it thins the oxide protrusions by etching mostly
sideways, so an 80 µm wide bonding pad is too wide to etch in this fashion.
Although ion mill planarization does flatten the oxide pillars, it simultaneously
renders the surface quite rough, pitting the surface with ∼10 nm deep craters with
∼100 nm periodicity (Figure 3.62). The limit of oxide planarization is 10 – 15 nm,
approximately the scale of the roughness from pitting.
For IM(4) and IM(5), in spite of what is shown in Table 3.15, I would suggest
doing these etches at φ = 0. This is because the milled holes are only 5 – 10
µm wide, and a 2 µm tall resist layer could shadow a region of nearly 1 µm along
89
the perimeter of these holes at φ = 20. Shadowing issues will make it difficult to
properly estimate an etch rate.
Ion milling criteria
By far the most critical steps of the nanopillar fabrication are the ion milling
steps. These etches define the bottom lead and bonding pad geometry, define the
nanopillar, and produce the top contact to the nanopillar. The criteria for each
step are different and I enumerate them here. Note that IM(1) and IM(2) would
be greatly simplified if the IBD had attached a secondary ion mass spectrometer
(SIMS) in order to analyze the etched layers in realtime (e.g. see [4]), as opposed
to relying on etch rates that require updating.
IM(1) – “Isolate Devices”:
For device isolation, you must etch into the substrate SiO2. The time required
to etch through all of the metals (tthrough) can be calculated providing all of the
etch calibrations are known. It is important to over etch in IM(1) since under
etching would fail to isolate the devices. Consequently, tthrough should be viewed
as a minimum etch time. There are two crude ways of getting direct feedback
during the etch to see if all metallic layers have been etched. The first is to try and
corroborate the elapsed etch time with the actual layer being etched, which you
can observe when looking through the chamber view ports. It is easier to see the
wafer color when the shutter is open. Identifying the etched material is really only
possible for Au and Cu layers since these are the only two metals with distinctive
colors, but it is a way of getting a mid-etch reference point.
The second is to identify the substrate SiO2 visually through the chamber view
ports. To do this effectively, etch for a total time tthrough, watching closely the
90
substrate color each time the shutter is open slightly before and after the calculated
tthrough. The substrate SiO2, because it is a piece of glass, appears different colors
depending on its thickness and the viewing angle. You can estimate the color
of the oxide from the viewing angle of the view port by holding a wafer (with a
similar thickness thermal SiO2) in a similar geometry. SiO2 etches fairly slowly in
the IBD, so color change due to thinning of the SiO2 is minimal.
IM(2) – “Define Pillar”:
This etch is entirely reliant upon the etch rate calibrations. Each pillar has
different etch requirements. For most nanopillar experiments, the pillar should
have both magnetic layers patterned. Due to the shadowing of the ion beam from
the pillar, there is an inevitable taper that occurs at the base of the etch. The
details of this taper and how it depends on the mill angle are not well known, so
it is a safe precaution to etch into the bottom lead so that the taper is not in
any magnetic layer. Etch into the bottom lead a depth equivalent to 1.5 × the
thickness of all the material above the bottom lead (magnetic multilayer and cap).
IM(3) – “Planarize SiO2”:
To planarize the SiO2 etch for a time equal to the pillar height ÷ 0.5 A/sec,
the approximate planar etch rate of the PECVD oxide. Pillar height should be
measured by AFM after pillar definition (IM(2)) at the central and extent die to
get an idea of the variation in pillar height. Measure the planarization of the pillars
with the AFM. They should be planarized to < 15 nm before continuing.
IM(4) – “Thin Oxide Above Pillar”:
Due to small angular deviations from a φ = 85 planarization angle from IM(3),
there may not be a consistent height of oxide above the top of the pillar from wafer
to wafer. This etch is carried out to thin the amount of oxide above the nanopillar
91
to a consistent value of 75 nm above the capping layer (Pt or Au) although ±25
nm is acceptable. Measure the etch afterwards and this will calibrate the PECVD
oxide etch rate for this wafer which, due to stoichiometric differences in the oxides,
could be different from wafer to wafer.
IM(5) – “Open Pillar”:
At this point the wafer should be scribed into chips. This etch opens up with
the pillar to allow for top electrical contact. Calculate the etch time using the etch
rate estimated from IM(4) and an oxide height equal to the total amount of oxide
above the capping later (Pt or Au) plus the planarization height. Feedback for
each iteration of the etch will come from the DC characterization of the devices
after the tops leads are deposited, where it will hopefully be clear if the etch time
was too long (i.e. mostly shorts, weak magnetic behavior) or too short (i.e. mostly
opens, pillar still covered in oxide). Usually the data are not nearly this obvious,
but average behavior of many samples can be a valuable metric and should give
feedback to help you hone in on the correct IM(5) etch time.
IM(6) – “Clean Top Contact”:
6 – 15 sec ion mill cleaning etch before top lead deposition.
Stripping resist
The procedure for stripping resist is shown in Table 3.16. The choice of Vbeam =
200 V for the IBD etches is a recent change and so resist stripping may take only
a few minutes as opposed to a few hours since the beam energy is lower.
92
Table 3.16: Strip photoresist that has been exposed to ion milling. This resiststripping recipe was first written for more energetic ion milling (Vbeam = 500V).Using both S1827 and new beam parameters for the IBD in Table 3.12 above,cross-linked resist may strip in less time.
(1) Acetone soak with intermittent agitation until large resist
flakes peel off (several hours)
(2) Transfer to new Acetone bath spraying with Acetone in transit
(3) Repeat Acetone soaking until resist is gone from unpatterned
wafer edges
(4) Sonicate for ∼15 sec
(5) Flush with Acetone (repeat (3)-(5) if necessary)
(6) Transfer to IPA bath spraying with Acetone in transit
(7) Flush with IPA
(8) N2 blow dry
93
Measuring etch depths
Measure depth of etches using the P10 profilometer at the CNF or the Alpha-step
500 profilometer in TOL. The most important profilometry measurements are those
that measure the height of the oxide above the nanopillar. This is easiest to do
just after stripping the resist from the PL(3): “HF Etch” step. The scans should
be taken across one bonding pad and over the device area. Take the difference
between the height of the oxide and the height of the bonding pad which, for the
C mask fabrication process, should be the same height as the top of the nanopillar
(not including the C mask).
It is best to measure this over many die, and if patience permits, over a few
devices per die. This is what is known as “wafer mapping”. One can measure the
height of the oxide over the pillar after the evaporation and liftoff of extra oxide
from the PL(4): “Protect Shorts” step, but is more difficult since now you have to
aim the tip so that it hits a 10 µm square. However, it is important to do a few
profilometry scans after the PL(4) lift off just to make sure that the resist above
the nanopillar did, in fact, lift off.
50 A SiO2 deposition
This PECVD step contains oxygen in the plasma which will likely etch the C
mask and so capping the pillar with 50 A of oxide protects the C. This deposition
can be done in the SC4500 EVEN hours evaporator or in the IBD. It is generally
recommended to deposit this 50 A of oxide in the CNF evaporator since four
3” wafers or three 4” wafers can be done simultaneously, whereas the IBD is
single wafer only. Beam parameters for SiOx deposition in the IBD are shown in
Table 3.17 and gas parameters in Table 3.18. Any oxide deposited in the IBD must
94
Table 3.17: Use these parameters for deposition of any material in the IBD.Rates are very stable and can be found in the IBD logbook.
Parameter Symbol Value
Beam current Ibeam 100 mA
Beam voltage Vbeam 500 V
Accelerator voltage Vaccel 200 V
Discharge voltage Vdis 40.0 V
Neutralizer current Ineut Ibeam (with shutter open)
Base pressure Pbase 1 – 8 × 10−7 Torr
Stage temperature Tstage +20C
Stage tilt angle φ 180 (∼20 from line-of-sight to target)
Stage rotation ROT ON
Shutter duty cycle duty always OPEN
be deposited with an Ar + O2 ambient. Pre-clean the IBD target by running the
deposition source for 5 min with the shutter closed before SiOx deposition.
PECVD oxide deposition
Although protocol for the PECVD tool is to perform a chamber clean for (dep time)
+ 10 min after every run, not everyone follows this rule. Check the deposition time
on the last process run. It must be > 10 min. If not, it is better to be safe and
clean the chamber for 1 hour to be certain the chamber is clean. Even if the last
run time is > 10 min, still perform a 10 min chamber clean. Next, before loading
the sample, do a 2 min seasoning with glass slides inside the chamber. Seasoning
is a brief deposition of the PECVD material to be coated on your wafer (oxide in
95
Table 3.18: Use these gas parameters for deposition in the IBD. PCM is theambient gas pressure read by the capacitance manometer. Deposition of any oxidein the IBD should always be done with the Ar + O2 recipe. Oxygen deficiency inthe film, which makes the film dark, glassy, and mechanically unstable, will resultif the oxide is deposited only in Ar.
Ar only
FLO3 (Ar) 2.24 sccm
SET3 2.32 sccm
PCM 0.48 mTorr
Ar + O2
FLO1 (Ar + O2) 1.92 sccm
SET1 5.96 – 7.12 sccm
FLO3 (Ar) 2.24 sccm
SET3 2.32 sccm
PCM 0.60 mTorr
96
this case). This has the effect of priming the gas lines as well coating the glass
slides that will hold your wafer in place and is believed to yield a better quality
film on the wafers. Include this 2 min seasoning time in the total clean time at
the end of your run.
For nanopillar wafers, deposit an oxide film at least 4× the height of the nanopil-
lar. This leaves enough oxide to prevent over etching during planarization. Deposit
oxide using “Process 1” with platen temperature 170C. Deposition parameters are
posted on the tool. Deposition rates should be recalibrated every few months. The
currently accepted rate for PECVD oxide at 170C is 32.0 nm/min.
HF etching
Use BOE (6:1) etch to clear the oxide above the bonding pads. PECVD oxide
etch rate in this solution is 7 nm/sec. Calculate an HF etch time = 2 × (height
of deposited PECVD oxide thickness)/(7 nm/sec). Sometimes there may appear
a slush-like sludge in the HF solution. My guess is that this is an indication of
a contaminated batch of HF, although the etch rates do not seem to be affected.
However, since over etching in this step poses no risk to the devices it is safe to
increase the etch time by another factor of 2.
3.3 The ion beam deposition system
The Ion Beam Deposition tool (IBD), acquired from the IBM T. J. Watson Fa-
cility at Yorktown Heights, NY, has both ion milling and ion beam deposition
capabilities, both of which I discuss in this section. The AJA sputter system is the
predominantly used metal-deposition tool because of its versatility (up to seven
materials deposited per bake out) and also because of the high purity of deposited
97
material. The latter is due to a low base pressure (∼ 10−8) and the tight con-
finement of the plasma within the sputter guns, which minimizes the amount of
cross-contamination coming from incidental etching and possible redeposition of
materials coating the chamber walls. The IBD system, in contrast, has two ion
beam sources which are, by definition, un-confined. Any coatings on the chamber
walls are then susceptible to being etched by these ion beams and may redeposit
onto a sample.
3.3.1 Deposition
Redeposition of chamber-coating materials does contaminate IBD-deposited films.
Two pieces of data that demonstrate this are a Rutherford backscattering (RBS)
analysis on an IBD-deposited AlOx film and the resistance as a function of temper-
ature of an IBD-deposited Cu film, shown in Figures 3.29 and 3.30, respectively.
The presence of trace amounts (∼3 at.%) of magnetic impurities in the RBS spectra
for AlOx is in complete agreement with the Kondo minimum observed in the resis-
tance versus temperature of IBD-deposited Cu. In Figure 3.30 the IBD-deposited
Cu resistance versus temperature is shown alongside that for AJA-deposited Cu,
which shows no Kondo minimum, highlighting the difference in contamination lev-
els of both systems.
Although magnetic alloys were deposited during its use at IBM, no magnetic
material has been deposited in the IBD in three years of continuous operation
at Cornell. It is more likely that the magnetic impurities in the deposited films
come from redeposition of a magnetic coating on the chamber walls which, itself,
originates from magnetic multilayer wafers ion milled in the same chamber. Fe,
Mn, and Co have all been etched in this chamber.
98
Adhesion layers (Ti), top leads (Cu), and dielectrics (SiO2 and Al2O3) are
the most commonly deposited materials. XPS measurements of AlOx films have
found identified as stoichiometric alumina (Al2O3) when deposited in Ar + O2 [5].
The electronic insulation properties of IBD-deposited Al2O3 films have yet to be
characterized.
3.3.2 Ion milling
A recirculating chiller was purchased for the IBD stage that could cool a 1:1 HPLC
Water:Ethylene Glycol mixture down to -20C, although it took nearly two hours
to do so. Consequently, an operating temperature of Tstage = -10C, which takes
less than one hour to achieve, was chosen. For the same beam parameters (Vbeam
= 500 V, Vaccel = 200 V, Vdis = 40.0 V, and Ibeam = 60 mA) etch rates in the IBD
were found to be approximately half those of the Veeco, due mostly to a difference
in current densities of the two sources (Veeco source area = π4(10cm)2, IBD source
area = π4(12cm)2, areal ratio ≈ 0.69). These parameters gave significant non-
uniformities in the IBD etch profile across a wafer. Changing the beam parameters
to Vbeam = 200 V, Vaccel = 200 V, Vdis = 40.0 V, and Ibeam = 70 mA does show
better etch profiles (Figure 3.31).
99
0
50
100
150
200
Counts
0.4 0.6 0.8 1.0 1.2 1.4
Energy (MeV)
O
Al
Si
Ca
FeMnCoMo
W
Figure 3.29: Rutherford backscattering (RBS) spectra of an IBD-deposited AlOxfilm. The signature of SiO2 (substrate) and AlOx are apparent in energies < 900keV. Heavier element contaminants are found at higher energies. Ca (∼4 at.%)likely comes from the target; FeMnCo (∼3 at.%), which are all indistinguishablefrom one another with the RBS technique, are likely due to chamber contaminantsredeposited on the sample that originated from magnetic material ion milled fromdevice wafers; Mo (∼0.3 at.%) and W (∼0.3 at.%) likely come from the etchingsource directly since the focusing grid is Mo and the neutralizer wire is W.
100
0 50 100 150 200 250 3007.47.57.67.77.87.98.08.18.28.3
404550556065707580
IBD top leads
Top
Lead
Res
ista
nce
[
Temperature [K]
AJA bottom leads
Bottom
Lead Resistance [
Figure 3.30: Device leads are measured as a function of temperature. The bottomleads are a thick Cu layer etched from a nanopillar multilayer stack deposited inthe AJA while the top leads are IBD-deposited Cu. Resistance values dependon film thickness and geometry. However, the minimum in the IBD-depositedCu resistance is the result of magnetic impurities, a Kondo minimum. Such aresistance minimum is not present in the AJA films owing to a much cleanerdeposition system.
101
0 10 20 30 400.75
0.80
0.85
0.90
0.95
1.00
N
orm
aliz
ed E
tch
Rat
e
Distance From Stage Center [mm]
Figure 3.31: IBD etch uniformity as a function of position away from stage centerfor different beam parameters: () Vbeam = 200 V, Ibeam = 70 mA and (•) Vbeam
= 500 V, Ibeam = 60 mA. For both, Vaccel = 200 V and Vdis = 40.0 V. The verticalline at the 32 mm stage position indicates where the most extent device lies ona wafer. Etch uniformity of the lower Vbeam () is 5% within the device extent,where it gets to worse than 10% for the higher Vbeam (•).
102
3.4 Aligned electron-beam lithography
3.4.1 Overview
This section discusses the method of aligned exposures in the Leica VB6 e-beam
lithography system. There are two levels of alignment accuracy, each requir-
ing different amounts of preparation to achieve. Global, or coarse alignment,
which can expose a point to within ∼1 µm of a desired location on the wafer,
requires one set of photolithography-defined alignment marks. Fine alignment can
achieve an accuracy of 10 nm but requires, in addition to the photolithography
marks for global alignment, e-beam lithography-defined alignment marks within
each exposure field. A mid-level alignment of 40 nm accuracy is also possible if
the e-beam lithography-defined fine alignment marks are instead formed by pho-
tolithography. The poorer accuracy is due to quality differences in the e-beam
versus photolithography-defined alignment marks. The procedure for definition of
mid-level alignment marks is described elsewhere [5].
The defining operational difference between global and fine alignment is whether
or not a stage move is required after an alignment procedure. Maximum deflection
of the electron beam allows for a maximum 327.68 µm square exposure field with-
out requiring a stage move. If the VB6 can align to high quality marks within this
exposure field, this defines a fine alignment as no stage motion was required after
the alignment. For global alignment, the VB6 aligns to marks positioned across
the entire wafer, to which a new coordinate system is mapped and all subsequent
exposures take place within this coordinate system with no further alignment.
Achieving fine alignment starts by mapping the global wafer coordinate system as
a reference as it moves the stage to each expected set of fine alignment marks and
103
performs a local alignment for each exposure field.
3.4.2 Alignment marks
In general, features with good contrast in an SEM may be useful as an alignment
mark. This includes features with topographical and/or material (i.e. atomic
number Z) contrast with the substrate. Depending on the application, the user
has a wide selection of material choices for alignment marks. Global alignment is
performed once per wafer and is usually done manually by positioning the stage
through keyboard input while monitoring mark position on the video monitor. Fine
alignment, however, is performed automatically by the VB6 for each exposure field.
Consequently, the guidelines for alignment mark definition are much more strict
for fine alignment than they are for global alignment.
For global alignment marks, material combinations confirmed to have worked
include KOH-etched windows in Si underneath continuous Si3N4 membranes, trenches
in a multilayer stack capped with 10 A of naturally oxidized Al and with thermal
SiO2 at the trench bottom, and lithographically defined Au on top of a multilayer
stack. Etched holes in Si are also claimed to work [6]. These examples span a
broad array of materials (metals, insulators, and semiconductors) highlighting the
relative simplicity of global alignment. Mark size can be rectangles with sides of
length 1 to 120 µm.
Global alignment accuracy is ultimately limited by the motion of the stage,
but there is also a contribution from the location of the marks themselves and
proper choice of alignment mark location can improve alignment. If the marks are
placed at the corners of a square, as opposed to, say, at the middle of the edges
of the same square, it means that when it comes time to find the marks in the
104
VB6, both x and y stage translation is required, and not just mostly x or mostly
y, thereby removing problems of mechanical hysteresis within the stage motion.
Also, global marks that are spaced further apart give better alignment due to a
smaller percentage error in their position relative to each other.
Automation of the fine alignment works by having the VB6 find the alignment
mark by scanning the electron beam across its edge while monitoring the intensity
of backscattered electrons. This requires a proper set of criteria that define what an
edge is to the VB6. Having good topographical and/or Z contrast of the alignment
mark compared to its surrounding substrate will help the VB6 identify the edge.
But it is the edge roughness and placement of the alignment marks, both of which
are completely dependent on the choice of lithography used to define them, that are
the two most important factors in achieving 10 nm alignment accuracy. Smooth
edges are achieved one of two ways. First is by evaporating the mark material into
a lithographically-defined hole where the resist has an undercut profile to prevent
undesirable edge effects such as fencing (Figures 3.32 and 3.33). Second is to
anisotropically etch a film of the mark material with a separate pattern transfer
layer that can be removed at a later stage (Figures 3.34 and 3.35).
E-beam lithography provides smoother edges and more precise placement than
photolithography and must be used to achieve the best alignment results. Fine
alignment marks must come in groups of four, all of which must fit inside one
exposure field (maximum 327.68 µm square), with each mark being a square 10
µm on a side. The marks should lie within a 40 µm clearance region, which is a
boundary around the marks that will be protected from all processing so as not
to introduce any topographical contrast, giving a uniform background from which
the VB6 can easily locate the mark edge.
105
(a)
(b)
Figure 3.32: Fencing. (a) Deposition into a resist feature with no undercut. (b)Resulting profile is elevated near the edges. This is called “fencing” and is a poorchoice for mark definition for automated e-beam alignment.
106
(a)
(b)
Figure 3.33: Alignment mark by lift off. (a) Deposition into a resist featurewith an undercut layer. (b) Resulting profile is smooth near the edges, making itpreferable for automated e-beam alignment.
107
(a)
(b)
Figure 3.34: Alignment mark by pattern transfer (1/2). (a) Begin with an e-beam defined Cr mask above the initially unpatterned C and Au. (b) Pattern theC with an anisotropic O2 plasma etch.
108
(a)
(b)
Figure 3.35: Alignment mark by pattern transfer (2/2). (a) Ion mill the Auusing C as a mask. (b) Remove the C with an O2 plasma.
109
3.4.3 Global alignment
All alignment procedures begin by manually locating the global alignment marks
on the wafer using the microscope in the VB6 room. With the resist spun, place
the wafer on the appropriate holder with the flat abutted against the two posts on
the wafer platter. Screw down the charge dissipation Cu tabs and then mount the
holder onto the optical microscope bed. Tightening down the Cu tabs will shift
the wafer slightly, so be sure to do that first before aligning to the holder under
the microscope. The holder and the microscope bed lie in the same rectilinear
coordinate system, so the wafer flat can be aligned to the holder through the
process of translating the microscope bed left to right while looking at the wafer
flat under the microscope. If the wafer flat is not parallel to the bed motion, rotate
the wafer until it is. Proper rotation alignment of the wafer to the holder greatly
simplifies alignment to the wafer once inside the VB6.
Next, move the microscope bed so that the holder-mounted Si chip is under the
view of the microscope. There are many Au markings on this chip, but find the
ones that look that those shown in Figure 3.36. These markings are the same for
both the 3” and 4” wafer holder. Locate circle1, as indicated in Figure 3.36. Make
sure the optical microscope is at its highest magnification and center circle1 in the
microscope cross-hairs. Zero the x and y coordinates of the bed (to the right of the
microscope). Then, in sequence, move to each of the four global marks, recording
the absolute coordinates of each ((xi, yi), i = 1 − 4) making sure that they are
centered in the cross-hairs and the microscope is at the highest magnification.
The x-y coordinates of the bed are in the VB6 coordinate system. I have found it
helpful to record these absolute coordinates on a sketch of the wafer which includes
the wafer flat, the general x-y axes of the VB6, and the mark coordinates relative
110
Figure 3.36: Locate circle1. These marks are identical on both the 3” and 4”wafer holders. They are Au patterned on Si. The mark from which to start thealignment procedure, circle1, is indicated as it appears under the optical micro-scope. The schematic is not to scale.
to the wafer center ((xi, yi), i = 1 − 4), as shown for an actual alignment in
Figure 3.37.
Being careful not to jar the wafer, load the holder into the load lock and go
through the process of transferring the holder to the chamber and performing the
VB6 calibration (jobcal). After calibration, type at the EMMA prompt
> MVSP CIRCLE1
This moves the stage so that the circle1 pattern is on the video monitor. Next
make sure circle1 is properly positioned on the video monitor, which is not the
monitor center but is rather indicated by a penned in ‘+’ on the screen. To do
this type
> SSEM
111
Figure 3.37: Record mark coordinates for e-beam alignment. Coordinates rel-ative to circle1 are (xi, yi). Coordinates relative to the wafer center are (xi, yi).This diagram is useful to copy down on paper each time an alignment is performed.Note that the coordinate system of the VB6 is 90 off from what would be intuitiveto the user. All coordinates here are written in the VB6 coordinate system.
112
> SMAG 2000
> MVRL .01 .00
The SSEM command changes the VB6 to SEM mode. SMAG changes the magni-
fication. It is best to do the alignment at the same magnification for all markings,
which will typically be 1000 – 2000. The two numbers following the MVRL, or “move
relative” command, are x and y translation of the stage (in mm). Move the stage
via the MVRL command until circle1 is properly positioned. When done, type
> SFAB
to place the VB6 back into fabrication mode. This prevents accidental exposure
when moving the stage to the global alignment marks.
Move to the first absolute mark (x1,y1)
> MVRL -7.7490 73.4645
taking numbers from Figure 3.37. Look for the mark in SEM mode.
> SSEM
If the mark is not immediately visible, reduce the magnification
> SMAG 200
If the mark is still not visible, systematically explore a local region either by
continual MVRL commands or by using the joystick. To use the joystick, click on
the Toolbox menu on the EMMA window, click on Joystick and use the mouse to
move the stage. It is helpful to store the initial stage position so that there is a
consistent reference point to come back to if you feel you are getting lost. To store
a position, type
> SP dummy1
where the name “dummy1” is just a temporary variable to store the reference point
position. To return to this reference point, type
113
> MVSP dummy1
Once the mark is found, adjust its position using the MVRL command until the
mark is aligned under the ‘+’ on the video monitor as circle1 was. Get back into
fabrication mode,
> SFAB
change to the directory to [VB.USERS.EMLEY2000.JOBS], and execute the global
alignment jobfile
> @COARSE ALIGN.COM
First the jobfile asks if circle1 has been found and if the first alignment mark has
been located. Both are done, so type
> CONTINUE
The jobfile then asks for the expected (Ei) and observed (Oi) relative positions
(xi, yi) of three of the four alignment marks where (0,0) is the wafer center. When
prompted for Oi, simply hit return and it will enter the default setting Oi = Ei.
Writing in italics below represents prompts from the program at the command
line. From the numbers in Figure 3.37
> Enter position E1 (x y) -8.000 -8.000
> Enter position O1 (x y) (default will set O1=E1)
> Enter position E2 (x y) -8.000 8.000
> Enter position O2 (x y) (default will set O2=E2)
> Enter position E3 (x y) 8.000 8.000
> Enter position O3 (x y) (default will set O3=E3)
> Move to the center of FIRST MARK. Type continue when done.
Since the first mark has already been aligned, type
> CONTINUE
114
Somewhat redundantly, it prints the following line
> ∗TYPE CONTINUE WHEN LOCATE THE ALIGNMENT MARK, SPO O1 IS NEXT∗and then goes into SEM mode. Type
> CONTINUE
and it stores the position of the first alignment mark O1. The jobfile then moves
the stage relative to this position to that entered for E2. The prompt displays
> ∗TYPE CONTINUE WHEN LOCATE THE ALIGNMENT MARK, SPO O2 IS NEXT∗and then goes into SEM mode. In the same manner as for the first mark, find and
align the second mark under the ‘+’. Type
> CONTINUE
and it stores the position of the second alignment mark O2. This is repeated for
the third mark, thus defining O3. Once the three mark positions O1, O2, and O3
are defined, the program maps the wafer coordinate system to these alignment
marks by the command pair in the jobfile
$ DWCO WAFER /EXP=(E1,E2,E3) /OBS=(O1,O2,O3)
$ DWMO WAFER /LOAD
It is possible now to check if the alignment is good by trying to find the fourth
alignment mark E4. Now that the wafer coordinate system has been mapped to
the alignment marks, use the MVPO command, which moves in absolute coordinates
now defined in the wafer coordinate system.
> MVPO 8.000 -8.000
> SSEM
where (x4, y4) are those from Figure 3.37. If the fourth alignment mark is within
∼20 µm of the ‘+’ on the video monitor, the alignment should be good enough
even for the possible next stage of fine alignment.
115
3.4.4 Fine alignment
Due to the tremendous flexibility in the usage of fine alignment in the VB6, I
will limit the discussion to methods for making aligned exposures of nanopillar
devices. Here I discuss the many important details of fine alignment, including
design of the alignment marks, design of the exposure pattern that will be aligned
to those marks, definition of the alignment marks on the wafer, and execution
of the fine alignment itself once the marks are defined. It turns out that there
are subtle yet important design differences in the mark and pattern when using a
positive tone resist, such as PMMA, versus a negative tone resist, such as hydrogen
silsesquioxane (HSQ).
Design guidelines for fine alignment marks: positive tone resist
During the alignment procedure the mark is exposed as it is being located by the
electron beam. This brings up the consideration that one set of alignment marks
is generally unusable after the alignment step as the next step after development is
usually an evaporation-then-lift off step that leaves the alignment marks partially
coated with the evaporant. However, for negative resists, exposure of the marks
actually encapsulates the mark under a protective layer of the exposed resist,
meaning that they may be used for subsequent alignments.
A fabrication process requiring multiple fine alignments using only PMMA is
limited to how many fine alignment marks can be fit into a single exposure field
without obstructing each other or any lead geometry for that device. The first
VB6 exposure should define both fine alignment marks as well as the device. To
avoid locating the wrong mark, the marks must be separated by at least 30 µm.
Best accuracy is achieved with the marks separated as far apart as possible.
116
Design guidelines for an aligned nanopillar: positive tone resist
Since each set of alignment marks can be used reliably only once, each aligned
nanopillar pattern must align to a new set of alignment marks. A worst case
scenario of multiple alignments is illustrated in Figure 3.38, where three sets of
alignment marks have been drawn on top of asymmetric leads. If the different
sets of alignment marks are arranged so that they lie on the perimeters of boxes
having the same center, it is easier to keep track of the multiple alignments as
they will all share a common origin (Figure 3.39). Likewise, it is much easier to
keep track of the center of the .FRE files if the geometry of the extent alignment
marks is included in the design for each aligned nanopillar. For an asymmetric
device like the one shown in Figure 3.38, the center of the .FRE file is not the
same as the center of the nanopillar. But if all of the aligned .FRE files have their
devices placed at a consistent location, numbers for exposure offset are consistent
from alignment to alignment. CAD for the subsequent exposures of the device of
Figure 3.38 is shown in Figure 3.40.
Design guidelines for fine alignment marks: negative tone resist
Using a negative tone e-beam resist such as HSQ alleviates the problem of requiring
separate alignment marks for each alignment step since the alignment marks can
be protected by exposing the resist around them. However, because PMMA/HSQ
is being considered as a bilayer resist that replaces the C layer as the ion mill mask
in a possible new fabrication process, we must account for limitations coming from
PMMA dissolving in Acetone. For example, if the first level of VB6 exposures
defined a device and alignment marks, there is no way these could be processed
separately since any photolithography would incorporate Acetone soaks that would
117
Figure 3.38: Alignment marks for multiple aligned exposures in positive toneresist. There are three sets of alignment marks, labelled as first (red), second(blue), and third (yellow). All three sets (plus a device pattern, if needed) mustbe exposed at once. Since these exposures will be on positive resist, each set ofmarks will be usable only for one alignment.
118
Figure 3.39: Shared origin of the different sets of alignment marks. The marksare designed such that they all have the same origin. This makes it simpler tocorrectly define the aligned coordinate system for multiple alignments since theywill all be defined with respect to a common origin. Due to the asymmetry in theleads, the device is not exposed at the aligned coordinate system origin.
119
(a)
(b)
Figure 3.40: Exposure pattern for positive resist multiple alignments. (a) Thealigned exposure pattern should always include extent features that encompassesthe device. For this example, the four green crosses, which are exposed over thefirst set of alignment marks, have an extent (dotted line box) which encompassesthe ellipse. The VB6 will expose the geometric center of the .FRE file at the centerof the aligned coordinate system, marked by a ×, which is not the center of theellipse. (b) The CAD should be drawn as it is here because there is a 90 offsetbetween the VB6 coordinate system and that of the VB6 room, as indicated.
120
destroy the exposed region. Unless it is confirmed that the pillar and alignment
marks can be defined in the same ion mill step, it is prudent to define the alignment
marks first (with positive tone PMMA) and then perform fine alignment to define
the pillar (in PMMA/HSQ).
The layout for the alignment marks is slightly different for the negative tone
resist than for the positive tone. This is because one of the goals of the exposure
is to completely expose each alignment mark as well as the 40 µm clearance region
around it. This will protect the alignment marks and the local substrate material
from being patterned, which will preserve the integrity of the alignment mark for
future alignments. Because of this concern, the alignment marks cannot be out at
the edge of the exposure window, but should be offset by ∼1/2 the width of the
clearance region. An example alignment mark layout is shown in Figure 3.41, which
is the one used in the proposed fabrication process of the 3-terminal nanopillar in
Ch.6.
Design guidelines for an aligned nanopillar: negative tone resist
The exposure pattern should include exposures at the clearance regions described
in the previous paragraph. The exposure compatible with the alignment marks of
Figure 3.41 is shown in Figure 3.42.
Define fine alignment marks
Defining fine alignment marks can only take place after global alignment where the
wafer coordinate system has been mapped to the global alignment marks. It is best
to use an e-beam resist bilayer, such as low molecular weight (MW) PMMA/high
MW PMMA [1], where the high MW overlayer (such as 950k PMMA) retains
121
(a)
(b)
Figure 3.41: Alignment marks for multiple aligned exposures in negative toneresist. (a) Asymmetric leads with a set of alignment marks. (b) The layout for thealignment marks includes space for the 40 µm square clearance region around eachmark, placing the entire exposure within a 320 µm exposure field. The alignmentmarks themselves set at the corners of a 280 µm square.
122
(a)
(b)
Figure 3.42: Exposure pattern for negative resist multiple alignments. (a) Theoverlayed pattern should always contain a feature that encompasses the device (i.e.the green ellipse). The four 40 µm green squares expose the clearance regions. (b)The CAD should be drawn as it is here because there is a 90 offset between theVB6 coordinate system and that of the VB6 room, as indicated.
123
the shape of the exposure and the low MW underlayer (such as 495k PMMA) is
overexposed and will develop to form an undercut. The undercut can be used to
form the alignment mark directly, by evaporating into the developed region and
lifting off, as shown in Figure 3.33. It is best to have ∼100 nm of alignment mark
material (Au, or other high-Z metals that are easily evaporated, such as Ta and
Pt), which is achievable provided the bottom undercut layer is thick enough. The
“rule of thumb” is for the deposited material to be at most 1/3 that of the undercut
layer thickness.
Alternatively, a thinner layer may be deposited into the e-beam exposed region
which lies on top of the alignment mark material. The thinner e-beam feature
can then be used to transfer the e-beam pattern into alignment mark material
underlayer by some anisotropic etch, as shown in Figures 3.34 and 3.35. This
method, however, requires more fabrication steps.
Align to existing fine alignment marks
Fine alignment can only be performed when fine alignment marks are defined on the
wafer, as described in the previous paragraphs, and a global alignment procedure
has mapped the wafer coordinate system to global marks. When the automated
fine alignment is executed the VB6 looks for the alignment marks, within a given
exposure field, one at time. For the “edge” locate algorithm (described below), the
VB6 first performs a coarse scan, which rasters the beam horizontally, comparing
signal intensity at consecutive points along the scan, until two parallel edges of
the alignment mark are found. The beam then rasters vertically until the other
two edges are found. At this point a fine search is performed in which the edge
intensity profile is fit with a line, and the intersection of the four lines from the four
124
edges is returned as the center of the alignment mark. This sequence is repeated
for the three other alignment marks.
The fine search is a sequence of line scans across the four edges of the square
alignment marks. Dimensions of the fine search are shown in Figure 3.43. These
parameters are important because they help quantify how the VB6 distinguishes a
mark. An example jobfile designed for fine alignment is [VB.USERS.EMLEY2000.JOBS]
3TERM ALIGN TOP.COM. This is the parent jobfile which configures the stage
height correction, selects the pattern to expose, sets the doses and clock speeds
and defines the positions of the devices within each die. A jobfile FMKMARK.COM is
called which defines the fine alignment focus mark and the search algorithm used
to find them.
Mark definition is set by the lines in FMKMARK.COM
$ QMARK DEL FMK
$ QMARK DEF FMK /DESC=10um square for focus locate
$ QMARK DEF FMK /SIGN=(rt=0.25,ct=0.1,filter=15,type=bright,locate=edge)
$ QMARK DEF FMK /GEOM=(h=10,w=10,mh=3,mw=3,mrkt=0.2)
$ QMARK DEF FMK /SCAN=(ovsam=8,lines=1,cslim=50,para=4,mlen=1.0,sres=1)
The values entered in the QMARK DEF FMK commands here have already been con-
firmed to work for Au alignment marks on thermal SiO2.
QMARK DEL FMK line deletes the pre-existing definition of the focus mark vari-
able name FMK. QMARK DEF FMK /DESC gives a text description of the mark and is
for information purposes only. QMARK DEF FMK /SIGN gives the signal parameters
and defines the algorithm used for determining the position of the alignment mark.
locate=edge assigns the “edge” algorithm to locate the mark, which compares the
rate of change of the backscattered electron signal. type=bright indicates that
125
Figure 3.43: Dimensions in the fine alignment jobfile are defined here. h and w
are the stated dimensions of the alignment marks. They are not necessarily theactual dimensions. How close h and w are to reality depends upon the quality ofthe lithography used to define the marks themselves. E-beam defined marks arethe highest quality. For this schematic, para = 4.
126
the mark is brighter than the background. rt=0.25 is the rise time of the signal (in
µm), defining the width of the alignment mark edge. ct=0.1 defines the minimum
contrast the VB6 should look for to determine edge location. filter=15 is used
to reduce the noise level.
QMARK DEF FMK /GEOM gives the geometry parameters which are shown in Fig-
ure 3.43. h=10,w=10 are the height and width, respectively of the alignment mark
(in µm). mh=3,mw=3 are the dimensions of the sampled regions of the mark edge
(in µm) and mrkt=0.2 is tolerance for dimensions h and w. QMARK DEF FMK /SCAN
gives the scan parameters. ovsam=8 defines the number of times the video level
should be read for averaging. lines=1 sets the number of line scans across the
same position at a mark edge. para=4 is the number of line scan positions within
the mh=3 and mw=3 dimensions. The total number of line scans per edge is lines ×para. cslim=50 is the coarse search limit (in µm), meaning the maximum window
the electron beam searches to find the alignment mark. mlen=1.0 sets the scan
length, which should be less than h and w. sres=1 is the scan resolution which
sets the size of the deflection between adjacent points during the coarse search.
This essentially represents how precise the coarse search should be, with 1 giving
the most precise search.
The jobfile [VB.USERS.EMLEY2000.JOBS] 3X3GRID ALIGN.COM is called
inside 3TERM ALIGN TOP.COM and does the alignment and exposure, but only
within a 3 × 3 array of die at the wafer center. This is merely to reduce the total
VB6 runtime since the fine alignment procedure increases the time per device by
nearly a factor of five from ∼6 sec (for an unaligned exposure) to ∼30 sec. Looking
at 3X3GRID ALIGN.COM, each exposure goes through the following sequence.
First, it moves to the stage position, which is still defined in terms of the global
127
alignment, by the MVPO and MVRL commands. Then a sub routine ”CHIP ALIGN”
is executed. First, the stage height is corrected with the AFLD command. Then
the alignment marks are located by
$ QLOCATE FINEMARK BL /MAIN=(X=-125,Y=-125) /FINE /LOG
$ QLOCATE FINEMARK TL /MAIN=(X=-125,Y=125) /FINE /LOG
$ QLOCATE FINEMARK TR /MAIN=(X=125,Y=125) /FINE /LOG
$ QLOCATE FINEMARK BR /MAIN=(X=125,Y=-125) /FINE/LOG
The /MAIN=(X=±125,Y=±125) numbers represent the expected positions of the
fine alignment marks relative to the center of the exposure field. The QLOCATE
commands execute the search for each mark and define the variables BL, TL, TR
and BR by the observed mark positions. The aligned coordinate system is defined
in terms of these observed mark locations by the commands
$ DWCO CHIP /EXP=(CBL,CTL,CTR,CBR) /OBS=(BL,TL,TR,BR)
$ DWMO CHIP /LOAD
Now that the chip coordinate system has been mapped to the fine alignment marks,
the pattern is exposed with the EPAT command, after which the system returns to
the wafer coordinate system by the DWMO WAFER /LOAD command. This sequence
is repeated for every exposure until the jobfile is finished.
3.4.5 Successful demonstration of fine e-beam alignment
An example of a successful fine alignment with the VB6 and the fabrication steps
taken to achieve it is shown in the following figures. Figure 3.44(a) shows an
SEM image and schematic of the pillar to be aligned. The layer deposited after
successful fine alignment, shown in Figure 3.44(b), shows ∼10 nm overlay accuracy.
Because of this accuracy limit, it is suggested that the patterns exposed during the
128
alignment be smaller than the patterns being aligned to. As such, overlaid devices
will be circumscribed within underlying patterns.
The fabrication procedure for the aligned exposures is described in Figures 3.45
through 3.56. This wafer was fabricated so that the top of alignment mark would
coincide with the top of the sputtered multilayer. It follows that doing this should
give the best exposure integrity as well as the best alignment since the focus for
the both alignment mark and the device are very close. However, the tolerance
for vertical offset between the top of the mark and the plane of resist where the
device will be exposed has not been explored. In order to have the tops of the
alignment marks and the sputtered multilayer lie in approximately the same plane,
the wafer began with photolithography, shown in Figure 3.45, that defined holes
in the multilayer stack, shown Figure 3.46. The layering began with an IBD Ti
50 A adhesion layer. Vacuum was broken for approximately 2 minutes to move
from the IBD to the AJA, where the rest of the stack was sputter-deposited;
underlayers/CoFeB 400 A/Cu 200 A/AlOx 15 A, where underlayers ≡ [Ta 50
A/Cu 50 A/CuOx 200 A]2/Ta 150 A. A thick layer of C (∼100 nm) was e-beam
evaporated and would serve as the ion mill mask, shown in Figure 3.47. This C
was removed by an O2 plasma from the holes where the alignment marks would
be defined, shown in Figure 3.48.
A similar photolithography mask was used to define what is called the alignment
marker base, a slab of material from which the alignment mark is etched, shown
in Figure 3.49. This mask was different from the one defining the C etch in
Figure 3.48 in that the square windows are smaller by 5µm along the perimeter.
This was done to prevent a step edge in the alignment marker base from forming in
the event of slight misalignment in the photolithography, potentially introducing
129
(a)
(b)
200 nm
200 nm
Figure 3.44: SEM images of VB6 fine alignment. (a) The nanopillar to bealigned to. The patterned layers are indicated in the schematic on the right. (b)SEM image of after the aligned exposure, development, evaporation of a Cr layerand liftoff. Overlay accuracy was estimated to be ∼10 nm, which is the quotedaccuracy limit of aligned exposures achievable in the VB6.
130
complications during alignment mark definition. The mask base layers were a Ti
50 A (adhesion)/Au 1305 A/C 500 A, all e-beam evaporated without breaking
vacuum. The C cap served as the ion mill mask when it came time to define
the alignment mark. Note how the net thickness of Ti and Au equals the total
thickness of the multilayer (not including the C cap).
A global align procedure was executed and, as discussed above, the first level
of e-beam lithography included the device pattern as well as the alignment marks
themselves, as shown in Figure 3.51. After development Cr 10 A/Au 250 A evap-
oration was followed by lift off. In this demonstration, the elliptical pattern was
not centered in the exposure due to geometrical constraints of the device leads.
The leads, similar to those in Figure 3.40, were patterned on this wafer, between
the fabrication steps shown in Figures 3.52 and 3.53, but are not shown here for
simplicity.
After the first level e-beam, the mark pattern was transferred into the base
in a two stage dry etch step as highlighted in Figures 3.34 and 3.35. First, the
C layers (for both devices and alignment marks) were etched anisotropically with
an O2 plasma, as shown in Figure 3.52, with the same etching criteria as for the
standard nanopillar process. Once the mask was patterned, it served as an ion
mill mask for underlying material. Since alignment mark definition ion mill time
was twice that of pillar definition ion mill time (in this case, etching to the top
of the CoFeB layer), the two etches were done separately. Photolithography was
used to protect the device mask while the alignment marks were etched, as shown
in Figure 3.53. This etch was done in the Veeco, with standard beam parameters
and Ibeam = 60mA, stage tilt φ = 25, and stage rotation on.
The alignment mark C mask was etched in an O2 plasma. The resist was
131
(a)
(b)
Figure 3.45: (a) Photolithography features that defined 40 µm holes in themultilayer where the alignment mark base material would be deposited. Placementof these features must reflect the positioning of the alignment marks to be definedin the first e-beam lithography step. (b) The geometry of these holes was suchthat their centers laid at the corners of a 300 µm square, close to the extent ofthe VB6 exposure field. Eventually, the 10 µm square fine alignment marks wouldbe defined inside these holes which would then serve as clearance regions of thee-beam alignment. These dimensions were small enough not to obstruct othersurface features.
132
Figure 3.46: The multilayer stack after lift off. The layering is described in thetext.
stripped and the alignment marks were then protected by photolithography while
the device was defined by ion milling in the IBD, at Tstage = -10.0C, φ = 45,
in an Ar + O2 ambient at PCM = 0.32 mTorr, stage rotation on, Vbeam = 500 V,
Vaccel = 200 V, Vdis = 40.0 V, and Ibeam = 60 mA. Total etch time was 100 sec,
which should have completed the etch approximately 20 A into the CoFeB layer.
This sequence is shown in Figure 3.54.
When both the nanopillar and the alignment marks were defined, fine alignment
to the pillar was performed with a typical PMMA resist bilayer. The fine alignment
procedure was executed using the “edge” location algorithm, as described above,
as shown in Figure 3.55. All four alignment marks were located successfully on
approximately 90% of the devices. The remaining 10% could only locate three of
the four, consistently not finding the fourth alignment mark (lower left mark in
Figure 3.51). It is not clear how this affected the resulting exposure, although the
133
(a)
(b)
Figure 3.47: (a) A thick (100nm) C layer was blanket e-beam evaporated ontothe sample to serve as a mask material for the nanopillar. (b) Photolithographyto remove this thick C mask layer.
134
(b)
(a)
(b)
Figure 3.48: (a) The thick C mask layer was etched from the holes by an O2
plasma. (b) Stripped resist.
135
(a)
(b)
Figure 3.49: (a) Photolithography defining Au mark base. The photolithographyholes were slightly smaller (35 µm) than that for Figure 3.48 to prevent compli-cations during alignment mark definition. (b) After evaporation and liftoff of thealignment mark base materials Ti(adhesion)/Au(alignment mark)/C(etch mask).The height of the Au was coincident with the height of the multilayer (without theC).
136
Figure 3.50: The first level e-beam lithography exposed the device pattern andthe alignment marks. After Cr 10 A/Au 250 A evaporation and liftoff, the singlefeature atop the multilayer was the device pattern, a nominally 120 nm × 240 nmellipse. The Au features on top of the alignment mark bases were the 10 µm squarefine alignment marks.
VB6 always returned that it executed the exposure successfully. The layout of the
overlaid exposure is shown in Figure 3.56, which deliberately exposed crosses on
top of the alignment squares. Overlay of the cross on top of the alignment squares
was an indicator (i.e. easy to see in an SEM) that the alignment procedure worked
at the ∼1 µm level. Poor alignment at this level prevents having to look for the
aligned nanopillar as it would already be known that the alignment failed.
137
Figure 3.51: The top-down view of the whole exposure pattern from Figure 3.50where the layout is better seen. The alignment mark centers laid at the cornersof a 300 µm square, exactly coincident with the bases. The exposure field 310 µmincluded the alignment mark width, but was still within the maximum exposurefield of 327.68 µm. The elliptical pattern was not centered in the exposure owingto geometrical constraints of the device leads (not shown).
138
(a)
(b)
Figure 3.52: (a) The C mask layers were etched in an O2 plasma, similar to whatis done in the normal nanopillar process. (b) Top down view showing the relativeplacements of the elliptical pattern, alignment marks, and alignment mark bases.
139
(a)
(b)
Figure 3.53: (a) Photolithography was used to isolate the alignment mark ionmill etch definition. (b) The etching was done at 25 with rotation.
140
(a)
(b)
Figure 3.54: (a) Photolithography protected the alignment marks for the nanopil-lar definition etch. (b) The nanopillar was etched in an Ar + O2 ambient mixture.Etch parameters are given in the text.
141
(a)
(b)
Figure 3.55: (a) A standard e-beam resist bilayer was spun onto the wafer. (b)A schematic of the beam looking for the alignment marks.
142
(a)
(b)
Figure 3.56: (a) A Cr layer was evaporated into the exposed PMMA. (b) Theoverlaid exposure. Crosses were designed to fall on top of the alignment marks,giving a crude but effective and easy way of determining if the alignment procedurefailed.
143
3.5 Chemical mechanical polishing
3.5.1 Overview
In general, for any wafer fabrication process, sequential etching and deposition
steps tend to exacerbate the topography of the wafer surface. Severe wafer to-
pography can adversely affect the quality of fine photolithography resolution, and
can even render them non-reproducible. Chemical mechanical polishing (CMP),
a technique which globally planarizes a full wafer, was thus developed for the
semiconductor industry, which run fabrication processes that can see hundreds of
photolithography steps [7, 8].
As the name implies, CMP is a polishing process which is part chemical and part
mechanical. The wafer is mechanically buffed by rotation against a polishing pad
which is itself attached to a spinning table (Figure 3.57). A slurry deposited onto
the polishing pad provides the chemical etch. Various slurries and polishing pads
have been developed to provide good planarization recipes for specific materials
prevalent in the semiconductor industry, such as dielectrics, poly-Si, W, and Cu.
Controlling the different contributions of chemical and mechanical etching can
change the nature of the etch [9] from one with good selectivity but faster rates
(more chemical) to one with slower rates but poorer selectivity (more mechanical).
The chemical etch is determined by the choice of slurry for any given material
to be polished, of which there are only a few available choices. The mechanical
etch is partly determined by the choice of the polishing pad, but other parameters,
table rotation speed, head rotation speed, and downward pressure, also play a role
and provide much more flexibility in changing the mechanical etch contribution.
The CNF CMP tool, a Strasbaugh 6EC, has polishing heads fitted for 3”, 4”, 6”,
144
Figure 3.57: Schematic of the CMP process. The wafer is held to the rotatingchuck by a slight vacuum. The spinning table has the polishing pad adhered toits top. The slurry is deposited continuously to the pad while the wafer is beingpolished. The chemical etch is determined mostly by the choice of slurry andthe material to be polished. The mechanical etch is determined by the choice ofpolishing pad, table rotation speed, chuck rotation speed, and downward chuckpressure.
145
and 8” wafers, although the best planarization uniformity is found for 4” wafers.
The tool is also typically configured to polish 4” wafers and so it is easiest simply
to comply with this standard. Doing so does mean a loss on throughput for wafers
requiring a CMP during e-beam exposures, since two 3” wafers can be exposed
during a single pump down on the VB6 versus only one 4” wafer, and during
evaporations, where four 3” wafers can fit into the SC4500 evaporators versus only
three 4” wafers.
3.5.2 Adhesion layer
A strong adhesion layer should be used to bond the multilayer to the Si/SiO2substrate.
This is because the CMP-induced stresses in the films may cause delamination. The
strongest adhesion layers (that are conveniently deposited in Clark Hall) are Ti or
Cr, evaporated, IBD-deposited, or sputtered, and should be 50 – 100 A thick [10].
If adhesion layer deposition occurs in a separate vacuum chamber from the AJA,
transfer into the AJA should occur within ∼5 min after venting, although transfer
from either the IBD or the Sharon evaporator can be done as quickly as 2 min.
3.5.3 Polishing dielectrics
The following section summarizes some CMP tests on three dielectric materials,
PECVD oxide (SiOx), PECVD nitride (SiNy), and IBD AlOx. Two films of each
type were deposited onto blank 4” Si substrates. PECVD films were grown with
platen temperature of 170C for both the oxide and nitride recipes. IBD AlOx was
deposited with parameters given in tables 3.17 and 3.18. Optical ellipsometry was
performed on each of these wafers in order to measure the index of refraction as a
function of wavelength. These data were then entered into an optical interferometer
146
Table 3.19: Deposition thickness uniformity of some dielectric materials thatare available with the on-site deposition tools. Deposition uniformity data aremeasured by optical interferometry at n = 37 points evenly distributed about thewafer. All films are 480 – 520 nm thick.
Film Deposition Thickness Uniformity (%)
PECVD nitride 4.0
PECVD oxide 6.3
IBD AlOx 1.2
in order to accurately measure the film thickness across each wafer. Table 3.19
shows the film thickness uniformity across the entire 4” wafer as measured by
sampling film thickness at n points evenly distributed across the wafer, where
uniformity = 2 × 100 × (standard deviation)/(mean).
All CMP recipes in this chapter use the IC-1400 polishing pad and the Semi-
Sperse (SS-12) slurry. These films were etched on the Strasbaugh 6EC CMP tool
with the “Oxide” recipe (chuck pressure = 10 psi, chuck rotation = 15 rpm, table
rotation = 35 rpm). Etch rate (mean / polishing time) and uniformity data,
averaged over two CMP runs on two different wafers, are shown in Table 3.20. One
PECVD nitride wafer was then polished using a different recipe “lowpressoxide”,
which applies less downward chuck pressure, but compensates by increasing the
rotation speeds (chuck pressure = 4.5 psi, chuck rotation = 80 rpm, table rotation
= 50 rpm). As shown in Table 3.21, very comparable etch uniformities can be
achieved simultaneously with a reduction of 20% in the etching rate between the
two recipes.
Comparing the data of tables 3.19 and 3.20, it is clear that choice of material
147
Table 3.20: CMP data for various dielectric films. Two wafers of each dielectricwere polished using the “Oxide” recipe, each for 45 sec. Etch rate and uniformitydata are measured by optical interferometry at n = 37 points evenly distributedabout the wafer.
Film “Oxide” Recipe Rate (nm/min) Etch Uniformity (%)
PECVD nitride 140 7.3
PECVD oxide 219 8.6
IBD AlOx 369 9.3
Table 3.21: CMP data for PECVD nitride. Data for the “Oxide” recipe areaveraged over etches on two wafers (45 sec polishing each) while the data for“lowpressoxide” (a 60 sec polish) was done on one of the wafers after it had gonethrough the “Oxide” recipe. Recipe parameters are described in the text. Etchrate and uniformity data are measured by optical interferometry at n = 37 pointsevenly distributed about the wafer.
Recipe Rate (nm/min) Etch Uniformity (%)
“Oxide” 140 7.3
“lowpressoxide” 112 6.3
148
plays a larger role in determining etch uniformity than does the overall deposi-
tion uniformity. However, it is also clear from Table 3.21 that proper choice of
etch parameters can reduce etching rate without compromising etch uniformity.
Exploration of etch parameters is discussed in the next section.
3.5.4 Effects of changing CMP process parameters
This section discusses some results of changing the CMP parameters while pol-
ishing IBD AlOx. This dielectric was chosen over PECVD nitride and PECVD
oxide, which give better CMP properties (Table 3.20), because it is much better
suited for the insulation material in the proposed HSQ mask nanopillar fabrication
process (section 3.10), because of its better deposition uniformity (Table 3.19). Pa-
rameters explored were the chuck and table rotation speeds and chuck pressure.
Although there are a few choices of slurry and polishing pad, the IC-1400 pad and
SS-12 slurry were chosen as the consistent pad/slurry combination since they are
the most frequently used and it is advantageous to minimize the inconvenience and
cost of having to change the pad.
Several 500 nm thick IBD AlOx films were deposited onto 4” Si wafers and
polished on the CMP tool with the “Al2O3recipe” (chuck rotation = X × 5 rpm,
table rotation = X × 6 rpm, and chuck pressure = Y × 1 psi), where X and Y
= 1 or 2. The AlOx index of refraction was characterized with ellipsometry, as
before. For X = Y = 2, one wafer was polished for three consecutive times (each
for 1 min) in order to see how etch rate and etch uniformity would change from
run to run. Data are shown in Figure 3.58, and it is clear that with this choice of
etch parameters, the etch rate can drift by ∼30% from run to run, although the
uniformity remains consistent, albeit somewhat higher than desired, at ∼20%.
149
1.0 1.5 2.0 2.5 3.036
38
40
42
44
46
48
50
0
5
10
15
20
25
Etch rate
Etch
rate
[nm
/min
]
CMP run
Uniform
ity [%]
Uniformity
Figure 3.58: CMP etching rates of IBD AlOx as a function of run number for X= Y = 2, each polished for one minutes. The etching rates are not consistent fromrun to run. The etching uniformity does not closely follow the etch rate. Etchrate and uniformity data are measured by optical interferometry at n = 121 pointsevenly distributed about the wafer.
150
Etching uniformity for these films is much worse and the etch rates much slower
than those quoted in Table 3.20. This shows that more aggressive polishing con-
ditions (larger rotation speeds (X) and chuck pressures (Y)) tend to yield more
uniform results at the expense of higher polishing rates. The rotation factor X was
varied from 1 to 2, shown in Figure 3.59. The etch rate increases, as expected, and
the uniformity gets slightly worse. In Figure 3.60, I show the result of altering the
chuck pressure. In this low pressure regime, the rates seems to increase linearly
(extrapolating to etch rate = 0 for chuck pressure = 0) while the uniformity gets
slightly better.
These data suggest that a good starting point for realizing reasonably slow
CMP etch rates (∼100 nm/min ≈ 17 A/sec) with decent uniformity (5 – 10%) is
to keep the rotation somewhat slow, but use moderate chuck pressures (∼5 – 6
psi). We note, however, that for X = 1 and 2, the table exhibited a “stick-slip”
motion during the polishing. This is because the table is a fairly massive block of
stone and the natural friction in the rotation bearing along with its own inertia
prevent smooth rotation for too-slow speeds. Smooth table rotation will provide
much more uniform polishing.
3.5.5 Comparison with ion mill planarization
In the current nanopillar fabrication process, the PECVD oxide insulating layer
is planarized by high angle ion milling. Planarization by ion milling works by
thinning oxide protrusions from the side. Simultaneously, however, it also widens
surface pits, and so film roughness is somewhat preserved during the planarization
etch [1]. CMP, in comparison, is a global planarization process which can yield a
much smoother surface, with roughness nearly an order of magnitude smaller.
151
1.0 1.5 2.0
1520253035404550
0
5
10
15
20
Etch rate
Etch
rate
[nm
/min
]
Rotation Factor X
Uniformity
Uniform
ity [%]
Figure 3.59: CMP etching rates of IBD AlOx as a function of rotation factor X forY = 2 (chuck pressure = 2 psi), both polished for two minutes. The etching ratesincrease while the uniformity is virtually unchanged. Etch rate and uniformitydata are measured by optical interferometry at n = 37 points evenly distributedabout the wafer.
152
1.0 1.5 2.0
25
30
35
40
45
50
0
5
10
15
20
25
30
35
Etch rate
Etch
rate
[nm
/min
]
Chuck Pressure [psi]
Uniformity
Uniform
ity [%]
Figure 3.60: CMP etching rates of IBD AlOx as a function of chuck pressure forX = 2, both polished for two minutes. The etching rates are nearly linear withchuck pressure in this low pressure regime. The etching uniformity drops almostby half. These data provide good insight as to how best to achieve a CMP recipewith reasonably low etch rates and good uniformities. Etch rate and uniformitydata are measured by optical interferometry at n = 37 points evenly distributedabout the wafer.
153
Figure 3.61 shows a 10 µm square AFM image of an as-deposited PECVD oxide
layer, which is the insulator used in the standard nanopillar fabrication process.
The as-deposited roughness is on the order of 10 nm (rms). An AFM image of
a similar oxide layer that was planarized by high angle ion milling is shown in
Figure 3.62 with a planarized nanopillar at the image center. The height scale is
the same for both figures, although Figure 3.62 is a 4 µm square scan area. The
ion mill planarized oxide shows small sharp protrusions with larger, bowl-shaped
pits, in contrast to the as-deposited film, which shows roughly even protrusions
and pits. More importantly, the scale of the roughness has decreased by nearly a
factor of two, from ∼10 nm (rms) to ∼ 5nm (rms) by ion mill planarizing.
CMP planarization on a similar PECVD oxide layer, however, shows a signif-
icantly smoother surface than that demonstrated by ion mill planarization (Fig-
ure 3.62). An 10 µm square AFM image of the CMP planarized film is shown
in Figure 3.63, where it shows the roughness to be ∼ 0.4nm (rms), better than
an order of magnitude smoother than the ion mill planarized sample. The CMP
planarization is also more global smoothing protrusions as well as pits. The sam-
ple in Figure 3.63 was planarized using the “lowpressoxide” recipe on the CMP
tool, with an IC-1400 polishing pad, SS-12 slurry, table rotation of 50 rpm, chuck
rotation of 80 rpm, and chuck pressure of 4.5 psi.
Amazingly, this smoothness can be preserved through subsequent ion milling
steps. As shown in Figure 3.64, the oxide layer gets no rougher even after 2000
seconds of ion milling at 0 incidence. This could be useful as a two stage oxide
CMP-planarization and ion mill-thinning etch process.
154
Figure 3.61: A 10 µm square AFM image of the PECVD oxide used for insulatingnanopillars in the C mask process. Roughness is ∼10 nm (rms) and the roughnessfeatures (protrusions and pits) are approximately symmetric. Normal PECVDoxide deposition parameters for Process 1 on the IPE PECVD tool, with Tplaten =170
155
Figure 3.62: A 4 µm square AFM image of a PECVD oxide layer with a nanopillarafter ion mill planarized. At the center of the image is a planarized nanopillarwith a base to peak height ∼12 nm, which is about as good a planarization as ispossible with high angle ion milling. The oxide surrounding the pillar is about halfas rough ∼5 nm (rms) as that of the as-deposited film. The lateral dimensions ofthe roughness are unchanged, although the shape indicates that the protrusionswere trimmed while the pits were made slightly deeper. Ion milling was done at φ= 85 and Ibeam = 60 mA for 30min with normal Veeco ion mill parameters (exceptfor the shutter being open continuously), shown in Table 3.14.
156
Figure 3.63: AFM image of a CMP-smoothed PECVD oxide using the “lowpres-soxide” recipe. RMS roughness (0.4 nm) is better than an order of magnitude thanthat for ion mill planarization.
157
0 500 1000 1500 20000.0
0.1
0.2
0.3
0.4
0.5
RM
S ro
ughn
ess [
nm]
Ion mill time [sec]
Figure 3.64: RMS roughness on the same PECVD oxide film as in Figure 3.63that has been ion milled at φ = 0 for different lengths of time. The time = 0 rmsroughness comes from the CMP-smoothing of the film (Figure 3.63). This smooth-ness is conserved when etching through the entire oxide film. RMS roughness ismeasured by AFM scans of the surface after each milling step. Normal Veeco ionmilling parameters were used (Table 3.14).
158
3.6 Radio frequency (RF) backsputtering
In addition to sputter deposition, the AJA is also configured to sputter etch the
sample itself, with a radio frequency (RF) plasma. RF plasma etching of the
sample, also known as backsputtering, allows for in situ cleaning and light plasma
etching of the sample before or during sputter depositions. This will prove useful
for fabrication processes requiring a brief clean of exposed metallic surfaces before
sputter deposition of overlayers, leading to better Ohmic contact between the two
layers. This is a very similar idea to using the ion mill to clean the top of a
nanopillar before top lead deposition. However, RF backsputtering allows for these
overlayers to be sputtered in the AJA, which yields films of much higher purity
than does the IBD.
In the AJA, the RF plasma is sustained between the substrate stage (cathode)
and the chamber (anode). Although the RF power supply goes up to 555W, all RF
backsputtering should be done at 30W so as not to damage any of the hardware
interior to the chamber. Like other RF plasma tools, the AJA has a matching
network to adjust tuning and loading to minimize the amount of power reflected
from the cathode, but it usually takes a few seconds to adjust the tuning and
loading capacitors (either manually or automatically). As such, the actual etch
rate on a sample will not be well estimated if the tuning and loading are not set
correctly beforehand on a sacrificial sample.
The sacrificial sample must be similar in size, shape, and substrate material to
the real sample to properly adjust the tuning and loading. Do not use the magnetic
stage as the field lines of the stage magnets will significantly affect the plasma and
cause major etch non-uniformities. To backsputter, it is easiest to operate the AJA
with the software in manual mode. Flow Ar at 20 sccm through the sputter guns
159
and establish an Ar pressure of 20 mTorr. Make sure the N connector output from
the matching network is connected to the RF power input on the stage column,
up at the top of the AJA.
With the RF power off, manually adjust the tuning and loading to initial values
0.3 and 0.4, respectively, then switch them both to auto. These nominal values
are close to the actual values that minimize the reflected power, but the matching
network will make slight adjustments once the power is turned on. With stage
rotation on, turn on the RF power supply to 30W and the tuning and loading
will automatically adjust themselves to minimize reflected power, but should come
to rest close to their initial values. The tuning and loading parameters do not
require readjustment after each wafer but should be adjusted once per day. The
backsputter parameters are summarized in Table 3.6. Once the tuning and loading
has been adjusted, change out the sacrificial sample for the real sample and repeat
the procedure without adjusting the load and tuning at all.
There is a possibility that some of the backsputtered material, which may
include metal, oxide, or resist materials, may condense on the sputter targets or
the inner walls of the sputter gun chimneys. Consequently, once the sample has
been etched and removed from the chamber, it is important to condition all of the
sputter guns by running them at their calibrated deposition powers 1 mTorr Ar.
160
Table 3.22: These are the parameters for RF backsputtering in the AJA. Adjust-ment of the matching network tuning and loading is described in the text and mustbe done before backsputtering an actual sample. F and R in the “Actual power”row are the amount of transmitted and reflected power to the stage, respectively.
Parameter Value
Ar flow 20 sccm (through guns)
Ar pressure 20 mTorr
Stage rotation 100%
Set power 30W
Actual power F ≈ 28W; R ≈ 1W
An RF backsputtering etch rate calibration was performed on Cu (with no
cap) that had been exposed to photolith processing and IBD AlOx. With the
parameters shown in Table 3.6 the Cu sample was etched for a total of 31 min.
Actual values of the etch were as follows. At t ≈ 0, F = 28W and R = 2. At t
= 9 min, F = 29.5W and R = 0. An average Cu/CuOx RF backsputter etch rate
is 42.4 A/min. The AlOx was etched for 5 min with F = 28 and R = 2 for the
duration, yielding an etch rate of 1.48 A/min.
3.7 Hydrogen silsesquioxane (HSQ)
3.7.1 History
Although not developed explicitly as a an e-beam resist, HSQ has become a widely
used negative tone resist for e-beam lithography tests of future, nanoscale device
applications such as hard drive read heads [11]. The origins of HSQ as an e-beam
resist come from two separate innovations in silicon oxide processing: demonstra-
161
tion of thermal SiO2 as an e-beam resist material and the invention of HSQ as a
spin-on form of silicon oxide for application as a self-planarizing interlayer dielec-
tric.
Silicon dioxide as an e-beam resist
Patterning thermal SiO2 as a positive tone resist was first shown nearly four
decades ago [12], and further refinement of the exposure technique has since demon-
strated ∼10 nm linewidth resolution with a minimum period of ∼50 nm in both
thermal SiO2 [13] and oxygen doped LPCVD SiO2 [14]. Direct patterning of the
SiO2 requires very high doses (104 – 105 × that of PMMA) making it prohibitive
for actual device fabrication. Nevertheless, its compatibility with standard semi-
conductor processing makes it a desirable e-beam resist if a means of lowering the
doses could be obtained.
HSQ as a spin-on dielectric
HSQ is a flowable, inorganic polymer that was first developed by Dow Corning for
use as a spin-on insulator, that was designed to replace current deposition processes
for interlayer dielectrics in the semiconductor industry. Its low levels of contamina-
tion, low dielectric constant, excellent gap fill, and very low defect density make it
a highly effective planarization material compatible with semiconductor industry
standards [15].
Processing of HSQ starts by spin coating the wafer, which is then baked for
one minute each at 150C, 200C, and 350C to bake off the solvent, melt the HSQ
into a glass, and flow the glass to improve planarization, respectively. The final
bake also drives out volatiles and the remaining material is almost a network of Si
162
and O atoms [16]. This network is mechanically very strong and has very similar
insulating and etching characteristics as SiO2. Since the HSQ planarizes naturally
when spun onto a wafer, it is an advantageous alternative to conformal deposition
techniques such as low temperature oxide (LTO) and chemical vapor deposited
oxide (CVD), which must be planarized by a separate process, such as CMP. Ease
of application and its chemical proximity to SiO2 is likely what inspired initial
investigate into HSQ as an e-beam resist.
3.7.2 HSQ as an e-beam resist
Although HSQ and SiO2 films are both e-beam resist layers and are made up of
Si and O atoms, their likeness does not go much further than that. The exposure
processes are different for both as seen by the fact that HSQ is a negative tone
resist and SiO2 a positive tone. However, both represent an effort to develop a
process for direct e-beam patterning silicon oxide. Nano-scale patterning of SiO2,
a ubiquitous material in semiconductor processing, would be a significant step
toward more advanced devices.
Chemical structure
Understanding the HSQ chemical structure is paramount to understanding how it
behaves as an e-beam resist. HSQ molecules have a cage-like structure containing
eight Si atoms, twelve O atoms, and eight H atoms. Each Si atom lies at the
center of a tetragonal arrangement of three O atoms and one H atom. Each of
three O atoms in the tetrahedron is shared between two tetrahedra, linking them
all together in the cage-like structure. There is only one H bond per Si atom and
so the HSQ chemical label is HSiO3/2 [16].
163
Sensitivity to an electron beam comes from the Si-H bond, which is much
weaker than a Si-O bond, and is severed during the e-beam exposure. Once cut, it
is believed that the now free bond on the Si reacts with local moisture, breaking
apart the H2O into and bonding to form an Si-OH bond. This silanol is unstable
and consequently decomposes into a stable siloxane bond, Si-O-Si. This final step
is what induces the molecules to crosslink, forming a mechanically strong film of
Si and O.
Comparison of HSQ with other organic resists
Although doses required to directly pattern deposited SiO2 films are prohibitively
large for use in industrial applications, HSQ doses are comparable to those of
more common organic resists, such as polymethyl methacrylate (PMMA), ZEP1,
and SAL2.
In comparison, HSQ has several advantages over these other resists. First,
linewidth fluctuations have been found to be smaller for HSQ (standard deviation
of linewidth ≡ σlw ≈ 1.8nm) than for ZEP (σlw ≈ 2.7nm) and SAL (σlw ≈ 4.0nm)
[17]. Reducing linewidth fluctuations is critical to the success of technologies that
rely on smooth, narrow, e-beam defined lines, such as hard disk drive magnetic
read heads which are predicted to approach a linewidth of ∼10 nm by the end of
the decade [11].
This reduced linewidth fluctuation occurs because of the fact that the ultimate
resolution of e-beam lithography is not the spot size of the e-beam tool, but rather
the aggregate size of the molecules in the resist matrix. During development of
the exposed resist, the solvent removes the aggregated resist molecules and so the
1atactic copolymer of methyl α-chloroacrylate and α-methylstylene2novolac-based chemically amplified resist
164
linewidth fluctuations reflect the size of these aggregates. HSQ molecules form
smaller aggregates than do the polymeric chains of the organic resists [17, 18] and
hence, linewidth fluctuations are smaller.
Second, HSQ has a much better etch resistance than the organic resists for most
dry etch processes. The e-beam turns the HSQ effectively into SiO2, which is a
much more physically and chemically robust material and can serve as a much more
effective etch mask for pattern transfer of lithography-defined features. At the same
time, it is this fact which highlights one drawback to using HSQ. Namely, stronger
etchants, such as HF, or more aggressive etching techniques, such as CMP [19] are
needed to remove the exposed HSQ.
3.7.3 Handling and processing
HSQ processing is summarized in Table 3.23. HSQ at the CNF, which at the time
of this writing is FOx-12, FOx-16, and XR 15413, is stored in a freezer in the
e-beam resist room. It must be thawed at room temperature for at least one hour
before spinning. Once the wafer is spun, bake the wafer on the 170C hotplate
for two minutes. This bake drives off the MIBK solvent but is not high enough to
melt the HSQ. For e-beam tests of HSQ, I have used doses found by other CNF
researchers which yield good results (1500 – 2000 µC/cm2 [20]). Once exposed, do
a post-exposure bake (PEB) for 5 min on the 170C hotplate before developing.
This bake is believe to enhance the replacement of silanol (Si-OH) with siloxane
(Si-O-Si) bonds as discussed above.
Avoid using glassware during the development and indeed during any step
3The FOx series is being phased out by the manufacturer, Dow Corning, whoclaims that XR 1541 is identical to FOx-12.
165
of the HSQ processing as the HSiO3/2 molecules will react with the glass and the
wafer will become contaminated with reactant particulates. Use a plastic beaker of
300MIF (0.26 N tetra-methyl-ammonium hydroxide (TMAH)). Development time
should be 10 – 15 min, depending on the thickness of the HSQ, without agitation.
Once the wafer is placed in the 300MIF, bubbles will begin to form on the wafer
surface, an indication that the unexposed HSQ layer is etching. When done, flush
with DI water for ∼20 sec, keeping the wafer at an angle so a continuous stream
of DI flows across the wafer to remove the developed HSQ. Try to keep all of the
wafer surface covered by the stream of water. While still wet, set on the e-beam
resist spinner and spin dry at 2000 rpm. Be sure to dry the spinner chuck once
finished.
For large aspect ratio HSQ structures (height:lateral width > 2), capillary
forces associated with the transition of the solvent rinse chemical (DI) from liquid
phase to gaseous phase is strong enough to topple the HSQ pattern. As such, it
is important, once the wafer is developed, to remove the solvent rinse using the
critical point dryer (CPD) at the CNF. This tool heats the solvent under pressure
to the point where the density of the liquid and gaseous form of the rinse fluid
(liquid CO2) the same, thereby minimizing the capillary forces. Steps for the CPD
are shown in Table 3.24.
3.8 Examining ion milled nanostructures with STEM
Ion milling, despite its ubiquity in the Cornell nanopillar fabrication recipe, remains
a very poorly characterized process. The only direct feedback is the planar ion mill
etch rate, measured by profilometry, but it reveals nothing about other local pro-
cesses like re-deposition or shadowing. The best way to characterize re-deposition
166
Table 3.23: Process parameters for HSQ.
Storage Keep in the freezer in the e-beam resist room.
To use, thaw at room temperature for at least 1 hour.
Handling Only use HSQ with plasticware. Do not use glass.
Spinning Spin-coat as usual.
Solvent bake on 170 hotplate for 2 min.
Exposure E-beam dose 1500 – 2000 µC/cm2.
PEB Post-exposure bake on 170 hotplate for 5 min.
Develop (1) 300MIF bath for 10 – 15 min without agitation.
(2) Remove from 300MIF, flush with DI for ∼20 sec.
(3) Transfer to DI water bath, spraying with DI in transit.
Drying Either spin dry at 2000 rpm on e-beam resist spinner.
Or dry with the CPD (Table 3.24).
Table 3.24: Critical point drying (CPD) for HSQ wafers. Use the CPD tool isthe height:lateral width ratio > 2. The wafer starts in the e-beam room hood, ina plastic beaker of DI after development in 300MIF.
(1) Cap the beaker and carry to the acid hood to the left of the CPD
(2) Transfer wafer to methanol bath, spraying with DI in transit.
(3) Fill the CPD basin with methanol.
(4) Transfer wafer to the CPD basin, spraying with methanol in transit.
167
is by direct imaging and chemical analysis through scanning transmission electron
microscopy (STEM). STEM images of individual nanopillars was attempted, but
because the specimen preparation technique used for STEM imaging is not suited
for “zero-dimensional” samples (i.e. nanopillars), it was much more feasible to im-
age “one-dimensional” lines. STEM imaging of ion milled lines was performed by
Andre Mkhoyan, a post-doc for John Silcox. I present the fabrication process for
lines imaged in the STEM and introduce simple model to explain the images. This
simple model is a good starting point for a more comprehensive model through
which STEM images may reveal quantitative information about the ion milling
process on the nano-scale.
3.8.1 Fabrication of HSQ lines
STEM specimens were lines ion milled from multilayer films using HSQ as an
ion mill mask. The details of the fabrication are described in the captions of
Figures 3.65 to 3.81.
3.8.2 STEM images of patterned lines
Bright and dark field STEM images of one of the HSQ masked lines are shown in
Figure 3.82 and the layers are identified in Figure 3.83, which shows a brightness
gradient of the image in Figure 3.82. Black lines have been added in the trilayer
as a guide to the eye delineating the different materials since the STEM dark field
images observe Z-contrast and Fe, Co, Ni, and Cu all have very similar Z. There
are some interesting, albeit unexpected, aspects of these lines. First is the profile
of the top of the AlOx layer just beneath the top lead Cu. Instead of being flat, it
has a mote-like topography, where there is a slight dip in the AlOx height before
168
Figure 3.65: Metallic layers were deposited onto oxidized Si substrates with thelayers underlayers/CoFe 100 A/Cu 100 A/Py 50 A/Cu 200 A/Pt 300 A//Ta 50A. The Ta cap was sputter deposited more than two weeks after the initial films.For this wafer underlayers = Ti 100 A//[Ta 50 A/Cu 20 A/CuOx 200 A/]2/Ta200 A is layered for smoothness since the Ta/Cu/CuOx multilayers reduce theoverall film roughness [21]. “CuOx” designates Cu grown in an oxygen ambient,not Cu-oxide. For both wafers the Ti 100 A adhesion layer was IBD deposited andtransferred to the AJA within 2 min of breaking vacuum.
169
Figure 3.66: The resist bilayer was then spun; PMMA 2% 495k (in Anisole) at4000 rpm (50 nm), baked for 15 min on 170C hotplate; XR 1541 (HSQ) at 1000rpm (spin speed charts claim 200 nm, optical interferometry measured a similarHSQ film at 167 nm thick when spun on a bare Si wafer), baked for 2 min on170C hotplate.
a mound that goes over the HSQ mask. This is likely caused by shadowing of the
AlOx from the HSQ mask itself during AlOx deposition. This indicates that the
tilt angle of φ = 180 may not be an acceptable deposition angle for the AlOx
insulation step. Also, the mask material seems to have remained, even after the
CMP planarization. Note that the PMMA layer has been dissolved away during
the STEM specimen preparation. However, the gap left is much less than the 50
nm at which it is claimed to have been spun on. This may indicate that the PMMA
was partially etched by the HSQ solvent (methyl-isobutyl-ketone (MIBK)). This
may indicate that polyimide may should be used as an HSQ underlayer.
The slightly parabolic dish shape just above the Ta layer is residual mask
material (HSQ), which has been filled in on the top with IBD AlOx. Redeposited
material (50 – 100 A thick) is very obvious on either side of this residual mask
and accounts for the measured increase in width of the mask lines after ion milling
170
(a)
(b)
Figure 3.67: (a) 100 nm wide × 300 µm long lines, spaced by 2 µm to filla region 1.8 mm long, were exposed on the wafer in the VB6 with a 2000µC/cm2 dose. Lines were exposed with the jobfile HSQ STEM LINES.COM inthe [VB.USERS.EMLEY2000.JOBS] directory. A PEB for 5 min on a 170C hot-plate was followed by development in 300MIF for 15 min (no agitation). Then thewafer was transferred to a DI water bath, flushing lightly with DI while in tran-sit, then transferred to a methanol bath, then to the critical point dryer (CPD),spraying with methanol while in transit. (b) Once dried, the PMMA was etchedoff in an O2 plasma (100W, 30 sccm, 30 mTorr, 10 min chamber clean) for 45 sec.
171
Figure 3.68: Top down schematic of HSQ lines. HSQ lines were exposed with 2µm spacing and were 300 µm long so as to fit inside a single exposure field. Nostitching of the lines was necessary. With this spacing, 900 lines were exposed inparallel forming a 1.8 mm long “comb” of lines. Five such “combs” were exposedacross the wafer. Total exposure time ∼2 hours per wafer.
172
Figure 3.69: SEM image of many of the lines. Slight curvature in the lines ispresent, perhaps due to slight swelling of the PMMA underlayer.
173
Figure 3.70: SEM image of one of the lines. The width, measured on the SEM,is 80 nm. The engineered width was 100 nm.
(Figure 3.74). The Ta has been completely encapsulated by the redeposition.
Where the Cu lead meets the line, redeposited sidewall material looks as if it has
been smeared slightly to the right, possibly due to the impact of the CMP step.
The fact that this redeposited material has pierced the AlOx shows that some of
the AlOx was polished away by the CMP. These sidewall redepositions may be
supporting the residual HSQ mask material, thereby hindering its removal by the
CMP. Future attempts at STEM imaging should be performed testing different
ion mill, AlOx deposition, and CMP parameters.
3.8.3 Shadowing effects: a simple model
I describe a very simple model first introduced in ref [22] for line-shaped masks, but
which I expand here to include pillar-shaped (i.e. cylindrical) masks as well. This
174
Figure 3.71: SEM image of one of the lines at a perspective. Grain structure ofthe sputter-deposited films is evident in the image.
Figure 3.72: AFM image of a PMMA/HSQ line. Height measured at 169 nm.
175
Figure 3.73: After development, the wafers were ion milled at φ = 135 (45
from beam) in the IBD with Pbase = 2.0 – 2.4 × 10−7 Torr, Vbeam = 200 V, andIbeam = 70 mA (other parameters shown in tables 3.13 and 3.12). The shutter wasopen continuously, etching for 526 sec (aiming for just to the bottom of the CoFelayer).
176
Figure 3.74: Vacuum was broken to do SEM and AFM measurement on theion milled samples before AlOx deposition. The stage temperature was warmedfrom T = -10.0C to +20.0C at the normal warming rate of the stage chiller(∼1C/min). SEM image of an ion milled line. Width is now measured to be100 nm, an increase by 10 nm on either side of the HSQ line, presumably due toredeposited material.
177
Figure 3.75: SEM image of an ion milled line at its edge. Evidence of shadowingeffects are subtle in this image and in Figure 3.74, but are observed as a narrowperimeter surrounding the HSQ line which shows slightly different grain structurethan the rest of the continuous film. A dark line down the spine of the HSQ maskis also something coming from the ion milling step.
178
Figure 3.76: SEM image of an ion milled line at a perspective. The dark linefrom Figure 3.75 is observed here as a slight concavity in the HSQ lines. This isconfirmed in subsequent AFM measurements (Figure 3.77).
Figure 3.77: AFM image of a line after ion milling. The total height, the sum ofthe remaining HSQ, the PMMA underlayer, and the etched metals, has decreasedafter the etch, now measuring 136 nm.
179
Figure 3.78: Before AlOx deposition, the target was conditioned for 5 min (i.e.source on, shutter closed). The AlOx was deposited (in Ar + O2) at φ = 180 (20
from line-of-sight to target) to a film thickness of 15 nm above the measured lineheight on the AFM. Deposition conditions are shown in tables 3.17 and 3.18.
180
(a)
(b)
Figure 3.79: (a) CMP planarization was performed using the “Al2O3recipe”recipe for 15 sec. AFM measurements show that 30 nm lines remain but processingwas continued for lack of time. (b) Spin resist for top lead definition.
181
(a)
(b)
Figure 3.80: (a) Spin PMMA (2% 495k at 3000 rpm) followed by S1827 (at 2000rpm) with respective bakes. Use mask “SpinMob 2” (from the C mask fabricationprocess) with 5 sec exposure on the 5× stepper. This mask has a large rectangularwindow directly over where the lines are. O2 etch PMMA for 2 min (100W, 30sccm, 30 mTorr) (Figure 3.80). (b) Top view.
182
(a)
(b)
Figure 3.81: (a) The top Ta cap was ion mill cleaned at normal incidence to theion beam (φ = 90) for 150 sec with standard beam parameters. (b) A 200 nmthick top Cu layer lead was deposited in situ after the Cu target was conditionedfor 5 min (i.e. source on, shutter closed).
183
(a)
(b)
Figure 3.82: Cross-sectional STEM image of a line fabricated with the processoutlined above. (a) Bright field (i.e. TEM) and (b) dark field (i.e. Z-contrast)images. The partially-shadowed edge is the curved edge indicated by the yellowarrow. Flatter regions beyond this curved edge were not shadowed by the mask andfollow the calibrated planar etch rates. The shape of this edge gives informationabout redeposition and shadowing in the ion milling process. STEM image analysiswill give valuable information that will help better understand the characteristicsof ion milled nanostructures. See Figure 3.83 for layer identification. This samplewas under etched since the targeted etch depth was to the Ta/CoFe interface whilethe actual depth is closer to the top of the CoFe layer.
184
Cu top lead
AlOx Pt
Cu
CoFe
Ta
Py
Cu
Ta
Cu/CuOxTa
Cu/CuOx
TaTi
SiO substrate2 400 A
AlOxresidual HSQ
Figure 3.83: The layers are identified by the estimated deposition thicknesses, se-quence, and corroboration by the STEM through known Z-contrast and measuredlayer thickness. This image is a spatial gradient of Figure 3.82(b), highlighting theboundaries between materials. The top-most Ta (above the Pt) is artificially col-ored white for better viewing. Its roughness is indicative of the top-most granularlayer in the SEM images taken before ion milling (see Figure 3.71). The thin (20A) Cu layers in the bottom lead are not distinguishable on this scale and are justincluded with the oxygenated CuOx layers. Black lines are drawn in as a guide tothe eye to delineate the CoFe/Cu/Py layers.
185
model is merely to understand the how the linear mask shadows the multilayers
from the ion beam by deriving a functional form of the sidewall profile. Reasonable
qualitative agreement is found between the simple model and the STEM image of
an ion milled line shown in Figure 3.82. Four significant details not covered in
this model, (1) sidewall redeposition, (2) mask etching, (3) multiple layer etch-
ing, and (4) time-dependence of the etch may account for quantitative differences.
Exploration of these four details would be very valuable since it may be possible
to compare model predictions with STEM images of lines to derive quantitative
information about the ion milling process at the nano-scale. Ultimately, this infor-
mation could be used to adjust ion mill recipes for nanopillars to achieve controlled
ion mill profiles.
Ion beam shadowing from a linear mask
Take the example where we have an ion mill mask, in the shape of a parallelepiped
that has an etch rate of identically zero. The shape of the mask is such that its
length is infinite while the height to width ratio is approximately one. A schematic
of such a mask is shown in Figure 3.84. The main idea of this model is to calculate
how much shadowing of the sample occurs, per rotation of the wafer, as a function
of position away from the mask edge. First consider shadowing from a single point
on the “cliff edge” of the line. As the wafer is rotated in the ion beam, from this
point a circular shadow is swept out on the plane, as shown in Figure 3.84. Since
this point represents all such points along the mask “cliff edge”, the cumulative
shadowing effect of the entire line is proportional to that of a single point.
Per wafer revolution, the amount of shadowing at a distance r away from the
mask edge is proportional to γ(r), the angular distance along which the point r is
186
Figure 3.84: This is a schematic that shows the geometry of the linear mask. Asthe wafer is rotated, each point on the mask “cliff edge” shadows a circular-shapedregion at its base with diameter dependent upon the height of the mask (h) andthe tilt angle (φ). The width of the mask is not relevant for this analysis.
shadowed. Shadowing occurs only out to rmax = h tan(φ), where h is the height of
the mask and φ is the tilt angle. A top down view of the mill mask line is shown
in Figure 3.85. The functional form of γ(r), the angular distance between the two
intersection points of the shadowing semi-circle and the line parallel to the mask
edge at a distance r, is found through simple geometry.
The amount of shadowing per rotation, γ(r)/2π, is plotted as a function of r
normalized to the position where shadowing ends, h tan(φ) in Figure 3.86. For
187
Figure 3.85: Here is the essential geometry for deriving γ(r), the angular distancealong which the point r is shadowed.
positions close to the mill mask, the material is shadowed approximately half of
the time. There is a gradual decrease in the amount of shadowing as r approaches
h tan(φ). γ(r) is continuous across r = h tan(φ) and is 0 beyond. Although the
functional form shown in 3.86 is for a single etched material with an zero etch rate
mill mask, it does resemble the etch profile of the STEM images of a multilayer,
shown above in Figure 3.82. The qualitative agreement is encouraging, and future
development of shadowing this model, coupled with STEM imaging of ion milled
lines, will give valuable feedback on the process of ion milling nanostructures.
Ion beam shadowing from a cylindrical mask
Now take the example where the mask material is cylindrical in shape, shown
schematically in Figure 3.87. The situation is slightly more complicated here be-
cause now that the mask has finite lateral dimensions, the ion beam can etch the
188
Figure 3.86: This figure shows the amount of shadowing per 2π radians plotted asa function of position away from the ion mill mask r. For points directly adjacent tothe mill mask r ≈ 0, the amount of shadowing is 1/2, meaning that such locationsare milled approximately half of the time. The amount of shadowing decreases forpositions further away from the mask and for r > h tan(φ), γ = 0.
sample on either side of the mask as well as from the “cliff edge” of the mask as was
the case for the linear mask (Figure 3.84). To estimate the amount of shadowing
at a point r away from the mask, consider the geometry shown in Figure 3.88,
where the amount of shadowing γ(r) now depends on the mask diameter d.
The shadowing angle γ(r) is again found through simple geometrical relations
and is shown in Figure 3.89. The ion mill profile shows more etching in the shad-
owed region than was the case for the linear mask (Figure 3.85) because the ion
beam can now etch from the sides of the mask. There is a discontinuous jump in
shadowing across rmax = h tan(φ) because the mask is of finite height.
Future modelling
The previous section serves as a primer for ion mill modelling. It shows that simple
geometrical considerations reveal a lot of practical information about shadowing
189
Figure 3.87: This is a schematic that shows the geometry of the cylindricalmask. As the wafer is rotated, the mask shadows a circular region around themask. Since the ion beam can etch from the sides as well as the top, the derivationfor the amount of shadowing is different from the case of a linear mask. Thediameter of the shadowed mask is d and its height is h.
190
(a)
(b)
Figure 3.88: (a) A snapshot of the system at one point in the rotation. The ionbeam is shadowed by the mask as shown, where the shadowed region has the widthof the mask itself. This shadowed region sweeps out like the hand of a clock aroundthe mask. (b) The position r (denoted by the ×) is shadowed entirely within theangular distance γ(r).
191
Figure 3.89: This figure shows the amount of shadowing per rotation plotted asa function of position away from the ion mill mask r. For points directly adjacentto the mill mask r ≈ d/2, the amount of shadowing is 1/2, meaning that suchlocations are milled approximately half of the time. The amount of shadowingdecreases quickly for positions further away from the mask. The decrease is muchfaster for a cylindrical mask as compared to a linear mask (Figure 3.86) becausethe ion beam etches from sides of the mask as well as from the top. There is adiscontinuous jump in shadowing at rmax = h tan(φ) because of the finite heightof the mask.
192
from the ion mill mask. Inclusion of effects such as sidewall redeposition, mask
etching, multiple layers, as well as explicit time-dependence will produce a highly
effective tool for characterizing the local ion milling details for nanopillars from
analysis of STEM images of lines. See ref. [22, 23] for possible directions for future
modelling.
3.9 Proposed future HSQ-based nanopillar fabrication
3.9.1 Purpose
What is the point in changing a process that already works? The C mask process
does yield enough devices per wafer to achieve good scientific results. However,
an HSQ-based fabrication process similar to the one employed by Hitachi Global
Storage Technologies (see for example ref. [19]) has many advantages over the
C mask technique. These are enumerated in Table 3.25. The critical difference
between the C mask technique currently in use at Cornell and the proposed HSQ
process is that the latter uses self-aligned contact. For the C mask technique, a
critical etch is performed to uncover the oxide from its insulating bed of PECVD
oxide, which relies on correctly estimating the etch time which itself lies within a
very narrow process window. Height non-uniformity of the PECVD oxide coming
from the ion mill planarization (∼10 nm, figure 3.62) and die-to-die etching and
deposition non-uniformities (<10 nm) complicate estimation of the critical etch
time. Although the C mask layer and other capping materials (e.g. Au or Pt)
are made thick to increase the process window for a successful critical etch time
to uncover the pillar, finding the critical etch time and making reliable electrical
contact with the pillar can be time consuming, usually requiring several iterations.
193
In contrast, planarization and uncovering the pillars in the HSQ process are
done simultaneously using CMP. This “self-aligned contact” procedure removes
critical ion milling and also allows for a diverse selection of device sizes on a single
wafer, which is a limitation for ion mill planarization. More importantly, though, it
has the potential for reducing wafer fabrication time and cost. Samples fabricated
with the HSQ process at Hitachi and measured at Cornell showed close to 100%
yield, where all devices were batch fabricated on a whole wafer, meaning that no
wafer dicing and iterative searching for a critical etch time was required. Such high
yield would relax the number of devices needed per wafer, which would reduce the
total time spent on the very expensive serial process of e-beam lithography.
194
Table 3.25: Benefits of the HSQ process over the C mask process currently usedat Cornell.
(1) Potentially higher yield
(2) Lower cost per wafer
(3) Yield not dependent on device size
(4) Quicker fabrication time
(5) Easier to adapt to more complex structures
3.9.2 Fabrication process
Due to the length of this proposed fabrication process, all of the text is incorporated
into the figure captions.
Table 3.26: These are the parameters of the PMMA/HSQ resist bilayer. HSQmixtures labelled FOx-12 and FOx-16 are being phased out, although XR 1541 isclaimed to be identical to FOx-12.
PMMA (bottom) HSQ (top)
Resist Type 2% 495k in Anisole XR 1541
Spin Speed (rpm) 4000 1000
Bake Temp (C) 170 170
Bake Time (min) 15 2
Thickness (nm) 50 200
195
Figure 3.90: This fabrication process begins with sputter deposition of the mul-tilayer onto an oxidized Si substrate. Wafers should be 4” diameter in compliancewith the standard operating mode of the CMP tool at the CNF. CMP will be anintegral part of this process and it is best to first deposit an adhesion layer that willstrongly bond the overlying stack to the SiO2, protecting the multilayer from po-tential delamination coming form CMP-induced stress. Sputter deposition of therest of the stack, starting with the underlayers, magnetic layers of interest, andfinally the cap, should be done in the AJA. The schematic in Figure 3.90 shows aCoFe/Cu/CoFe trilayer spin-valve structure, although tunnel junction structurescould be deposited as well. The capping material needs to be Ta [24], although atpresent it is not entirely clear why this is so.
196
Figure 3.91: Once sputtered, the resist bilayer of PMMA/HSQ should be spun.The bottom PMMA is used not as an ebeam resist, but rather is speculated to actas a glue between the Ta and HSQ. A glue that, as will be shown later, cannotbe too strong since it must allow for the HSQ mask to be physically removedby CMP without damaging the nanopillar. Resist spinning parameters are givenin Table 3.26. Note, however that the HSQ solvent is MIBK. This is a solventfor PMMA and will likely affect the underlayer. This is consistent with STEMimages of HSQ-masked lines (Figure 3.83). It may be better to use an underlayerof polyimide which does not etch in MIBK.
197
(a)
(b)
Figure 3.92: Un-aligned e-beam lithography can be carried out in the samefashion as for the C mask process. As with PMMA, it is best to perform a dosetest for each unique layers although a dose of 2000 µC/cm2 has successfully yieldedHSQ pillars, even at for thick (500 nm) HSQ layers. After the exposure, do thePEB. (a) The HSQ pillar after e-beam exposure, PEB, 300MIF development, andcritical point drying (if necessary). The PMMA is not etched in the 300MIFdeveloper and must be etched in an O2 plasma, which is done at 100W, 30 sccm,30 mTorr, for 10 sec after plasma color changes from white to greenish yellow,typically 1 min. (b) The PMMA/HSQ bilayer after the O2 plasma etch.
198
Figure 3.93: The next step is to define the nanopillar with ion milling. It isnot possible to perform any photolithography in between e-beam development andpillar definition with the ion mill since the PMMA underlayer will dissolve in anacetone strip bath. Ion milling should be done in the IBD with Tstage = -10.0C.
199
Figure 3.94: Once ion milled, the nanopillar should be insulated with IBD AlOx.This can be done in situ, but should be done with the stage temp warmed to T= +20.0C. There is a worry that the PMMA/HSQ mask may topple due tothermal stresses if the stage temperature is raised too quickly after ion milling.A heating rate of 1C/2min did not give any toppling problems. The idea ofthermal expansion has not been confirmed, but it is suspected that rapid thermalfluctuations led to HSQ masks coming off a wafer when shipped from San Jose,California to Ithaca, New York in the middle of winter. Thermal fluctuations needto be considered when carrying wafers to and from Duffield Hall during winter.Deposition occurs at φ = 180, approximately 20 away from facing the target(≈ 160) and with stage rotation. Looking at the topography in Figure 3.82,however, φ = 180 is not and optimized angle.
200
Figure 3.95: Photolith 1 (PL(1)): “Isolate Devices”. Now that the nanopillaris safely encased inside the AlOx insulation layer, photolithography to define theleads can now take place. PL(1) should be done with the thicker S1827 spun ontop of the P20 adhesion layer. This photolith defines the device footprint. The300MIF (TMAH-based) developer etches AlOx, but since this step is meant toremove the AlOx, any incidental etch of from the developer can only help.
201
(a)
(b)
Figure 3.96: (a) PL(1): “Isolate Devices”, top view of the wafer. (b) Tone anddesign of the mask used for PL(1). Black is Cr, white is glass.
202
Figure 3.97: Etch the AlOx in hydrofluoric acid (HF).
203
(a)
(b)
Figure 3.98: (a) Ion mill the metal underlayers to electrically isolate the devices.(b) After resist strip, measure the height of the oxide h1 above the substrate. Theoxide height h1 will be the target for AlOx deposition to fill in the region that wasjust etched away.
204
Figure 3.99: PL(2): “AlOx Refill”. The same mask for PL(1) can be used forPL(2). Resist layers for PL(2) must include a protective PMMA layer, that pre-vents unwanted etching of the AlOx, an undercut profile resist, such as LOR10A,and S1813. S1813 and LOR10A will develop in 300MIF, but the PMMA must beetched in an O2 plasma. The PMMA layer can be the same as what was used forthe bottom e-beam resist layer, so the same etch parameters can be used to removethe PMMA here. The S1813 will continually etch in the O2, so do not wait for theplasma to change colors, just etch for the same time as in Figure 3.92(b).
205
(a)
(b)
Figure 3.100: (a) PL(2): “AlOx Refill”, top view of the wafer. (b) Tone anddesign of the mask used for PL(2), which can be the same mask used for PL(1).Black is Cr, white is glass.
206
(a)
(b)
Figure 3.101: (a) IBD-deposit AlOx to a thickness just slightly above h1. De-position at φ = 180 along with the undercut resist profile should allow for fullsidewall insulation. (b) Since the LOR10A reacts adversely with Acetone the resistmust be stripped in a heated bath of PG nanoremover. Check the topography ofthe union of the two AlOx layers. The refill AlOx should at least be higher thanthe insulation AlOx.
207
Figure 3.102: CMP planarization and HSQ lift off. In principle, the CMPsmooths the AlOx while at the same time removing the HSQ masks from all ofthe pillars on the wafer. Having the AlOx refill height be as close to h1 is veryimportant. If it is much higher than h1, uncontrollable etch non-uniformities couldgreatly hinder HSQ removal and potentially reduce yield. If it is much lower thanh1, delamination of the insulating AlOx from the bottom lead material may occur.
Figure 3.103: PL(3): “Define Bonding Pads”. Unlike the C mask process, thebonding pads are etched to the bottom of the pillar definition etch, just as the leadsare. PL(3) deposits a non-oxidizable metal to define the bonding pads, allowingfor good electrical contact to the leads. Resist layers here are P20/S1827 becauseprocessing will see HF as well as ion milling.
208
(a)
(b)
Figure 3.104: (a) PL(3): “Define Bonding Pads”, top view of the wafer. (b)Tone and design of the mask used for PL(3). Black is Cr, white is glass.
209
Figure 3.105: HF etch of AlOx above where the bonding pads will be defined.Discoloration where the HF etch meets the top of the underlayers indicates thepossibility that the underlayer may corrode or etch in HF.
Figure 3.106: Corrosion in the bonding pad region is cleaned by ion milling.
210
Figure 3.107: There are no Au or Pt targets in the IBD, so the wafer must betransferred into the AJA for sputter deposition of these Au or Pt. As mentioned forthe adhesion layer, transfer from the IBD to the AJA can be done in < 2min, andas such Ohmic contact to the freshly ion milled bonding pad area with sputteredAu or Pt may be good enough. To be slightly more rigorous, a brief RF backsputtercould be performed to clean off any moisture or oxide that accumulated in the 2min exposure to air, although the backsputter clean may be deemed unnecessaryin the future. Sputter Cu 200 A/Pt (or Au) 300 A. If adhesion of the bonding padsis problematic, deposit a 100 A film of Ti or Cr before the Cu/Pt as an adhesionlayer. This schematic shows the device after lift off.
Figure 3.108: PL4: ”Define Top Leads”. This step needs PMMA underneaththe S1813 since the AlOx needs to be protected from the 300MIF.
211
(a)
(b)
Figure 3.109: (a) PL4: ”Define Top Leads”, top view of the wafer. (b) Tone anddesign of the mask used for PL4. Black is Cr, white is glass.
212
Figure 3.110: Inside the IBD, the top Ta capping layer needs to be removed byion milling. This layer, although oxidized itself, is a good oxidation barrier forunderlying material, so there should be a clean layer of Cu underneath that thision mill is trying to expose.
213
(a)
(b)
Figure 3.111: (a) Deposit thick (200 - 300 nm) Cu film as top lead material. (b)Lift off, and the process is complete. Top view of completed device.
214
3.9.3 HSQ nanopillar images
This section shows some SEM images taken of e-beam exposed HSQ circles and
ellipses. All HSQ pillars have the same 50 nm PMMA bottom layer (2% 495k in
Anisole).
FOx-12
FOx-12 resist was spun onto Ti 100 A/Cu 700 A/Ta 40 A that had already been
coated with the PMMA underlayer at 3000 rpm (∼92 nm thick) and baked for 2
min on the 225C hotplate. The wafer was exposed with a dose of 1600 µC/cm2,
followed by a PEB for 2 min on the 225C hotplate. The wafer was developed in
300MIF for 2 min with no agitation, transferred to a DI water bath where it was
dipped in and out several times. The wafer was then spun dry at 3000 rpm for
60 sec. The PMMA underlayer was etched in an O2 plasma (100W, 30 sccm, 30
mTorr) for 2 min. FOx-12 pillars were ion milled at in the Veeco (φ = 45, Ibeam
= 60 mA) continuously for 2 min and are shown in Figure 3.112. Nanopillar to
nanopillar shape fluctuations are shown for one device size in Figure 3.113.
215
51nm x 98nm 56nm x 128nm
49nm x 165nm
(a) (b)
(c)
Figure 3.112: Different measured sizes of FOx-12 nanopillar shapes after the ionmilling. The engineered sizes are (a) 40nm × 120nm elongated hexagon, (b) 50nm× 150nm ellipse, (c) 50nm × 200nm ellipse. The imperfections in shape are visible,although the elongated hexagon (a) gives the most elliptical result. Note that thescales are different for each image.
216
51 nm x 167 nm
(a) (b)
(d) (c)
66 nm x 170 nm
49 nm x 169 nm 52 nm x 162 nm
Figure 3.113: Measured length by (maximum) width for four elliptical HSQ(FOx-12) nanopillars, processed on the same wafer as those in Figure 3.112. Theengineered lateral dimensions are all 50nm × 200nm ellipses. The longer dimen-sion is more reproducible while there are fluctuations on both the width and thenanopillar shape. Note that the scales are different for each image.
217
FOx-16
FOx-16 resist was spun onto Ta 500 A that had already been coated with the
PMMA underlayer at 3500 rpm (∼500 nm thick) and baked for 2 min on the
225C hotplate. The wafer was exposed with a dose of 1200 µC/cm2 followed by a
PEB for 2 min on the 225C hotplate. The wafer was developed in 300MIF for 5
min with no agitation, transferred to DI water bath (sprayed with DI in transit),
then transferred to methanol bath (also sprayed with DI in transit), then placed
in the critical point dryer (CPD) while spraying with methanol in transit. After
critical point drying, the PMMA was etched in an O2 plasma (100W, 30 sccm, 30
mTorr) for 5 min. The FOx-16 pillars were ion milled continuously at in the Veeco
(φ = 45, Ibeam = 60 mA) for 2 min and are shown in Figures 3.114 and 3.115.
3.9.4 Unknowns
There are four main unknowns within the proposed HSQ process that need to be
resolved before it can be successfully implemented as the primary Cornell nanopil-
lar fabrication process. First is that a reliable CMP recipe does not exist for the
planarization step. Initial tests on the CMP of dielectric materials attempted to
find an etch recipe with a controllably slow etch rate. For this application, I do not
think this is the right philosophy to have. Reducing CMP parameters drastically
increased etch non-uniformity as well as run-to-run variation in the etch rates as
seen in Figure 3.58. This makes a slow CMP etch a poor choice for the CMP
planarization of Figure 3.102. A better philosophy in searching for a tenable CMP
etch recipe is to concentrate more on etch uniformities, rather than etch rates. At
the point in the process of the CMP step, the highest point on the wafer are the
AlOx pillars above the actually HSQ-capped nanopillars. This means that a very
218
height = 465 nm
(a) (b)
height = 379 nm
height = 488 nm
(c)
Figure 3.114: Different circular pillars of HSQ (FOx-16) spun on at 500 nmthick. The height of the pillars = HSQ mask + PMMA underlayer + etched depthis shown for each. The image is taken at a 40 incidence, so height is measuredas vertical length on the image h divided by sin(40). The measured diametersfor these pillars are (a) 102 nm, (b) 395 nm, and (c) 89 nm. The HSQ mask in(a) shows bending and some slanting, possibly from thermal stresses as this waferwas exposed to freezing temperatures during the walk from Duffield to Clark Hall.This pillar correspondingly shows a shorter height since some of the HSQ thicknessis lost in the bend. (b) is not very circular and shows some “tails” near its base.These are attributed to a problem in the conjugate beam blanking in the VB6(mentioned in section 3.2.2). (c) is possibly the best representative on this wafer,showing very little bending, only slight slanting, and a large height. Note that thescales are different for each image.
219
height = 457 nm
(a) (b)
height = 460 nm
height = 488 nm height = 447 nm
(c) (d)
Figure 3.115: Different elliptical pillars of HSQ (FOx-16) spun on at 500 nmthick. The height of the pillars = HSQ mask + PMMA underlayer + etched metalis shown for each. The image is taken at a 40 incidence, so height is measuredas vertical length on the image h divided by sin(40). The measured short axisfor these elliptical nanopillars are (a) 59 nm, (b) 147 nm, (c) 69 nm, and (d) 84nm. As with the circular HSQ masks in Figure 3.114 some elliptical pillars showbending and slanting, as in (a), possibly due to thermal stresses. (b) is a slightlylarger ellipse and is more square on its edges. There are also some “tails” near itsbase, possibly due to the VB6 beam blanking issue (mentioned in section 3.2.2).(c) has slightly less curvature than in (a) and is correspondingly taller. (d) is verystraight but not as tall as (c), which is not well understood at this time. Note thatthe scales are different for each image.
220
uniform, and hence a very rapid etch, need only be applied for a very short time in
order to planarize the and remove the HSQ masks. Take, for example, the AlOx
CMP recipe in Table 3.20, where etch uniformity across a full wafer is better by
a factor of two than for the slower CMP recipe shown in Figure 3.58. Although
there is nearly an order of magnitude difference in the etch rate, the wafer would
only need to be polished for a few seconds, rather than a few minutes. Finding the
proper balance of etch rate and etch uniformity will take some more exploration,
but it is a much more sound approach to polish at higher, more uniform rates for a
shorter time than to use slower rates, where etch uniformities may be prohibitively
large.
Second, the best choice of HSQ mask thickness is not known. From AFM
images of HSQ lines exposed to ion milling (compare Figure 3.72 with 3.77), it
seems that the ion mill rate of HSQ is on the order of metallic etch rates, which
limits its effectiveness as a mill mask unless it can be made reliably thick enough
to satisfy a wide range of ion milling needs. As seen with the FOx-16, the HSQ
can be made very tall (500 nm), but slanting in the mask material is problematic
for making nanopillars.
Third, it is not known what the optimum height of the insulating oxide should
be (Figure 3.94). Advice from Jordan Katine [24] is to deposit only to the top
of the multilayer, but one must keep in mine that his fabrication tool set is likely
tuned to optimum performance. Lack of tight control over the CNF tools will alter
their performance, fact that should be fully appreciated and taken into account.
My feeling is that the top of the HSQ mill mask should be higher than the top of
the insulation AlOx layer above the underlayers. The ion mill etch rate of e-beam
exposed HSQ needs to be calibrated in order to determine how the HSQ mask and
221
insulation AlOx heights line up. STEM images of HSQ-masked lines (Figure 3.82)
indicate that proper AlOx deposition and multilayer ion milling well help in the
CMP lift off process.
Fourth is the quality of the contact at the top of the nanopillar after removing
the Ta cap. In previous versions of the fabrication process, there was always a non-
oxidizable metal (Au or Pt) capping the pillar. As such, the ion mill “cleaning”
step just before top lead deposition was on the order of 10 sec. To ion mill a 50
A Ta cap that has been oxidized would require over a minute in etch time. With
topography expected to be such that the top of the pillar is at the bottom of a
shallow hole in the AlOx (as illustrated Figure 3.108), shadowing effects from the
hole, redeposition of the local AlOx, and redeposition of the Ta/TaOx cap onto
the nanopillar may pose problems in making a reliable contact to the pillar.
222
References for Chapter 3
[1] Albert F.J., The Fabrication and Measurement of Current Perpendicular tothe Plane Magnetic Nanostructures for the Study of the Spin Transfer Effect,Ph.d., Cornell University (2003).
[2] Robertson J., Diamond-like amorphous carbon, Materials Science and Engi-neering R 37, 129–281 (2002).
[3] Rooks M.J., Kratschmer E., Viswanathan R., Katine J., Fontana R.E., &MacDonald S.A., Low stress development of poly(methylmethacrylate) forhigh aspect ratio structures, J. Vac. Sci. Tech. B. 20, 2937–2941 (2002).
[4] Secondary ion mass spectroscopy, Technical report, Semicon Far East,http://www.semiconfareast.com/SIMS.htm.
[5] Ozatay O., Ph.D. thesis, Cornell University (in preparation).
[6] Guillorn M., private communication, unpublished (2004).
[7] Fried D.M., private communication, unpublished (2005).
[8] Campbell S.A., The Science and Engineering of Microelectronic Fabrication,chapter 11, p. 254, Oxford University Press (1996).
[9] Woodie D., CMP primer, Technical report, Cornell University,http://www.cnfusers.cornell.edu/doc/CMP%20Primer.pdf.
[10] Benjamin P. & Weaver C., The adhesion of evaporated metal films on glass,Proc. Roy. Soc. London A 261, 516–533 (1961).
[11] Driskill-Smith A.A.G., Katine J.A., Druist D.P., Lee K.Y., Tiberio R.C., &Chiu A., Electron-beam lithography for the magnetic recording industry: fab-rication of nanoscale (10 nm) thin-film read heads, Microelectronic Engineer-ing 73-74, 547–552 (2004).
[12] O’Keeffe T.W. & Handy R.M., Solid State Electronics, volume 11, pp. 261–266, Pergamon Press, London (1968).
[13] Allee D.R. & Broers A.N., Direct nanometer scale patterning of Sio2 withelectron beam irradiation through a sacrificial layer, Appl. Phys. Lett. 57,2271–2273 (1990).
[14] Pan X. & Broers A., Nanometer scale pattern generation in desposited Sio2
with electron beam irradiation, J. Appl. Phys. 71, 6189–6191 (1992).
[15] Dow Corning Corporation, Midland, Michigan 48686-0994, Information AboutFOx-1x and FOx-2x Flowable Oxides, form no. 10-713a-98 edition (1997).
223
[16] Dow Corning Corporation, Midland, Michigan 48686-0994, FOx (R) FlowableOxide Baseline Processing Overview.
[17] Namatsu H., Takahashi Y., Yamazaki K., Yamaguchi T., Nagase M., & Kuri-hara K., Three-dimensional siloxane resist for the formation of nanopatternsswith minimum linewidth fluctuations, J. Vac. Sci. Technol. B 16, 69–76(1999).
[18] Nagase M., Namatsu H., Kurihara K., Iwadate K., Murase K., & Makino T.,Nano-scale fluctuations in electron beam resist pattern evaluated by atomicforce microscopy, Microelectronic Engineering 30, 419–422 (1996).
[19] Lacour D., Katine J.A., Smith N., Carey M.J., & Childress J.R., Thermal ef-fects on the magnetic-field dependence of spin-transfer-induced magnetizationreversal, Appl. Phys. Lett. 85, 4681–4683 (2004).
[20] Tanenbaum D., private communication, unpublished (2005).
[21] Fuchs G.D., Ph.D. thesis, Cornell University (in preparation).
[22] Gokan H. & Esho S., Pattern fabricaiton by oblique incidence ion-beam etch-ing, J. Vac. Sci. Technol. 18, 23–27 (1981).
[23] Yamauchi N., Yachi T., & Wada T., A pattern edge profile simulation foroblique ion milling, J. Vac. Sci. Technol. A 2, 1552–1557 (1984).
[24] Katine J.A., private communication, unpublished (2004).
CHAPTER 4
SYNTHETIC ANTIFERROMAGNETIC LAYERS IN MAGNETIC
NANOPILLARS
4.1 Introduction
The reversal of a thin ferromagnetic layer by application of a spin-polarized current,
or spin transfer effect (ST), has been extensively studied in systems with the famil-
iar Cofixed/Cu/Cofree current-perpendicular-to-plane (CPP) pseudo-spin valve
composition [1–4] as well as other magnetic trilayers [5, 6]. The prevailing theo-
ries [7, 8] indicate that the spin-polarized current applies a spin torque, local to the
Cu/Cofree interface, that can induce a dynamical response from the Cofree mag-
netization. Such dynamics, although important for the study of ST, are parasitic
for more passive devices such as CPP giant magnetoresistance (GMR) read heads,
where the Cofree layer is sensitive to stray fields from magnetic bits on a hard
drive medium [9]. ST-induced dynamics would give erroneous signals in nanoscale
devices and so it is advantageous to minimize the effects of a spin torque in such
devices.
Here I present the results of an experiment where a third, oppositely aligned
magnetic layer (Cobottom) has been incorporated into the CPP spin-valve struc-
ture adjacent to the Cofixed layer. I investigate field H and a DC current I -
induced switching of the Cofree layer in structures with layer composition Cu(100)/
Cobottom(11.5)/ Ru(0.7)/ Cofixed(8)/ Cu(6)/ Cofree(2)/ Cu(2)/ Pt(30) (in nm),
where all three Co layers are patterned in a nanopillar geometry. Interlayer ex-
change coupling through the thin Ru spacer gives a strong AP alignment of the
two adjacent Co layers [10]. The Cobottom/Ru/Cofixed trilayer thus forms a syn-
224
225
thetic antiferromagnet (SAF), where magnetostatic fields from the two Co layers
are in opposition and the overall dipolar coupling to the Cofree layer is reduced.
4.2 Experiment
All layers are DC sputter-deposited in a 1 mTorr Ar ambient onto thermally ox-
idized Si wafer substrates. Base pressures are ≤ 3 × 10−8 Torr. Electron beam
lithography, reactive ion etching, and evaporation are used to define a mask which
protects the underlying layers during an ion beam etch step. The etching is timed
to stop partway through the thick Cu buffer, patterning all three Co layers. SiO2 is
deposited with a plasma enhanced chemical vapor deposition process. Photolithog-
raphy, subsequent ion beam etching steps, and sputter deposition define top and
bottom leads in a 4-point CPP configuration. Resistance measurements are made
at T = 295 K using a Wheatstone bridge and lock-in amplifier technique with a 25
µA excitation current iex. Here, negative I indicates electron flow from the SAF
to the Cofree layer.
Figure 4.1(a) shows the device GMR (I ≈ 0) with H applied in-plane. The
continuing decrease in dV/dI at the maximum H is the gradual breaking of the
SAF AP alignment [11]. To distinguish between switching events for |H| < 1.0
kOe, I use a Stoner-Wolfarth simulation where total energy (Zeeman, anisotropy,
interlayer exchange, and dipole field) is minimized for all three layers at each 4
Oe increment in H. This simulation confirms the different magnetic configurations,
shown pictorially in Figure 4.1(a). For H ≈ 150 Oe the switch from high to
an intermediate resistance state corresponds to the reversal of the Cofree layer.
A smaller coercivity is expected for the Cofree layer due to its smaller shape
anisotropy Hshapean (∝ thickness). The jump back to high resistance at H ≈ 600 Oe
226
is the reversal of the Cobottom layer, which in turn switches the Cofixed layer via
the strong interlayer coupling. I do not observe any temporary deviation from AP
alignment within the SAF that may occur during this jump in resistance.
In Figure 4.1(b) the magnetic field is scanned over an asymmetric range, -1000
Oe < H < 300 Oe, in order to isolate the switching of the Cofree layer. The offset of
this hysteresis loop is taken as the dipole field on the Cofree layer |HSAFdip | ≈ 220
Oe. The two thicknesses of the SAF magnetic layers are chosen specifically to
minimize the combined dipole field halfway through the Cofree layer. Imperfections
in the magnetic layers during fabrication are most likely responsible for HSAFdip 6= 0.
Dipole field calculations from surface currents on an isolated magnetic disk (i.e. no
SAF pair) show |Hdip| ≈ 400 Oe halfway through the Cofree layer. The resistance
changes in Figure 4.1(b) shift in H as a bias I is applied (shown in Figure 4.3), a
further indication that the minor loop is the Cofree layer switching since the SAF
is too thick to be affected by the spin torque [12].
Looking at the resistance area product (∆R · A) from the GMR of 35 SAF
samples, I find (∆R · A)SAF = 0.45 ± 0.07 mΩ · µ m2. For 59 normal samples
without the Cobottom/Ru layers but with identical thicknesses for the rest of the
trilayer, I measure (∆R · A)normal = 0.94± 0.19 mΩ · µ m2, almost a factor of 2
larger. This reduction of ∆R ·A for the SAF samples is attributed to the reduced
polarization of the electrons that pass through and are reflected from the SAF
trilayer compared to the case of a single Co fixed layer. Both the Cobottom and
the Cofixed layers in the SAF are considerably thinner than the room temperature
spin-diffusion length of Co (`Cosf ≈ 38 nm [13]), and the Ru coupling layer is also
much thinner (0.7 nm) than its spin-diffusion length (∼14 nm [14]). Consequently,
all of the interfaces of the SAF play a role in the spin-filtering [15] and collectively
227
(b)Hdip
(a)
Cofree
Cu
SAF Cofixed
Cobottom
Figure 4.1: (a) High-field GMR of the SAF nanopillar. Simulated devices areshown to indicate the alignment of the three magnetizations as the field is sweptfrom negative to positive values. (b) Low-field GMR of the same device. The fieldis scanned asymmetrically to isolate the switching of the Cofree layer. The dipolar
field (HSAFdip ≈ −220 Oe) is taken as the offset of the hysteresis loop, indicated by
the dashed line.
228
determine the net spin polarization of the current that impinges onto the Cofree
layer in these near-ballistic ST devices.
4.3 Discussion
While the spin-filtering that results from the electronic structure of the two Co/Ru
interfaces [16, 17] and any bulk spin-dependent scattering that does occur can be
expected to modify the effect, the two oppositely aligned magnetizations of the
SAF pair will clearly reduce the spin current amplitude that passes through or,
depending on the bias current direction, reflects off the SAF. Since the magnetore-
sistance signal ∆R ·A is, in the ballistic limit, linearly dependent upon the effective
polarization ηeff of this current, the reduced magnetoresistance signal from SAF
devices indicates that ηeff ≈ 12ηCo where ηCo is the polarization produced by the
spin filtering of a single fixed Co layer.
I note that the ∆R ·A for normal samples here is larger than for those reported
in [12]. I suspect that this difference is due to the fact that the samples here and
those in [12] were prepared in different sputter systems which yield multilayer films
with different interfacial qualities. The Co layers in this study had 37% larger grain
sizes (from x ray diffraction measurements) and larger rms interfacial roughness
(∼ 3×, from AFM measurements) than those in [12]. A detailed understanding of
the role of interfacial quality on ∆R · A is still lacking, however.
Not surprisingly, I find that the ST switching is also affected by the reduced
ηeff from the SAF. In Figure 4.2(a) I show a ST scan for a SAF sample at low field
(I ramp rate = 0.5 mA/sec) and a similar scan from a normal sample (1.0 mA/sec)
Cofixed(40)/ Cu(10)/ Cofree(3)/ Cu(2)/ Pt(30) (in nm), where the Cofixed layer
is unpatterned, in Figure 4.2(b). I plot the current density J normalized to the
229
(a)
(b)
-150 Oe
-200 Oe
Figure 4.2: dV/dI versus current density normalized to the free layer thicknessJ/t for (a) a 90 × 190 nm elliptical SAF sample and (b) a 70 × 130 nm ellipticalnormal sample. The resistance values in (b), a two-point measurement, includelead resistance ∼6 Ω and contact resistance ∼9 Ω.
Cofree layer thickness t because this is the value most directly related to the spin
torque [12].
For four SAF samples, I measure ∆Jc/t = 7 ± 1 × 107 A/(cm2× nm), while
for 24 normal samples, ∆Jc/t = 3.0 ± 1.0 × 107 A/(cm2× nm), an increase by a
factor of ∼2.3. Here, ∆Jc ≡ J+c −J−c , where J+
c (J−c ) is the critical current density
for switching Cofree P to AP (AP to P) with Cofixed. There is a small difference
in the Cu spacer thickness between the SAF and normal devices, but this would
230
only account for a 2% change in ηeff, which is well within our uncertainties. The
polarization of the conduction electrons that exert a spin torque on the Cofree layer
may depend on the direction of the current flow. For J−, electrons traverse the fixed
layer (single Co or SAF) and are thereby spin filtered to produce a current with
polarization η− that impinges on the Cofree layer. For J+, the incident electrons
that are spin-filtered by the Cofree layer traverse the Cu spacer and impinge onto
the fixed layer. From there a portion are reflected back to the Cofree layer, after
being re-polarized η+ by the spin-filtering effects of the fixed layer (single Co or
SAF), and exert a spin torque on the Cofree layer. For simplicity I assume that the
effective polarization of the electron current exerting a spin torque on the Cofree
layer is the same in both cases, ηeff = η+ = η−.
From [8], ∆Jc/t ∝ α[g(0, η)−1 + g(π, η)−1]. Here α is the Gilbert damping
parameter and g(θ, η) is a measure of the spin torque that is exerted on the free
layer as a function of its alignment with the fixed layer and is a monotonically
increasing function of η. Assuming ηeff ≈ 0.4 and 0.2 for normal and SAF devices,
respectively, and that α is the same for both types of devices, I plug these ηeff values
into the g(θ, η) expression derived by Slonczewski [8] and find(∆Jc/t)SAF
(∆Jc/t)normal≈ 2.5,
in reasonable agreement with the data.
I investigate the dependence of the ST I -H phase diagram on ηeff by measuring
the coercivity Hc of the Cofree layer as a function of J/t for both normal and
SAF samples, shown in Figure 4.3. The normal samples (same as those shown in
Figure 4.2(b)) have an unpatterned Cobottom layer which has a naturally smaller
Hc, making it simpler to identify the Cofree layer switching. These plots mark
the respective boundaries between the bistable P/AP and P regions, as measured
in other experiments with non-SAF samples [3, 18, 19]. The important point of
231
Field (Oe)dV
/dI (Ω
)
P/AP
P/AP
P
P
Field (Oe)
dV
/dI
(Ω)
Figure 4.3: Hc plotted versus J/t, the current density normalized to the Cofreelayer thickness, for normal (•) and SAF () samples. The data illustrate theboundary separating the bistable P/AP region and the stable P region as shown.Insets show how Hc was measured.
Figure 4.3. is that the slope of Hc vs. J/t is much larger for normal samples than
for SAF samples, highlighting the weaker influence, on the Cofree reversal, of the
reduced ηeff in SAF samples. Spin torque-induced excitations of a nanomagnet
have been described by thermal activation models where either the effective barrier
height or the temperature is modified by ηeff [6, 20, 21] and so a reduced ηeff
correspondingly has a weaker influence on the activation process.
232
4.4 Summary
In summary, Cobottom/Ru layers were added to the familiar Cofixed/Cu/Cofree
CPP magnetic nanopillar composition. The Cobottom and Cofixed layers are
AP due to exchange coupling through the Ru spacer and succeed in reducing
unfavorable dipole fields on the Cofree layer. It is clear that these AP magnetic
layers also reduce the spin polarization of I from the bulk Co value or single
Co spin-filter value. Such reduction of the current polarization is advantageous for
nanoscale devices seeking to reduce the effects of a spin torque, such as CPP-GMR
read heads [9], where the reduction in ∆R ·A due to the SAF can be countered by
partially oxidizing the magnetic interfaces [22].
233
References for Chapter 4
[1] Katine J.A., Albert F.J., Buhrman R.A., Myers E.B., & Ralph D.C., Current-driven magnetization reversal and spin-wave excitations in Co/Cu/Co pillars,Phys. Rev. Lett. 84, 3149–3152 (2000).
[2] Albert F.J., Katine J.A., Buhrman R.A., & Ralph D.C., Spin-polarized cur-rent switching of a Co thin film nanomagnet, Appl. Phys. Lett. 77, 3809–3811(2000).
[3] Grollier J., Cros V., Hamzic A., George J.M., Jaffres H., Fert A., FainiG., Youssef J.B., & Legall H., Spin-polarized current induced switching inCo/Cu/Co nanopillars, Appl. Phys. Lett. 78, 3663–3665 (2001).
[4] Wegrowe J.E., Kelly D., Jaccard Y., Guittienne P., & Ansermet J.P., Current-induced magnetization reversal in magnetic nanowires, Europhys. Lett. 45,626–632 (1999).
[5] Sun J.Z., Monsma D.J., Abraham D.W., Rooks M.J., & Koch R.H., Batch-fabricated spin-injection magnetic switches, Appl. Phys. Lett. 81, 2202–2204(2002).
[6] Urazhdin S., Birge N.O., Pratt W.P., & Bass J., Current-driven magneticexcitations in permalloy-based multilayer nanopillars, Phys. Rev. Lett. 91,146803 (2003).
[7] Berger L., Emission of spin waves by a magnetic multilayer traversed by acurrent, Phys. Rev. B 54, 9353–9358 (1996).
[8] Slonczewski J.C., Current-driven excitation of magnetic multilayers, J. Magn.Magn. Mater. 159, L1–L7 (1996).
[9] Nagasaka K., Seyama Y., Kondo R., Oshima H., Shimizu Y., & Tanaka A.,CPP operational mode of GMR head, Fujitsu. Sci. Tech. J. 37, 192–200(2001).
[10] Parkin S.S.P. & Mauri D., Spin engineering: Direct determination of theRuderman-Kittel-Kasuya-Yosida far-field range function in ruthenium, Phys.Rev. B 44, 7131–7134 (1991).
[11] Parkin S.S.P., More N., & Roche K.P., Oscillations in exchange coupling andmagnetoresistance in metallic superlattice structures Co/Ru, Co/Cr, andFe/Cr, Phys. Rev. Lett. 64, 2304–2307 (1990).
[12] Albert F.J., Emley N.C., Myers E.B., Ralph D.C., & Buhrman R.A., Quanti-tative study of magnetization reversal by spin-polarized current in magneticmultilayer nanopillars, Phys. Rev. Lett. 89, 226802 (2002).
234
[13] Piraux L., Dubois S., Fert A., & Belliard L., The temperature dependence ofthe perpendicular giant magnetoresistance in Co/Cu multilayered nanowires,Eur. Phys. J. B 4, 413–420 (1998).
[14] Eid K., Fonck R., Darwish M.A., W. P. Pratt J., & Bass J., Current-perpendicular-to-plane-magnetoresistance properties of Ru and Co/Ru in-terfaces, J. Appl. Phys. 91, 8102–8104 (2002).
[15] Upadhyay S., Louie R.N., & Buhrman R.A., Spin filtering by ultrathin ferro-magnetic films, Appl. Phys. Lett. 74, 3881–3883 (1999).
[16] Rampe A., Hartmann D., Weber W., Popovic S., Reese M., & G G. .
[17] Zhong L. & Freeman A.J., Induced magnetism of 4d transition metals: Rhand Ru/Fe(001) overlayers and sandwiches, J. Appl. Phys. 81, 3890–3892(1997).
[18] Myers E.B., Albert F.J., Sankey J.C., Bonet E., Buhrman R.A., & Ralph D.C.,Thermally activated magnetic reversal induced by a spin-polarized current,Phys. Rev. Lett. 89, 196801 (2002).
[19] Kiselev S.I., Sankey J.C., Krivorotov I.N., Emley N.C., Schoelkopf R.J.,Buhrman R.A., & Ralph D.C., Microwave oscillations of a nanomagnet drivenby a spin-polarized current, Nature (London) 425, 380–383 (2003).
[20] Li Z. & Zhang S., Thermally assisted magnetization reversal in the presenceof a spin-transfer torque, Phys. Rev. B 69, 134416 (2004).
[21] Koch R.H., Katine J.A., & Sun J.Z., Time-resolved reversal of spin-transferswitching in a nanomagnet, Phys. Rev. Lett. 92 (2004).
[22] Nagasaka K., Seyama Y., Varga L., Shimizu Y., & Tanaka A., Giant mag-netroresistance properties of specular spin valve films in a current perpendic-ular to plane structure, J. Appl. Phys. 89, 6943–6945 (2001).
CHAPTER 5
PERFORMING TIME-RESOLVED SPIN TORQUE-DRIVEN
SWITCHING MEASUREMENTS AT VARIABLE T
5.1 Introduction
This chapter describes how to perform time-resolved measurements of spin torque-
driven phenomena at controlled bath temperatures T . The breadth of techni-
cal detail inherent to such a complicated measurement is considerable, from the
hardware involved to perform the measurement to the software written for data
analysis. Here, I document the intricate, custom procedures used in this kind of
measurement that may be useful for similar projects in the future.
5.2 Hardware setup
For time-resolved spin torque-driven switching measurements at variable T , we
are merging two non-trivial measurement techniques — high speed electronics and
cryogenics. In this section, I discuss the topic of making a reliable high frequency
4-point1 contact that can withstand essentially unlimited thermal cycling. This
includes construction of a high frequency dipper, designs for custom sample holders,
and a detailed procedure for making electrical contacts by Au ribbon bonding.
In order to achieve reliable contacts for the time-resolved switching measure-
ments of this experiment, it was decided to use bonded contacts as opposed to
probes. Using bonds, however, demanded use of an entirely different cryostat
setup as the bond contact, which is essentially a permanent contact to the sample,
1Using a 2-point connection with a directional couple was attempted, but toomuch signal distortion was introduced by the coupler.
235
236
was incompatible with the probe contact cryostat (Desert Cryogenics Tabletop). A
bond-compatible dipper was made for a separate cryostat (Janis STVP-300). The
permanent contact of the bonds mandated the design and constructing of special
sample holders to mate with this new dipper. A schematic of the setup is shown in
Figures 5.1, and detailed instructions on how to duplicate these pieces of hardware
are discussed in the following sections.
5.2.1 Construction of the high frequency dipper
This high frequency dipper is used with a Janis cryostat (model STVP-300), of
which there are two in the Clark Hall basement2 Construction of a high frequency
dipper requires several pieces of hardware, listed in Table 5.1. A schematic of the
Janis cryostat with the dipper inserted is shown in Figure 5.1. The top flange
on the Janis cryostat is a 1.5” sanitary flange. An adaptor between this sanitary
flange and a KF flange must be machined for two reasons. First is that sanitary
flange hardware (O-ring seals and clamps) are less prevalent within the lab and
will be more difficult to replace if a set of O-ring seals or clamps are lost. Secondly,
the adaptor allows a place to mount the plug for the thermometer that measures
sample temperature. To machine the parts for the adaptor, first shorten the two
stainless steel (S.S.) pipes to 2.5” each, machined on both ends. Cut a 3/4” hole
in the side of the 1.25” diameter S.S. pipe, symmetrically placed along its length
and diameter. Cut a 3/4” hole in the S.S. NW-25 KF blank-off, and a 1.25” hole
in both the S.S. NW-40 KF blank-off and the S.S. 1.5” sanitary end cap. Note
that the 3/4” pipe may need to be cut slightly concave at one end to be flush with
the inner wall of the 1.25” pipe. Have the pro-shop weld all the pieces together as
2A third has been purchased and should be available in early 2006.
237
Figure 5.1: This schematic shows the general layout of the dipper construction.
shown in Figure 5.2.
To construct the dipper, first drill two (44 drill bit) holes in the brass NW-40
KF blank-off, placed symmetrically about the center, separated by approximately
5/8”. Machine the 1” diameter teflon rod to the specifications in Figure 5.3. This
teflon piece serves as a spacer that will guide and stabilize the high frequency semi-
rigid coaxial lines (henceforth coax lines) while inside the cryostat. It will also be
helpful in future soldering steps. The holes drilled in the teflon for the coax lines
(0.085” diameter) should fit them snugly.
The geometry of the Janis is such that the input and output coax lines lie on
the same side of the sample. This requires that one of the lines be bent by 180.
238
Table 5.1: Parts needed to construct a high frequency dipper. All parts, asidefrom the Cernox thermometer, can be found in the Physics Stock room. A propanetorch is available in the Center for Nanoscale Systems (CNS) shared facility.
Brass NW-40 KF blank-off
Brass NW-25 un-bored KF stub
Stainless steel NW-40 KF blank-off
Stainless steel NW-25 KF blank-off
Stainless steel 1.5” sanitary end cap
Stainless steel 3/4” pipe, 1/16” thick
Stainless steel 1-1/4” pipe, 1/16” thick
Semi-rigid coaxial cable (85-S.S.)
Flux (Eutectic Castolin Flux 157)
Non-rosen core (i.e. solid wire) solder
44 drill bit (for brass)
Propane torch
K connectors (K101M-085 from Anritsu)
Teflon 1” diameter rod
8-pin hermetically sealed military grade connector
Cernox thermometer (CX-1070-AA-4L from LakeShore)
Low thermal conductivity insulated stainless steel wire
239
Figure 5.2: This schematic shows how the parts of the adaptor should be weldedtogether. The top of the adaptor will eventually be capped by the brass NW-40while the bottom is what mates to the top of the Janis cryostat. The NW-25 iswhere the plug for the thermometer, which will be soldered to the brass NW-25stub, will be connected.
240
Figure 5.3: Machine the teflon piece to these dimensions. The 0.050” holes areto make sure there is an unobstructed path for the He gas to escape. The 1/8”slot is a place for the wires connecting the thermometer to go. The 0.085” holesare where the semi-rigid coaxial lines go. The fit must be snug.
Since bends in the coax lines may introduce signal reflections and loss, it is best to
place the 180 bend on the incoming side of the sample so that any loss affects only
the incoming pulse without reducing the switching signal from the nanomagnet on
the outgoing side of the sample. The incoming pulse will be calibrated once the
dipper is assembled, so losses due to the bend will be accounted for. Take two
sections of 85-S.S. semi-rigid coaxial cable (each ∼3 feet in length), designate one
as the input line, and bend one end into a 180 loop with a diameter ∼3/4”. Bend
the S.S. with fingers (if possible) or with pliers, taking care not to kink the lines
as these will add to signal loss. Available space for this 180 bend is very limited
inside the cryostat, so it is best to have the curvature extend into 3-dimensions,
as shown in the images in Figure 5.4.
Solder a K101M-085 connector to the end of the input coax line which has the
241
(a)
(b)
Figure 5.4: Images of a working 4-point contact high frequency dipper. (a) Ablank holder with the Cernox thermometer inserted into its holder, the output lineon the left, and the curved input line on the right. Note that the 180 curve issmooth, with minimal sharp bends or severe kinks. This requires bending the coaxline into 3-dimensions. (b) Opposite perspective from (a) but now with the teflonspacer shown on the right.
242
180 bend. The instructions for this are in the same drawer as the K101M-085
connectors themselves in the CNS shared facility. Solder a second K101M-085
connector on the (unbent) output coax. Feed the unsoldered ends of the input and
output coax lines through the teflon spacer and then through the 44-sized holes in
the brass NW-40 blank-off. Attach a blank device holder between the two K101M-
85 connectors so that their relative positions are correct for soldering. To establish
the proper absolute position of the two coax lines, slip the welded adaptor around
180 loop and the 1” teflon spacer and quick-connect it to the brass NW-40 end
cap. Lower the dipper assembly into the cryostat where it should be obvious when
the lowest point of the dipper (namely at the 180 bend) hits the bottom of the
cryostat. Clamp the adaptor onto the cryostat top, then extract the two coax lines
about a 1/4”. With a sharpie, mark this position on the coax lines and remove
the entire dipper from the cryostat. These marks represent the proper positioning
of the coax lines for soldering.
To solder, remove the adaptor and clamp the teflon piece in a vice so that the
coax lines hang vertically. Position the teflon spacer high on the coax lines so
that as it rests in the jaws of the vice, the brass NW-40 blank-off lines up with
the sharpie marks indicating the proper positioning of the solder joints. Clean off
the sharpie marks with IPA. Paint some flux onto the exposed coax line and the
surrounding brass, then place a short pieces of solder, bent into a donut shape,
around the coax line where it juts out from the brass. The propane torch is used
to melt the solder. However, the dielectric within the coax line can expand and
rupture the outer wall with too much heat load. To avoid this, sweep the tip of the
propane flame across the soldered area, never leaving the direct flame focused on
the coax line for more than ∼0.5 sec, and uniformly heat the brass. The solder will
243
eventually melt and coat both the S.S. and the brass, forming a good conductive
bond. Repeat for the other coax line. When both lines are soldered, clean the
area with IPA, and then indicate somehow which line is the input and which is the
output. On the two ends of the coaxial lines protruding from the brass, solder K
or 2.4mm male connections. The connectors for the sample holder are already K,
so there is no benefit in frequency range by going to 2.4mm, although the majority
of low-loss high frequency cables in the CNS shared facility are 2.4mm.
The next step is to solder the hermetically sealed plug. Bore a hole in the
brass NW-25 un-bored KF stub large enough so that the 8-pin hermetically sealed
military grade connector fits loosely in the hole. Paint some flux onto the ridge of
the bored hole and place a line of solder around the intersection of the KF stub
and the connector. Place these pieces on a Corning hotplate with the heat setting
at 5. After a few minutes the solder will melt. Remove the piece from the hotplate
and cool with an IPA spray. Follow the instructions that come with the Cernox
thermometer to solder it to the hermetically sealed connector via the low thermal
conductivity S.S. wire. Kwik-connect the thermometer plug and the soldered coax
lines to the adaptor.
5.2.2 Custom high frequency holders
Design
Custom samples holders are necessary in order to mate with the input and output
coax lines on the Janis dipper described above. Machine plans for this holder are
shown in Figures 5.5, 5.6, 5.7, 5.8, 5.9, 5.10, and 5.11. All parts are made out of
brass. The 0.2” rectangular groove down the middle of the holder is where the
sample, along with two coplanar waveguides (CPW) on either side, will be placed.
244
The CPW, which is necessary to propagate the high frequency signal to and from
the sample, is a piece thin piece of plastic (Duroid 0.010” thick) sandwiched in
between two slabs of Au-plated Cu, each ∼0.002” thick. On one side of the CPW,
what is referred to as the “top” side, the metal plane is subdivided into three
separate sheets, two grounding sheets surrounding a narrower signal strip. The two
ground sheets are connected to each other electronically by vias to the continuous
metal plane on the “bottom” side of the CPW. The designs for the CPWs, which
are necessary to submit an order from the manufacturer (Modular Components
National, Inc. www.modularcomp.com) are shown in Figures 5.12 and 5.13. If
a CAD version of the designs is required, they are located in a folder on the
simulation computer in the CNS shared facility C:/ Documents and Settings/
Buhrman/ My Documents/ Nathan/ HF-hardware/.
245
Groove
Input connection hole
Figure 5.5: Machine drawings for sample holder. Perspective. Brass.
246
Sid
e vie
w
1.0
92"
0.0
56"
0.7
52"
0.3
40"
0.2
"+
0.0
03"
-0.0
"(
)
dim
ensi
ons
sym
met
ric
wit
h l
eft
side
0.1
20"
+/-
.001"
0.1
25"
+/-
.001"
0.1
40"
+/-
.002"
0.2
75"0.0
29"
0.0
91"
+/-
0.0
02"
0.0
32"
+/-
0.0
03
dia
,
thru
to c
avit
y
0.0
28
" +
/- 0
.001 d
ia, th
ru
0.0
78"
+/-
0.0
01 d
ia, to
dep
th
conce
ntr
icit
y t
o w
ithin
+/-
0.0
01
Figure 5.6: Machine drawings for sample holder. Side view. Brass.
247
square corner
with cutter
1-72 tapped (4)
2-56 tapped (4)
Top view
Figure 5.7: Machine drawings for sample holder. Top view. Brass.
248
0.05"
0.075"
+/- 0.001
0.555 +/- 0.001"
0.58"
0.630"
0.415 +/- 0.003"
0.315 +/- 0.001"
0.215 +/- 0.003"
End view
Figure 5.8: Machine drawings for sample holder. End view. Brass.
249
Figure 5.9: Machine drawings for sample cover. Perspective. Brass.
250
0.2
44
"
0.6
56
"
Bottom view
0.9
00
"
0.05"
0.10"0.53"
0.58"
0.63"
1-72 clear (4)
Figure 5.10: Machine drawings for sample cover. Side view. Brass.
251
0.0
5"
0.1
5"
End view
Figure 5.11: Machine drawings for sample cover. End view. Brass.
252
Figure 5.12: Coplanar waveguide (CPW) purchased from Modular ComponentsNational, Inc. with the following dimensions shown here (in mm). Include thefollowing information with the purchase order: Coplanar waveguide on Duroid6002 (10 mil thick). Both sides copper clad with gold-plating (thickness 1 mil).Top layer patterned with CPW: central trace 410 micron, gap 90 micron. Viasfrom the top grounds have to be made to insure a good contact with the bottomground (see CAD for placement). Total thickness approximately 14 mils. VIAsshould be on both ground planes, with mirror symmetry about the CPW. 10 mildiameter VIA.
253
Top ground sheets
Signal strip
End view
Bottom ground plane Vias
(a)
(b)
Figure 5.13: (a) A perspective and (b) end view of the CPW to help in visual-ization.
254
Assembly
All the parts necessary for assembling the custom holder are shown in Table 5.2.
The K103F flange connectors mate to the S.S. coax lines of the dipper. The signal
is conducted from the flange connector through the K100-10 glass bead conductor
and K110-1 sliding contacts to the CPW. Au ribbon bonds link the CPW to the
sample. This chain of connections is shown in Figure 5.14.
The first step in assembling these holders is to solder the CPWs onto the holder.
Place a holder on a corning hotplate and set the temperature to setting 5. Place a
drop of flux and solder in the 0.2” groove and wait for the solder to melt. Once it
melts, paint the solder around with the flat end of a swab and square the corners
on all sides of the groove. Be sure not to accidentally cover the hole for the input
connection with solder. Remove from hotplate and spray cool with IPA. Rub the
groove with cotton swab soaked with IPA to remove the flux residue. Place some
flux in the now solder-covered groove and place one CPW in the groove with one
end abutting the corner so that one end of the signal strip is adjacent to the input
connection hole. Orient the CPW so that the end with vias further from the edge
faces the sample. With the flat end of a swab, gently press down on the CPW,
pushing it into the corner as much as it will go. Push down on the two ground
sheets, not the signal strip so as not to unnecessarily damage it. Put the holder
back on the hotplate and wait for the solder to melt. Remove from hotplate and
cool with IPA spray. Check the holder under the microscope to see if solder has
crept up the edges of the grounding planes. Check also that the waveguide is
abutted to the corner of the holder. A small gap between the signal strip and the
input connection hole of < 100 µm is acceptable. Repeat this procedure for the
other CPW at the opposite end of the holder.
255
Figure 5.14: This schematic shows the linkage of electrical connections from theK103F flange mount to the CPW. The center conductor fits around one end of thepin from the glass bead. The sliding contact fits around the other end of the pinand contacts the signal strip of the CPW via the tab.
256
Table 5.2: Parts needed to construct one high frequency holder. Part numbersare from Anritsu.
Custom machined brass holder
Custom machined brass cover
Two custom coplanar waveguides (CPW)
Two K110-1 sliding contacts
Two K100-10 glass beads
Two K103F flange connectors with center conductors
01-306 Glass bead supporting fixture
Solder
Flux (Superior No. 30)
The next step is to install the glass bead. Rim the hole of the holder with flux
first, then place the bead in the hole with the longer end of the pin facing out.
Use the glass bead supporting fixture to hold the conducting pin of the glass bead
concentric with the input connection hole. Cut a short strand of solder and place
it in the top hole above the bead. Place entire holder on the hotplate and wait
for the solder to melt and bubble for a few seconds. Remove from hotplate and
carefully cool the holder with IPA, just spraying small amounts of IPA onto the
flat brass portions of the wafer, waiting for them to evaporate, and repeat until
cool to the touch. Cooling too quickly can crack the glass bead. Yank lightly on
the conducting pin to see if the glass bead has been soldered successfully. The
solder does not need to coat all of the bead, only hold it secure. Slip the center
conductor over the pin of the glass bead protruding from the holder, then slide
the K103F flange connector over the center conductor and screw tightly onto the
257
holder. The threaded holes on the holders are too shallow and two to three threads
will need to be cut off from K103F mounting screws (2-56). Repeat for the other
side of the holder.
An Anritsu K110-1 sliding contact is a cylindrical shell with a tab at one end
and is necessary to connect the pin with the signal strip of the CPW. This piece
has a cylindrical shell which fits tightly over the pin and a flat tab that makes
contact with the signal strip of the CPW. These steps of the holder assembly are
best done under a low-magnification optical microscope. First place the sliding
contact on the signal strip, tab side down and facing the sample. With a small flat
head screwdriver, push the cylindrical shell of the sliding contact onto the pin. The
friction contact between the two is sufficient for quality high frequency conduction.
Mash the tab down if it is not already in contact with the CPW signal strip.
Check conductivity between the CPW signal strip and the ground planes to make
sure the two are isolated. To solder the sliding contact to the CPW, place a small
amount of flux on top if the tab of the sliding contact. Use only a very small
amount of solder, ∼1/4 of a thin (∼200 µm) cylindrical sliver of solder, and drop
it on top of the flux. Using too much solder could spill over CPW and short the
signal strip to the grounding planes. The delicate soldering job should be done
with a very fine soldering iron (set to medium power) or a hot finger. Clean the
hot finger by heating and rubbing the tip against the bare brass. Then touch the
hot finger to the signal strip in proximity to the sliding contact tab. The solder
should melt very quickly but make sure it coats the tab and a small region of the
signal strip. Check Ohmic connections and repeat for the other CPW.
258
Sample mounting
First clean the solder-covered surface between the two CPWs, where the sample
will sit, with an IPA-soaked swab. Scribe the sample into the chip size to be
mounted on the holder. With careful scribing, the chip can be diced into single
devices. This is not recommended, however, because it is harder to control a single-
device chip during the scribing and orientation of the device can easily be lost. It
is best to scribe chips having two or four devices because a bigger chip is less likely
to be popped off during the bonding process.
Apply a small amount of silver paint to this region with the brush-like edge of
a broken wooden swab. Place the chip on the silver paint and press down with the
scriber. Wait a few minutes for the silver paint to dry, then test its strength by
pushing very lightly on a side of the chip with the flat end of a swab. Alignment
of the sample is purely visual. If the adhesion strength is not strong enough, the
chip can be ripped off the holder during the bonding process. Once the chip is
adhered, use a pressurized air can to lightly blow off any dust from the sample.
Now the sample is ready to be bonded.
5.2.3 Procedure for Au ribbon bonding
Bonds versus probes
Using high frequency probes is a very common practice. However, for time-resolved
measurements, positional shifts in the probes lead to drifts in the measured sig-
nal due to changing reflection and transmission coefficients of the high frequency
contact [1]. For persistent signals, such as those shown in ref. [2], signal drift and
other distortions are correctable because the envelope of the signal is expected to
259
be fairly constant over the duration of the measurement. However, drifts are gen-
erally not correctable for transient signals and the transient switching data in [2]
came from scans which suffered no shifting contacts. A measurement showing such
a distortion is shown in Figure 5.15. The undistorted pulse shape (Figure 5.15(a))
was measured with the sample contacted by ribbon bonds. It shows a rapid rise
time (∼200 psec), some abberations coming from the pulser, cables, and amplifier
(shoulder in the step edge and a slowly oscillating background) which prevent ideal,
square-step-edge behavior, with an approximately flat amplitude for the duration
of the pulse. Although the distorted pulse (Figure 5.15(b)) shows similar turn-on
behavior, the amplitude continues to rise over the pulse duration. This upwards
drift in amplitude has been attributed to unstable probe contacts.
Before bonding
Ribbon bonding is done on the Marpet Enterprises Inc. (MEI) wire bonder located
in the clean TOL facility in D-corridor in the Clark Hall basement. An image
identifying important parts of the bonding tool is shown in Figure 5.16. Bonding
is achieved by providing a brief pulse of acoustic energy to a wire which then
adheres to the sample. The sonicating head resides at the end of a boom that tilts
downwards to contact the sample. This tool must be operated with two hands.
The left hand moves the stage platform, which positions the sample underneath
the head. On the stage platform control handle, there is a silver button that
automatically drops the boom. This automatic bonding action is uncontrolled and
should be avoided for ribbon bonding. Manual lowering of the head is done by a
lever controlled by the right hand.
The tool is typically set up as an Al wire bonder and so the Al spool must
260
time [ns]
ribbon bonds
probe contacts
(a)
(b)
Figure 5.15: The measured pulse shape from (a) electrical connection to a sampleby Au ribbon bonds, showing an approximately square step and a flat amplitudeand (b) probe contacts, showing a pulse with an increasing amplitude over thepulse duration.
261
hole
cotton
posts
clampsonicating head
boom
spool spindle
"V" hook
Figure 5.16: This is an image of the bonding part of the wire/ribbon-bonder toolin the clean TOL room in Clark Hall basement. Important parts of the bonderthat are referenced in the bonding procedure text are identified here.
262
be removed before and then replaced after Au ribbon bonding is finished. One
should allot at least 2 hours to change out the Al for the Au. Two tools will prove
extremely useful for this procedure. First is a fine set of reverse-action tweezers,
which I find are much easier to use for threading the Au ribbon than normal
tweezers. Such a pair is presently a tool at the MEI bonder. Second is a set of
fine sewing scissors with very narrow shears. Scissors on a Swiss Army Knife work
very well. Larger scissors are insufficient since the width of the shear is too large
and will bend the Au ribbon during cutting.
Static discharge across the sample is always a concern, but there are two very
important steps one can take to minimize this risk. First is to not use a fabric
covered chair. These chairs can generate static electricity merely by moving in the
seat. Use an all metal or plastic covered seat. Secondly, one must be attached
to ground at all times when the sample is out of its anti-static bag. Attach the
grounding wrist strap to a metal screw on the light source to the left of the MEI
bonder and wear it for the duration of the bonding procedure.
Configure the bonder for Au ribbon
Nitrile gloves are much easier to work with than vinyl when exchanging the spools.
As it is wound up like a spring, the Al wire has a lot of tension in it and so if
the wire end becomes unanchored, the tension will cause the wire to unravel on
the spool, rendering it unusable. To remove the Al wire spool, place one finger
on the wire end, pinching it against the rest of the coiled wire. With the other
hand, remove the cotton piece and open the clamp by toggling the switch marked
“CLAMP” on the left hand side of the tool.
With both hands, each alternately holding down the free end of the Al wire,
263
Figure 5.17: This is an image of the back of the Al wire sonicating head. Theround hole is specific for cylindrical wire.
rotate the Al wire spool so as to retract it from the sonicating head. Once the wire
is clear of the posts, remove the spool from the spindle and scotch tape the free
end of the Al wire to the spool so that it does not unravel. Before the Au spool can
be mounted, the sonicating head must be changed. The reason for this is because
the Al wire is cylindrical and so its sonicating head has a cylindrical hole to guide
the wire underneath the tip of the sonicating head. Au ribbon is rectangular and
so a special Au sonicating head with a rectangular hole or via must be installed.
Images of the different sonicating heads are shown in Figures 5.17 and 5.18.
The sonicating head is held in place by a single set screw at the tip of the boom.
Once this screw is loosened the sonicating head will fall so be sure to have a grip
264
via
Figure 5.18: This is an image of the back of the Au ribbon sonicating head. Therectangular hole, or via, is specific for ribbon.
265
on the sonicating head before loosening the set screw. Replace the Al sonicating
head with the Au one, making sure the flat part of the sonicating head is facing
out, and position the head so that it protrudes from the boom by ∼0.5”. More
precise positioning of the sonicating head will occur later once the Au ribbon spool
is mounted. The set screw should be finger tight.
Au is more ductile than Al and so there is less of a risk of the Au ribbon
unravelling on the spool. Nonetheless, it is safest to handle Au and Al with the
same care and not let the free end unravel. There should be a piece of scotch tape
holding the Au wire to the spool. Remove the tape and place the Au spool on
the spindle oriented with the free end on top of the spool. Position the Au ribbon
between the two posts and thread it through the hole in the boom. The hole has a
slight lip on which the Au ribbon can catch and bend. Once inside the hole, gently
rotate the spool to forward the Au ribbon through. Once the free end appears on
the bottom side grab it with the tweezers and pull enough Au ribbon through so
that it fits around the front of the “V” hook and lies within the grips of the clamp.
Toggle the “CLAMP” switch again to close the clamp grips on the ribbon.
Take the piece of cotton and sharpen its tip by rolling it in between your thumb
and fore finger. If there are any stray strands of cotton, cut them off with the fine
scissors. Feed the cotton into the hole and position it in between the two posts,
adjacent to the Au ribbon. Since the Au ribbon has been handled by tweezers,
it will be slightly disfigured, so open the clamp and pull the free end through a
few inches then re-clamp. Snip off the end with about 1” of Au ribbon protruding
from the hole.
Grip the free end of the Au ribbon in the tweezers and, with steadiness of hand,
aim the tip of the free end into the rectangular via of the sonicating head. This is
266
the most difficult part of the Au ribbon mounting procedure, and it is likely that it
will need to be repeated several times during the ribbon bonding of a sample. Once
the ribbon is successfully through the via, pull it out from the front side and snip
off the ribbon with about 0.25” protruding from the sonicating head. Bend the
ribbon so that it points upward at a ∼90 angle to the ribbon suspended between
the hole in the boom and the via in the sonicating head.
Here is where fine positioning of the sonicating head should be done. Essen-
tially, the goal is to minimize the bending of the Au ribbon between the hole in
the boom and the via in the sonicating head. Make sure to handle the sonicating
head in the same way as mentioned above if repositioning is necessary. Once the
head is positioned correctly, ribbon bonding to the sample can now commence.
Ribbon bonding the sample
Au ribbon has a rectangular cross section of 50.8 µm × 12.7 µm. The bonder
works by making an initial bond to the surface followed by a second bond after
which the ribbon is severed. Bonder settings should be as follows. Set the weight
of the boom to maximum, so that the weight is drawn out towards the user as
much as possible. Ch.1 sets the first bond (on the CPW) and should have power P
= 10.0 (maximum) and time t = 10 (maximum). Even though power and time are
maximum, this does not produce the most secure bonding conditions as the Au
ribbon sometimes pops off. Two very important tips to make secure bonds are the
following. First is to “paint” the ribbon slightly with the sonicating head. Painting
is done just after the sonicating burst by, with the head kept in contact with the
surface, gently pulling the sample out (towards the user) by only about 1–2 times
the ribbon thickness (∼ 15µm) so that the weight of the boom mashes more Au
267
ribbon into the surface. Too much motion, however, and the head will pinch the
ribbon off. Second, after the bond is made, raise the boom tilt lever slowly with
the right hand because the clamp takes a few tenths of a second to open once the
lever is raised. If raised too fast, this delay could cause the unopened clamp to
pull off the Au ribbon before it has a chance to open. When done correctly, the
sonicating head should still be in contact with the surface when the clamp opens.
Ch.2 sets the second bond (on the sample bonding pads) and should be set
at power P = 4.5 and time t = 5.0. These levels represent a minimum energy
input that still allows the Au to bond strongly to the bonding pads of the devices,
which is typically Pt. It is important to minimize the bonding energy because too
much acoustic energy can destroy the sample. Part of the success of this bonding
technique, and why it has lead to 100% yield to date, is because much of the bond
strength comes from the “painting” of the ribbon, rather than the sonication.
Place the sample, which is mounted in a custom high-frequency holder, onto
the bonding stage with the trigger clamp. Then place the bonding stage onto the
stage platform (which is the translatable plane allowing the stage to be positioned
under the sonicating head). The bonding stage height may need to be adjusted in
order for the sample to be at a good height to allow the proper amount of clearance
for the sonicating head. The K-connectors should be shorted at both ends of the
holder with shorting caps. The cap diameter is larger than the width of the holder
itself, so the holder will be slightly suspended by less than a millimeter above the
bonding stage, but this poses no problems for bonding.
The first bond should be a practice one on the CPW. When performing this
practice bond and all subsequent bonds it is crucial to keep the ribbon directly un-
derneath the sonicating head. This ensures that the bonds are where the sonicating
268
head touches down and not pushed off to the side. Such alignment is made simpler
by manually rotating the bonding stage so that the initial and final bonding points
lie along a line parallel to the ribbon itself.
After this test bond, the clamp will push down on the ribbon and, if the son-
icating head alignment is correct, the ribbon will be pushed through the via and
protrude from the head a distance of about the width of the ribbon itself (∼50
µm). The most common problem in ribbon bonding will occur right at this point
and is indicated by the ribbon not being pushed through the via at all. If this is
the case, the positioning of the sonicating head is not correct. Check the shape of
the ribbon in the gap between the hole in the boom and the sonicating head. If the
ribbon is bent or kinked, re-positioning of the head is necessary. Only when after
a ribbon bond the Au ribbon is pushed out of the head by ∼1 width of the ribbon
can one reproducibly make good bonds. For mechanical stability of the bonds, it is
important that the 0.25” length of ribbon protruding from the sonicating head not
be part of an actual bond to the sample but should come off during the practice
bond.
If the Au ribbon is still not reliably being pushed through the via after several
adjustments of the sonicating head, it if possible that the head is clogged. This may
be caused by broken pieces of ribbon stuck in the via or contamination originating
from the CPW surfaces that have gummed up the via. The latter case usually
coincides with the Au ribbon having difficulty sticking to the CPW, because the
surface contamination prevents good adhesion of the ribbon, even at the highest
powers. The metallic planes of the CPWs are Au-plated Cu, but over time the Au
plating can become caked with moisture-based contaminants. This contamination
layer can prevent the ribbon from adhering to the CPW, even with the techniques
269
described here, and can also gum up the via on the sonicating head, obstructing
motion of the ribbon through the via. Obstructions can be removed from the via
by clearing with a blast from the compressed air can. To clean the CPW, the best
way is simply to scrape off the contamination with a razor blade to open fresh
CPW material.
Once practice bonds to the CPWs are successful, remove them with tweezers.
Bonding to the sample should commence by bonding to all of the four ground
sheets to their respective bottom leads first. When bonding to the sample, make
sure that some of the bonding pad is visible in the microscope in front of the head.
This is because the edge of the head closest to the user does the sonicating and one
must be certain the head is directly above metal, not oxide. Au ribbon tends not to
stick to the oxide at these bonding energy levels. Signal strip to top lead bonding
should occur last. Once all of the lead bonds are made, screw on the sample cover
and place the sample in an anti-static bag. Images of a ribbon bonded sample
ready for measurement are shown in Figures 5.19, 5.20, and 5.21
5.2.4 Connecting to the dipper
When connecting the holder to the dipper the main idea is to keep the sample
grounded at all times. The dipper itself should remain connected to ground and
one should remain connected to a grounding strap throughout as well. Connections
on the room side of the cryostat should be shorted with shorting caps, thereby
shorting all of the connections in the dipper.
Attach the thermometer mount to the sample holder and remove the shorting
cap from the side of the holder facing the input line. Attach the sample holder
to the input K-connector. Remove the other shorting cap, which faces the output
270
Figure 5.19: This is an image of a completed holder with shorting caps screwedto the K103F female flange connectors on both ends. The CPWs are visible to theleft and right of the sample, which is the darker square in the middle of the holder.The sample cover is in the upper left of the image. A U.S. dime is included forsize comparison.
271
Figure 5.20: This optical microscope image shows the Au ribbon bonds linkingthe signal strip and grounding planes of the CPW to the top and bottom leads ofthe sample, respectively. The silver paint adhering the sample to the holder is alsoplainly visible underneath the chip.
272
Figure 5.21: A high magnification optical microscope image of the same device asin Figures 5.19 and 5.20 showing the two top lead connections (left to right leads),and the four bottom lead connections (diagonal leads). Despite there being sixtotal bonds, the nomenclature is for the four bottom lead, or ground, connectionsto be labelled as two. Hence, this is an image of a four-point high frequencycontact.
273
line and mates with the K-connector. This can only be done by bending the two
semi-rigid coax lines. Over time the solder joints of these two K-connectors will
become worn out and may need to be re-soldered. The orientation of the sample
holder should be that shown in Figure 5.4(a) in order to allow enough space for the
thermometer mount. With an adjustable wrench, hold the sample holder firmly
while tightening both K-connectors with the proper torque wrench (100 N·cm).
Place the Cernox thermometer in the mount.
The Janis cryostat rests on a custom table where it hangs in between the pole
faces of an electromagnet. Other users of the cryostat have their own dippers that
have different sample heights within the cryostat cavity, and so adjustment of the
table height may be necessary. At the time of this writing, there are only two
other dippers that are used within the Janis cryostats. One is a DC dipper and
the other is a high-frequency dipper with only a two-point connection. Due to the
curved input (Figure 5.4), space limitations within the cryostat restrict the sample
positioning to be higher than it would be for the other two dippers. Consequently,
the table must be lowered by 1”. This is accomplished by unscrewing the holding
nuts on the four threaded posts (13 threads/inch) underneath the table by 13 turns
each.
5.2.5 High frequency setup
A schematic of the measurement setup is shown in Figure 5.22. The sample hangs
at the end of the dipper, suspended between the poles of a GMW model 5403 elec-
tromagnet which is powered by a Kepco 20-20 BOP, producing maximum fields
|Happ| ≈ 3 kOe. The incoming signal pulse is from a 65-psec rise-time Picosec-
ond Pulse Labs (PPL) 10,070A pulse generator which provide a pulse of 10 nsec
274
maximum duration, the reset pulser is a 5-nsec rise-time Agilent 33250A Arbitrary
Waveform Generator, and the two are coupled together at the input by a power
divider. A bias T is connected at the input for optional addition of a DC current
IDC. It was found that adding an IDC to the measurement distorted the falling
side of the switching signal. Unless absolutely necessary, I would not recommend
including a DC current. If no DC current is applied, the DC input of the bias T
should be capped with a 50 Ω terminator. The output signal is amplified by a
+25 dB PPL model 5883 40 Gb/sec (inverting) amplifier and read on a Tektronix
TDS 8000B Digital Sampling Oscilloscope (TDS). Both pulse generators have 20
dB attenuators attached to their outputs (not shown in Figure 5.22) to minimize
risk of damaging the Tektronix inputs, which can be damaged for inputs beyond
a 1 V maximum range.
5.2.6 Voltage calibration at the sample input
Voltage amplitude at the sample input is calibrated as a function of pulse generator
output voltage at room temperature only. A low loss 2.4mm cable, connected by a
K-to-2.4mm F-F adaptor to the incoming line on the dipper (at the point indicated
by the blue × in Figure 5.22), was connected directly to the Tektronix for voltage
measurement. The pulse amplitudes were estimated by taking a visual average
of all the bumps in the flat range of the voltage pulse shape (see Figure 5.23(a)).
Output voltage from the pulse generator was the number returned when queried
with a GPIB interface command ÷ 10 (to account for the 20 dB attenuator).
Measured pulse amplitude at the sample as a function of output voltage from
the pulse generator is shown in Figure 5.23(b). The ratio of the two is nearly
1/2, which is expected since the power divider that couples the two pulsers at the
275
Sig
nal
pu
lse
Res
et p
uls
e
Po
wer
div
ider
Bia
s T
+2
5 d
B
Am
pli
fier
ID
C
50 G
Hz
Dig
ital
Sam
pli
ng
Osc
illo
sco
pe
Var
iab
le t
emp
erat
ure
cry
ost
at
sam
ple
4.2
- 2
90
K
Mag
net
ic f
ield
:
H
(m
ax)
= 3
kO
eap
p
Figure 5.22: High frequency setup schematic. Hardware is described in the text.The green arrow is the incoming side of the sample, while the red arrow is theoutgoing side. The blue × is the point where the incoming pulse was measured forcalibration.
276
input halves the voltages of both. The losses measured at the sample input are
attributed mostly to reflections from incidental kinks in the 180 bend. The losses
are measured by a network analyzer connected to both sides of the outgoing line
of the dipper and are found to increase smoothly from 0 to -5.1 dB between 0 and
20 GHz.
5.2.7 Sample alignment
Aligning the sample to the holder is done during the silver painting step just before
ribbon bonding. This is a purely visual alignment, but can usually be within 5 if
a line between two adjacent devices is used to orient the chip relative to the groove
edge. Rotational alignment in the cryostat, done by corroborating the orientation
of the sample with the orientation of the adaptor, is again purely visual but can
be achieved to within 5.
5.3 Performing time-resolved pulsed current measurements
5.3.1 Preparation
DC characterization of samples mounted on the high frequency dipper is possible
with the usual Wheatstone bridge technique by employing K-to-BNC adaptors. It
is best to do DC characterization all at once because I found that changing the
cable connections frequently meant a faster arrival at sample death. Once DC
characterization is complete, connect the input line to the power divider (which
couples the signal and reset pulsers together) and the output line to the amplifier.
The amplifier must be powered by a +8V, -5V power supply and, if functioning
properly, should be drawing ∼0.2 A of current. When powering the amplifier,
277
Slope = 0.4639
(a)
(b)
Mea
sure
d a
mp
litu
de
[m
V]
Am
pli
tude
[mV
]
Figure 5.23: Calibration of the voltage pulse. (a) The measured voltage ampli-tude was estimated as the middle of all the bumps and wiggles in the flat voltagerange of the voltage pulses. The one shown here is not from the calibration, butis shown to illustrate the method. (b) Measured versus output voltage calibra-tion. Pulse generator output voltage was the number returned when queried witha GPIB interface command ÷ 10 (to account for the 20 dB attenuator).
278
exceeding these levels can destroy it, so turn up the voltage slowly. Make the sure
amplifier is powered for at least one hour before doing any pulsed measurements
because it takes some time for all the internal circuitry to thermally equilibrate
and large distortions will appear in the data if used before this.
5.3.2 Software control
The time-resolved, pulsed-current experiments are controlled by LabView code.
The last version used in F20 is in the C:/Documents and Settings/Buhrman
Group/Desktop/High Freq VIs/TDS 8000b/ directory, and is named “Time Ma-
chine 2.8.vi”. The low-noise input is channel 4, so the channel 4 soft button on
the front panel of the TDS should be illuminated. To run the program, the pulse
shape must be correctly placed on the screen of the oscilloscope. To do this, en-
able the signal pulser output (with Panel Control 1.3.vi program), then set the
acquisition mode (Acq. Mode on the TDS touch screen) to Sample. The signal
should be flickering, meaning it is actively scanning the input. If it is not, press
the Run/Stop touch screen button to run the sampling acquisition (it will turn
green when pressed). If nothing happens, it is possible that the trigger level is
not set properly. The default setting is 0 mV-rising edge, meaning that the TDS
oscilloscope looks for the rising edge of the trigger at a threshold of 0 mV. The
trigger level is 1 V, so manually adjust the trigger level to 500 mV (using the touch
screen buttons). This should result in a signal being displayed on the screen. Ad-
just the voltage and time settings so that a waveform similar to that shown in
Figure 5.15(a) appears on the screen. Data is acquired only for the pulse turn on
and its duration, so make sure the falling edge of the pulse is not on the screen.
Once the pulse is adjusted on the screen set the acquisition mode to Average.
279
The data entered on the front panel of Time Machine 2.8.vi that is pertinent
to transient dynamics data is shown in Table 5.3. Data Points is the number of
samplings per trace, Scope Ave. is the number of averages taken on the TDS,
Soft Ave. is the amount of times LabView calls for the TDS to take Scope Ave.
averages. pause [ms] is an internal delay to allow for GPIB communication. Probe
adjustment is a toggle switch that asks for user input to adjust the probes. This
is a relic from when this program was used in a probe-bearing cryostat and should
be set to “Off” for use with bonds. 20 dB attenuator? is a toggle switch that
accounts for presence of the 20 dB amplifier on the output of the PPL signal pulse
generator and should be se to “Yes”.
Switching/Precession is a toggle switch that, when set to Switching prints out
only the Time Machine 2.8.vi front panel and not an additional page of Fourier
transformed data that is more useful for persistent dynamics measurements. Ab-
solute value? is a toggle switch asking if the scans recorded on the TDS should
be absolute valued when passed into LabView. Return Field to 0? if toggled to
“YES” returns the field back to zero after every scan, which needlessly increases
the amount of time per measurement. The algorithm for jitter correction does not
work properly in Time Machine 2.8.vi and so “Auto jitter correction” should be
toggle to “NO”. Field values Init, Final, and Base should all be equal for switching
measurements, and should be set to the value of the GMR minor loop offset (i.e.
dipole field from the fixed layer) as determined from DC measurements.
Control of the pulse generators is by a separate program in the C:/Documents
and Settings/Buhrman Group/Desktop/High Freq VIs/Psec10070/ directory, and
is named “Panel Control 1.3.vi”. This vi controls the PPL signal pulse (amplitude
and polarity), the Agilent reset pulse (amplitude, trigger delay, and pulse width),
280
Table 5.3: Front panel input values for Time Machine 2.8.vi. These are the valuesrelevant for transient dynamic measurements.
Input Value
Channel 4
Data Points 250
Scope Ave. 1000
Soft Ave. 10
pause [ms] 20
Probe adjustment (toggle) Off
20 dB attenuator? (toggle) Yes
Switching/Precession (toggle) Switching
Sample name
Absolute value? (toggle) No
Return Field to 0? (toggle) No
Auto jitter correction (toggle) No
Init Field [G] = Final Field [G] = Base Field [G]
281
and the outputs for both (Enable/Disable for the PPL and Toggle output On/Off
for the Agilent). The polarity of the reset pulse is automatically made opposite to
the signal pulse. set amplitude is the value read from a GPIB query to the PPL that
does not take into account the 20 dB attenuator, although the number displayed
in the Time Machine 2.8.vi output incident V [mV] does account for the 20 dB
attenuation (i.e. ÷ 10). The input value for the reset pulse amplitude (Vpulse
[mV]) does not account for the 20 dB attenuation and should be 1001 − 2001
mV. Trigger delay should be 60 nsec, and pulse width 100 nsec. Make sure both
pulser outputs are enabled. NEG, or negative polarity means P to AP switching ,
POS, or positive polarity means AP to P switching (for normal bottom lead/fixed
layer/spacer/free layer/top lead nanopillar geometries). It is important not to
alter the nomenclature or convention for the pulses in the software as the filename
structure is used extensively in the analysis software (see section 5.3.5).
5.3.3 Data acquisition
The data acquisition method, originally published in [3], is shown in Figure 5.24.
The data are acquired first by taking a baseline scan, which is where there is a
signal pulse but no reset pulse. In this sequence the reset pulse amplitude is set to
1 mV (which is further attenuated to 0.1 mV by the 20 dB attenuator), which is
far too small to switch the nanomagnet. The output of the reset pulser is not set
to 0 mV because that would automatically shut off the output and thereby change
the input impedance of the pulser which will introduce large voltage offsets in the
switching data3. As a reference, Figure 5.24(a) shows a DC spin transfer scan.
The baseline sequence is where the sample is pulsed between positions 1 and S,
3Input impedance of the reset pulser is 50Ω when the output is on.
282
but, as there is no reset pulse, the nanomagnet does not switch and so all that is
measured is the background. The corresponding pulse train (top of Figure 5.24(b))
contains only the fast rise time signal pulse. The baseline (and signal) sequence
is averaged over 10,000 traces, each of which is a construction of 250 points. For
a 10 nsec observation window, the time-resolution (i.e. sampling bin-width) is 40
psec, which is sufficient for spin torque switching measurements.
For the signal sequence, the much slower reset pulse is added to the pulse train
(bottom of Figure 5.24(b)), where now the sample is pulsed from 2 to S, which
records the switching event, returning to 1 whereupon the reset pulser, which is
triggered by the PPL signal pulse generator, pulses to R and the nanomagnet is
reset, and ready to receive another sequence of signal-then-reset pulses. Employing
two separate pulse generators is merely out of convenience as the pulse generator
could reset the nanomagnet with a GPIB command to invert pulse polarity. How-
ever, GPIB commands operate on a 10 ms timescale, whereas the trigger of the
PPL signal pulse generator cycles every 100 µsec, thereby decreasing measurement
time by two orders of magnitude.
An example of the recorded signal and baseline scans for a parallel (P) to
anti-parallel (AP) spin torque switching event is shown in Figure 5.25(a). The
majority of the voltage reflects the measured excitation pulse, and the signal and
baseline scans are nearly identical except for a small gap between them around
time = 1 and 2 nsec. Taking the difference between the signal and baseline scans
gives the actual switching event, shown in 5.25(b). This small difference between
signal and baseline scans is the voltage contribution from the GMR of the device
|I ·∆R|, where the current pulse amplitude I is found from an equivalent circuit
relation [2] I = Vp/Req(T ), where Vp is the pulse amplitude from the PPL signal
283
(b)
(a)
1
2R
S
Baseline Sequence: 1 - S - 1
Signal Sequence: 2 - S - 1 - R - 2
Measurement window
Figure 5.24: Pulsed current acquisition sequence for time-resolved measurementof the magnetization reversal. (a) The DC spin transfer scan as a reference, showingpositions 1, 2, S, and R which correspond to different nanomagnet orientationsaccessed during the signal and baseline pulse trains. (b) Top: The pulse train forthe baseline scan, which includes only the fast rise time signal pulse and pulses thenanomagnet between 1 – S – 1. Bottom: The pulse train for the signal scan, whichhas now included the reset pulse so that the nanomagnet is pulsed between 2 – S– 1 – R – 2. Measurement is only made for the pulse rise and duration (pulsed 2– S), as shown by the measurement window.
284
pulse generator (calibration shown in Figure 5.23(b)), and the equivalent resistance
Req(T ) =1
2[2Rx(T ) + RT (T ) + RB(T ) + 50Ω] (5.1)
where Rx(T ), RT (T ), and RB(T ) are the sample, top lead, and bottom lead resis-
tances, respectively, each of which is measured separately as functions of T , and
∆R = Rx(AP) − Rx(P). It is important to measure the 4-point resistance Rx(T )
for the sample mounted on the high frequency dipper, but RT (T ) and RB(T ) are
purely dependent on lead geometry thickness and so it is sufficient (and much more
convenient) to measure these on an adjacent sample. This can be done very quickly
using DC probes inside the Desert Cryogenics tabletop cryostat, measuring the 4-
point resistances of the top Rt(T ) and bottom Rb(T ) leads as functions of T . Note
that, for the resistances in equation 5.1, RT (T ) = 12Rt(T ) and RB(T ) = 1
4Rb(T ).
5.3.4 Data analysis
LabView code has been written to handle all of the data analysis. On the F20
computer, these files are in the C:/ Documents and Settings/ Buhrman Group/
Desktop/ High Freq VIs/ Data Analysis/ directory. It is very important to save
all of the data in a very specific directory structure, which is shown schematically
in Figure 5.26, because this LabView analysis code will not be able to locate
the raw data files if a different file structure is used. These are case and spacing
sensitive. There must be eleven temperature folders, for data taken at up to eleven
temperatures, in the “Measurements” folder and they must be labelled as “4.2K”,
“20K”, “40K”, “80K”, “120K”, “160K”, “200K”, “240K”, “290K”, “other1”, and
“other2”. The software works by seeking out eleven folders, even though eleven
285
10
300
100
150
200
250
50
2 3 4 5
Sig
nal
, B
asel
ine
[mV
]
Baseline
Signal
time [ns]10
0
2
3
4
5
P
AP
1
2 3 4 5
|Sig
nal
- B
asel
ine|
[m
V]
(a)
(b)
Figure 5.25: (a) The baseline (blue) and signal (red) data for a P to AP switch-ing event. The curves are averages of 10,000 individual traces, each of which isconstructed from 250 individual samplings. So this plot represents 5 million nano-magnet reversals. They are nearly identical except for the small sliver of whitebetween them around time = 1 to 2 nsec. The difference between these two is theswitching event (b), where the signal peak is fully P alignment and the return tozero is the fully AP alignment.
286
temperatures may not have been explored, the folders must be present. If there is
not data for all temperatures, leave the ones with no data empty. “other1” and
“other2” are useful if there are other temperatures one has data for (e.g. “60K”).
If this is the case, the “other1” folder must be renamed to the new T value.
Time-shifting
Shifting the signal and baseline scans in time is generally necessary because the two
are usually displaced in time due to jitter noise in the triggering of the oscilloscope.
Time-shifting the data is necessary for Time Machine 2.8.vi because the Auto jitter
correction code does not work properly, although a later version, on the CNS3
computer (“Time Machine 2.16.vi”), has a working automatic jitter correction
code. Time-shifting and automatic jitter correction do the same thing, namely shift
the signal relative to the baseline scan, until the rising edges of the two coincide.
To time-shift, open the “Read data from file 1.3.vi” LabView file. Select each
pulsed switching raw data file individually and time-shift the data as necessary.
An example of a time-shifted switching event is shown in Figure 5.27. Usually, the
raw data has some upwards or downwards spike indicating the need for a time-shift
(Figure 5.27(a)). The criteria for a proper time-shift is to make this spike as small
in amplitude as possible, at least on the 0.1 psec scale shifting in the program. It
is important that this spike not be placed in the mid-way point of the current step
as this will cause errors when trying to estimate the rise time. Keep the spike as
close to 0 (but still positive) as possible. The time-shifted data are saved in new
folders “Sorted by Vpulse” and “Sorted by IDC” in the newly created “Adjusted
Data” folder.
287
Dynamics
Time Domain
Switching
IDC (e.g. "0 mA")
IDC (e.g. "0 mA")
Vpulse (e.g. "211.38 mV")
Sorted by IDC
Sorted by Vpulse
Temperature (e.g. "4.2K")
dVdIGMR
AntiPar to Par
Final Form Data N% Vrise
Adjusted Data
GMRraw data----------
Pulsedswitchingraw data------------
Pulsedswitchingfinal formdata------------
SMTraw data----------
Collapsed Data
Root
Measurements
Par to AntiPar
Figure 5.26: Directory structure necessary to execute the analysis software. Theroot is not specific. All files with curly brackets are not necessary for theanalysis software to work, but are helpful to store other data (e.g. DC data).All files in italics represent a set of similar folders and are distinguished by somenumber representing an experimental control parameter (e.g. “4.2K” or “0 mA”).All files labelled in black must be created by the user, whereas the blue-labelledfolders will be automatically created and sorted by the software. The N is thefraction of the voltage rise taken as the temporal reference point from which tomeasure the switching time. Typically N = 50. An identical directory structureshould exist for “Par to AntiPar” as well.
288
(a)
(b)
Figure 5.27: Example of time-shifting data. (a) Raw (|Signal – Baseline| data.(b) The same data with the Signal data shifted in time by 3.1 psec. The arrowindicates the downwards peak that was reduced by the time-shifting. Other thanreducing this spike, difference introduced by time-shifting are minimal.
289
Measure tswitch
Once all of the data has been time-shifted, it is now ready to be analyzed. From
the same /Data Analysis/ directory open the “Switching Time 1.6.vi”, “Sectioned
polynomial fitting 1.2.vi”, and “Print And-Or Save 1.2.vi” LabView files. The
Fraction of rise signal input is N from Figure 5.26, indicating the fraction of the
pulse turn-on that the switching time is measure with respect to. There is no need
for it to be other than N = 50. Run Switching Time 1.6.vi and it prompts the
user to select a directory. Go to the “Adjusted Data” directory, where the new
folders “Sorted by IDC” and “Sorted by Vpulse” now sit, and click on Select Cur
Dir. Run the program once without saving the data to make sure there are no
errors, then rerun with the Save? button activated. There is no need to print the
output.
The program reads in one file at a time from the “Sorted by Vpulse” folder and
divides the total wave form in half, split on either side of the voltage maximum
Vmax. The signal to the left of Vmax is analyzed to find the time t1 at 0.5·(Vmax -
Vo), where Vo is the “initial voltage” determined as the mean of the all initial data
points that occur before a voltage threshold of 1.0 mV is exceeded. This threshold
is a reasonable estimate of the noise floor in the data and is mostly independent of
T , switching direction, I, and sample. The signal to the right of Vmax is analyzed
to find the times at 0.9·Vmax, 0.5·Vmax (t2), 0.4·Vmax, 0.3·Vmax, 0.2·Vmax,
and 0.1·Vmax. The switching time is then defined as tswitch ≡ t2 − t1. The width
in the switching transition (named “dispersion” inside the program) is measured
between 0.9·Vmax and one of 0.4·Vmax, 0.3·Vmax, 0.2·Vmax, or 0.1·Vmax. The
selection is merely for flexibility.
290
Error analysis
As tswitch is measured, its error is determined within the same execution of Switch-
ing Time 1.6.vi. The data are read into Sectioned polynomial fitting 1.2.vi where
they are sectioned into 0.5 nsec intervals (although this interval size is adjustable).
Each interval is fit with a nth order polynomial, where n is stepped out from 3 to
13. At each n value, the second derivative of the polynomial is taken to define the
inflection points. All of the inflection points from each of the 0.5 nsec intervals, as
well as the value of the signal at maximum time τ and at (Threshold)·Vmax, are
then fit with an decaying exponential function An exp(−t/Bn), where Threshold is
defined below. For each n, the best fit decaying exponential is subtracted from the
data at each measured time value ti. The absolute value of these differences are
then all summer together but with a starting time corresponding to the signal value
(Threshold)·Vmax, with Threshold = 0.90, typically. This starting point of the
summation accounts for the natural differences in shape between the exponential
fit and the data at earlier times.
f0(n) =∑
i
|Vi − An exp(−ti/Bn)| (5.2)
The minimum value of f0 is divided by the time interval of the summation τ , so
that the error in signal is ∆y = f0(nmin)/τ . Error in tswitch is then found by
estimating the respective times ∆t± between 0.5·Vmax and 0.5·Vmax ± ∆y. The
error in tswitch is then ∆tswitch ≡ 12(∆t+ + ∆t−). If there are multiple traces for
a single voltage amplitude, then the final error in tswitch is the quadrature sum of
each ∆tswitch and the standard deviation of tswitch. At the end of running Switching
Time 1.6.vi, the data tswitch, ∆tswitch, both of these values from individual traces,
291
signal amplitude (named “signal swing” in the code) Vmax−Vo, the dispersion
data, and the best fit exponential decay are all saved in new files within the new
folder “Final Form Data 50% Vrise”.
Collapsing the data
Now that tswitch ±∆tswitch has been determined, pulsed voltage amplitudes must
now be converted into current amplitudes, as prescribed by equation 5.1. This
is done with a sorting program named “Collapse Data 1.1.vi”. This program
is designed to find all of the errors that occurred during the complex analysis.
The algorithms are not infallible and some errors do occur. Before collapsing
the data, all of the Rx(T ), Rt(T ), and Rb(T ) numbers need to be entered into
the front panel input. Note that the numbers entered for top and bottom lead
resistance are Rt(T ), and Rb(T ), as they are measured in the Desert Cryogenics
tabletop cryostat. Pulse Conversion is the calibrated 0.4639 from Figure 5.23(b)
and equation 5.1 is programmed into the code. The pulsed voltage amplitude,
which has been preserved in the filename between all of the programs, is now used
to calculated the current amplitude. Only one switching polarity can be done at
a time.
The recommended procedure is to include one temperature at a time by clicking
on the adjacent green include? icon. If there are any errors in the final form data,
they will be logged. The majority of errors will be due to incomplete switching,
where the signal does not fully return to zero within the measurement window.
“<Inc. Swing> *Dispersion*” will be the returned error. These can be ignored
as they are only saying that not all of the dispersion values (i.e. 40%, 30%, 20%,
and/or 10% of Vmax) were successfully estimated. Other common, but unim-
292
portant, errors are “<NaN error> *Dispersion*”, “<Inc. Swing> *SS*”, “<NaN
error> *SS*”, where SS means signal swing, and the NaN error meaning that es-
timating algorithm for the value (SS or Dispersion) failed. This is usually the case
when the signal-to-noise (SNR) is low, at high temperatures and/or at low pulsed
current values.
The errors that do need to be dealt with are logged as “<NaN error> *Tswitch*”.
This means that tswitch is not being correctly calculated somewhere in the analysis.
Next to the error message in the log is the final form data which gave the error.
Looking up each errant file and correcting it is the prescribed method. The most
common problem is that the fitting algorithm failed somehow and now the final
form has returned a not-a-number (NaN) for tswitch. If there are multiple values
of tswitch in the final form data file, but only one showing NaN, simply delete the
bad data point and adjust averages and errors accordingly. Also, for the SS, make
the NaN equal to the average of the other SS data.
293
References for Chapter 5
[1] Pozar D.M., Microwave Engineering, chapter 4, Wiley, 2nd edition (1997),ISBN 0471170968.
[2] Krivorotov I.N., Emley N.C., Sankey J.C., Kiselev S.I., Ralph D.C., &Buhrman R.A., Time-domain measurements of nanomagnet dynamics drivenby spin-transfer torques, Science 307, 228–231 (2005).
[3] Koch R.H., Katine J.A., & Sun J.Z., Time-resolved reversal of spin-transferswitching in a nanomagnet, Phys. Rev. Lett. 92, 088302 (2004).
CHAPTER 6
TIME-RESOLVED SPIN TORQUE SWITCHING AND ENHANCED
DAMPING IN PY/CU/PY NANOPILLARS
6.1 Introduction
Experiments [1–4] have shown that a spin-polarized current passed through a nano-
magnet can excite a dynamic response as the result of a spin torque applied by
the conduction electrons [5, 6]. The potential for technological impact of this spin
transfer (ST) effect has inspired research in DC current-induced microwave oscilla-
tions [3, 7] and hysteretic switching [1, 2, 4] in current perpendicular to the plane
(CPP) nanopillars and nanoconstrictions. Typically, ST switching data is obtained
through the use of slow current ramp rates (∼1 mA/sec), but fast pulses (∼1010
mA/sec) access the regime where the major contribution from temperature T is
not to thermally activate the moment over a current-dependent barrier [8, 9], but
only to set, through thermal fluctuations, the precise orientation of the nanomag-
net moment at the initiation of the pulse and to scale the T dependent magneti-
zation. This spin torque-driven regime [10] is advantageous for the quantitative
examination of the spin torque parameters due to the computational accessibility
of numerically integrating the Landau-Lifshitz-Gilbert (LLG) equation for short
durations.
Here I present time-resolved measurements of the spin torque-driven switching
event in Cu 100/Py 20/Cu 6/Py 2/Cu 2/Pt 30 (in nm, Py = Ni81Fe19) CPP spin-
valve nanopillar structures at bath temperatures T = 4.2 K to 160 K. I compare the
experimental results with LLG simulations in the macrospin approximation and
find good agreement between simulation and measurement. This both confirms the
294
295
applicability of the macrospin approximation in the spin torque-driven regime and
facilitates the quantitative determination of T -dependent spin torque and magnetic
damping parameters. At higher T the strength of the spin torque exerted per
unit current is found to be in reasonable numerical accord with recent model
calculations, and that the damping parameter α0 for the nanomagnet excitations is
both anomalously high, as suggested by previous pulsed current measurements [11],
and T -dependent. The strong T variation of α0, in conjunction with anomalous
behavior of the nanomagnet switching fields HS,i(T ) in some devices, points to the
presence of an adventitious antiferromagnetic oxide layer around the perimeter of
the nanomagnet that has a major effect on the nanomagnet dynamics driven by a
spin torque.
6.2 Experiment
The nanopillar devices employed in this study were fabricated by ion-beam etching
sputter-deposited films, where e-beam lithography-patterned C etch masks define
the pillar shape and lateral size (see section 3.2 of this dissertation). Measurements
were performed inside a continuous He-flow cryostat which was suspended between
the pole faces of an electromagnet supplying in-plane magnetic fields |Happ| < 3.0
kOe. Alignment of the easy axis of the elliptical device with Happ was accurate to
within ±5. DC giant magnetoresistance (GMR) and ST sample characterizations
were done using standard Wheatstone bridge with lock-in techniques. A slow ramp
rate ST scan is shown in Figure 6.1 for sample 1, a 60×190 nm ellipse. In Figure 6.2
I show parallel (P) to anti-parallel (AP) switching events for sample 1, averaged
over 10,000 switches, taken at pulsed current amplitudes I = 1.07 mA and 2.13 mA
at T = 40 K as open and solid squares, respectively. The measured signal is a time-
296
T = 40K
H = 385 Oeapp
Figure 6.1: Slow ramp rate spin torque switching of a nanomagnet as measuredby the GMR effect for sample 1, a 60×190 nm ellipse, at T = 40 K and Happ =385 Oe, which opposes the dipole field so that Happ + Hdip ≈ 0. Arrows indicate
the scan direction.
297
T = 40K
P to AP
1.07 mA2.13 mA
tswitch
H = 385 Oeapp
[nsec]
mz
Figure 6.2: Pulsed spin torque P to AP switching measured for the same sampleat I = 1.07 mA () and 2.13 mA (). The data (symbols) have been normalizedto mz = ±1 for simple comparison with the simulated macrospin switching (lines).Pulse shape distortions are due to the setup (see ref. [12]).
298
resolved voltage drop |I ·∆R| from the giant GMR of the sample as the free layer
switches from P to AP orientation, where ∆R ≡ Rx(AP)−Rx(P) and Rx is the 4-
point device resistance. The data have been normalized to mz = +/−1 (minimum
resistance / maximum resistance) for simple comparison with simulated switching
events, described below, which are shown as solid and dashed lines [13]. The pulsed
data were measured with the techniques described in Ch.5 of this dissertation. The
abrupt (∼200 psec) jump from mz = -1 to 1 at time = 0 is not a switching event
but is simply the rising edge of the current pulse. The more gradual transition
between P (mz = 1) and AP (return to mz = -1) is the envelope coming from
averaging over thousands of individual switching events, each of which follows a
trajectory determined by initial conditions that are randomized by the stochastic
thermal fluctuations of the free layer. The switching time tswitch is defined as the
time elapsed between 50% of the signal rise and 50% of the signal drop as indicated
in Figure 6.2 [12].
6.3 Macrospin simulation
To obtain a quantitative understanding of the ST switching, I have simulated
the nanomagnet dynamics by numerical integration of the LLG equation in the
macrospin approximation with the inclusion of a Slonczewski-type spin torque
term.
dm
dt= γ[m× ( ~Heff + ~HLang(T ′))− α(θ)m× (m× ( ~Heff + ~HLang(T ′)))
− Ihg(θ)
eMs(T ′)(area · d) sin θm× p× m] (6.1)
Here γ is the gyromagnetic ratio, m is the unit directional vector of the free layer
macrospin, p is the spin polarization axis, θ is the in-plane angle between them,
299
g(θ) is the spin torque function, Ms(T ) is the free layer magnetization, as mea-
sured separately for a continuous 2 nm Py film in a Cu/Py/Cu trilayer that was
exposed to the same heat treatments as the nanopillars, d is the nanomagnet thick-
ness, area = π4ab is its lateral area with dimensions a and b that are estimated by
OOMMF micromagnetic simulations [14] (see below), and ~Heff is the sum of ex-
ternal ~Hext, in-plane anisotropy ~HK, and out-of-plane anisotropy ~H⊥ fields. ~Hext
is the sum of the magnetostatic dipole field from the fixed layer ~Hdip and the in-
plane applied field ~Happ from the electromagnet, which is adjusted to compensate
for ~Hdip so ~Hext ≈ 0, but see below.
The initial conditions of the free layer simulation were set by its initial angle
(θi) relative to the fixed spin-polarization axis p,
θi = θ0 + θmis + θrand(T ) (6.2)
where θ0 = 0 for P to AP and 180 for AP to P switching and θmis represents
any systematic angular misalignment between free and fixed layer moments due to
the setup and was generally set to 0, but see below. The random angle θrand(T )
is treated as a Gaussian with a standard deviation
σθrand=
√kBT/2E0(T ) (6.3)
where
E0(T ) = E0(4.2 K)[Ms(T )/Ms(4.2 K)]2 (6.4)
is the uniaxial anisotropy energy. Typical values of σθrandare ∼1 for T = 4.2 K,
∼3 for T = 40 K, 5 for T = 120 K, and 6 for T = 160 K. Randomization of
θi is crucial in closely approximating the experiment since the measured switching
behavior is the average of many thousands of events, each having a thermally
300
determined initial angle. E0(4.2 K), a, and b are estimated from T = 0, 2D
OOMMF simulations of Py elliptical disks having HK and ∆R values similar to
those measured at 4.2 K. E0(4.2 K) typically ranged from 0.6 to 1.1 eV. The lateral
area is estimated to 12% uncertainty with this method, nearly a factor of 2 better
than the inherent shape variation among otherwise identical elliptical patterns due
to lithographic fluctuations.
The in-plane anisotropy field is also T -dependent,
HK(T ) = HK(4.2 K)[Ms(T )/Ms(4.2 K)] (6.5)
Ohmic heating effects during the current pulse are taken into account by locally
raising the temperature of the device to
T ′ =√
T 2 + 10.23(K/mV)2(Rx(T ) · I)2 (6.6)
where the coefficient of (Rx(T ) · I)2 is 34( e
πkB)2 [9]. A Langevin field ~HLang(T ′)
accounts for thermal fluctuations during the dynamic trajectory, fluctuating ran-
domly in 3-dimensions with a standard deviation
σHLang=
√2α0kBT ′µ0/γMs(T ′)(area · d)∆t (6.7)
where ∆t = 1 psec is the simulation time step [8, 10, 15]. Gilbert damping is
assigned an angular dependence
α(θ) = α0[1− ν sin2 θ/(1− ν2 cos2 θ)] (6.8)
where ν = 0.33 for Py/Cu/Py nanopillars [16], but the addition of this angle-
dependent damping term had only a small effect on the simulation results.
The spin torque function is approximated by
g(θ) = A sin θ/(1 + B cos θ) (6.9)
301
where A and B are phenomenological parameters [11, 17, 18]. In my simulations
I use α0, A, and B as T -dependent fitting parameters to match the simulated
with the measured values of 1/tswitch versus I for each T , where α0 is allowed to
be different for the two switching directions. In Figure 6.2 I plot the average of
2000 simulated P to AP switching events at T = 40 K alongside the normalized
data for sample 1 with the best fit simulation yielding A = 0.5, B = 0.11, and
α0 = 0.048. Since the current step in the simulation turns on instantaneously,
an average pulse half rise time of 112 psec, measured from data such as those
in Figure 6.2, has been added to all simulated tswitch. I plot measured 1/tswitch
versus I for AP to P and P to AP switching at T = 160 K, 40 K, and 4.2 K for
sample 2, an 80×180 nm ellipse, together with best fit simulations, all of which
are averages over 2000 events, in Figures 6.3 and 6.4, respectively. Simulations
out to long switching times (1/tswitch < 0.1 nsec−1) allow for good estimates of
the 1/tswitch → 0 intercepts I±c0(T ), which are the critical currents (+ = P to AP)
defining the onset of spin torque-driven switching. These should depend on the
spin torque and damping parameters as I±c0(T ) ∝ (α0/η±)M2
s (T ) [10], where η±
are the effective spin polarizations of the current for the two switching directions
[η± = A/(1 ± B)]. A striking result from these measurements is the strong T -
dependence of I±c0(T )/M2s (T ) (Figure 6.5), which varies by more than 60% over
the entire T range, where the upturns at low T indicate a strong dependence of
damping, spin torque, or both.
302
AP to P
4.2 K
40 K
160 K
Figure 6.3: Measured (symbols) and simulated (lines) 1/tswitch versus I for sample2, an 80×180 nm ellipse, at T = 160 K (), 40 K (•), and 4.2 K () for AP to Pswitching.
303
P to AP
Ic0+(4.2 K)
4.2 K
40 K
160 K
Figure 6.4: Measured (symbols) and simulated (lines) 1/tswitch versus I for sample2 at T = 160 K (), 40 K (•), and 4.2 K () for P to AP switching. Simulationsto 1/tswitch ≈ 0 yield estimates of the intercepts I±c0(T ), as shown for T = 4.2 K.
304
P to AP
AP to P
Figure 6.5: I±c0(T )/M2s (T ) versus T . P to AP () and AP to P (•) switching.
In Figure 6.7 the best fit values for α0, A, and B (assuming θmis = 0) are
plotted as functions of T for sample 2. Uncertainties in the fit parameters, ∆α0
= 0.0035, ∆A = 0.025, and ∆B = 0.045, are found through an exploration of
parameter space about the best fit values. Accounting for these T -dependences,
the theoretical prediction of I±c0(T ) ∝ α±0 (T )M2s (T )(1 ± B(T ))/A(T ) agrees with
the measurement to within 10% over the entire range of T , shown in Figure 6.6. All
four devices that were extensively studied show an amplitude and T -dependence
of α0 very similar to that of Figure 6.7(a); a gradual but significant increase with
decreasing T below 160 K, above which the devices are thermally unstable, followed
by a stronger increase starting below 60 K - 40 K where the best fit values of α0 also
305
P to AP
AP to P
Figure 6.6: Critical currents normalized to spin torque parameters.I±c0(T )A(T )/[α±0 (T )M2
s (T )(1±B(T ))] versus T , varies by < 10% over the range ofT = 4.2 K to 160 K.
suggest differences between the two switching directions. For T < 60 K, the trends
in the T -dependence of the spin torque parameters A and B vary from sample to
sample, but for T > 60 K both consistently show a very mild dependence on T as
illustrated in Figure 6.7(b). For the four samples studied in detail I found that at
40 K A ranged from 0.5 to 0.68 and B varied from 0.11 to 0.35. In general I also
found that A would decrease by 10 or 20% in going from 40 K to 160 K while B
would typically vary by 10% or less.
These values of A and B and the variation with T > 60 K can be compared with
the results of a two-channel model [19] with which the measured GMR parameters,
306
R(T ) and ∆R(T ), can be used to predict the spin torque parameters [20]. This
model predicts A = 0.52, B = 0.36 at 40 K, with A decreasing to 0.47 at 160 K and
B remaining essentially constant. This is in reasonable accord with the data, given
the experimental uncertainties in nanomagnet size and alignment. I note here that
the best fit values for α0, A, and B as functions of T do depend on the choice
of misalignment angle θmis, which is not directly controlled in the measurement.
However, simulations performed out to θmis = 10 show that, although α0, A,
and B all trend towards smaller values for larger θmis, the T -dependences are not
strongly affected (see Figures 6.23, 6.24, and 6.25).
If, instead of simulating the nanomagnet switching in zero magnetic field, I
include a non-zero constant external field Hext 6= 0 both the spin torque and
damping parameters change. The resulting best fit values of α0, A, and B are
shown for sample 2 at 4.2 K in Figure 6.8, where I have simulated at Hext = -100
Oe and -200 Oe. The sign of the field is such that it points in the direction of
the dipole field from the fixed layer and favors AP alignment. HK = 261 Oe for
sample 2 so the device is still in the regime of hysteretic switching. From the best
fits to the measurement, as Hext gets increasingly negative B gets larger indicating
that the spin torque per unit current for the two switching directions gets more
and more asymmetric (see e.g. the spin torque function g(θ), Figure 6.20). This
is understood as the effect of Hext < 0 strengthening AP alignment and thereby
increasing the amount spin torque necessary to induce AP to P reversal. The
P state is simultaneously weakened and consequently the amount of spin torque
necessary to induce P to AP reversal is reduced. Effective spin polarizations versus
Hext are shown in Table 6.1.
307
(a)
(b)A
B
Figure 6.7: Best fit parameters (a) damping α0 (for P to AP () and AP to P(•) switching) and (b) spin torque parameters A () and B () as functions of Tfrom matching simulated with measured values of 1/tswitch versus I for each T forsample 2.
Table 6.1: Effective spin polarization η± = A/(1±B) (+ = P to AP) as a functionof Hext. A and B values are from best fits of the macrospin LLG simulation toT = 4.2 K (1/tswitch versus I) data for sample 2 (see Figures 6.7 and 6.8). Thistable shows that, for increasingly negative Hext (< 0 favors AP alignment) morespin torque is required to induce an AP to P reversal, while less is required for Pto AP.
Hext P to AP: η+ = A/(1 + B) AP to P: η− = A/(1−B)
0 Oe 0.52 0.99
-100 Oe 0.46 1.03
-200 Oe 0.34 1.08
308
P to AP
AP to P
(a)
(b)
ext
Figure 6.8: Effects of changing the net external field Hext from 0 to -200 Oe. Forthese simulations, I have matched the simulated and measured value of 1/tswitch
(similar to the fits shown in Figures 6.3 and 6.4), now with a non-zero Hext.Damping decreases with negatively increasing Hext, and the spin torque functiontends toward being more asymmetric (i.e. larger B).
309
6.4 Discussion
The significant T -dependence of α0 is attributed to the presence of a weak an-
tiferromagnetic (AF) layer on the sidewalls of the nanopillar. Although no such
AF layer was deliberately deposited, the exposure of the nanopillars to air after
ion mill definition undoubtedly oxidized the sidewalls, thus allowing for AF Py
oxide to form and weakly exchange bias the ferromagnetic layers. An example of
direct evidence for this adventitious exchange biasing is shown in Figure 6.9, where
switching fields HS1 and HS2, defined in the inset, from 20 field scans at each T ,
are plotted from 4.2 K to 160 K for sample 3, an 80×180 nm ellipse, a previously
unmeasured device cooled in Happ = 0. Note that HS1 varies more rapidly with T
than HS2, which is indicative of an exchange bias that strengthens with decreasing
T (particularly rapidly below 40 K) and promotes AP alignment, i.e. a bias set by
the dipole field from the fixed layer.
Another key point, illustrated in Figure 6.10 is the large variation that develops
in HS1, and to a lesser extent in HS2, upon multiple minor loop scans after the
device is cooled to low T . Initially, the device switches repeatedly with nearly
the same switching fields (see the minor loop scan in the inset of Figure 6.9), but
after six or seven magnetic reversals the switching fields begin to fluctuate greatly
from reversal to reversal, indicating stochastic variations in the net strength of the
oxide pinning field. While the effects of the random pinning field are particularly
pronounced at 4.2 K they are observed up to 160 K, indicating that some degree
of magnetic ordering within the AF persists over this entire T range and also that
each reversal of the free layer nanomagnet has an irreversible perturbing effect on
the magnetic structure of the AF oxide. It is important to note that the slow
ramp rate current-driven switching events at low T for these devices show good
310
reproducibility, with little variation from one sweep to another, as should be the
case because ST switching currents are less sensitive to field variations than are the
switching fields. The strength of this low T AF exchange biasing varies from device
to device, with some samples showing no random variations in HS,i. Fluctuations
of this sort are not believed to have affected any previously-published conclusions
from our group. Nevertheless, the fluctuations visible in some samples indicate
clearly the presence of an AF layer that should influence the properties of all
nanopillar ST devices.
Exchange biasing in AF/Py films has been demonstrated to dissipate dynamic
magnetic energy through a two-magnon scattering process arising from local vari-
ations in the interfacial exchange coupling [21–23]. Over the course of the pulsed
I measurements, the free layer is switched hundreds of millions of times, which
the HS,i data indicate should result in the AF layer being on average magnetically
ordered but with a finer, more randomized local magnetic structure that leads
to strong damping. The rapid increase in damping, observed over the same low
T range where both the unidirectional AF pinning field and the critical currents
I±c0(T ) also increase rapidly, is attributed to an increasing portion of the AF oxide
layer becoming blocked, thereby simultaneously increasing the amount of interfa-
cial exchange coupling variation seen by the free layer nanomagnet as it moves
in its dynamic switching trajectory, consistent with the two-magnon model. The
process of inducing randomization in the AF by the nanomagnet reversal itself
may also lead to enhanced damping. The slight deviation in α0 for the two switch-
ing directions at low T is speculated to come from fine differences in AF ordering
due to small differences in the net magnetic field (the magnetostatic dipole fields
from both layers and Happ) between P and AP orientation. The unidirectional
311
pinning field is present over the entire T range, with diminishing amplitude with
increasing T . As the AF grains become unblocked, an additional damping mecha-
nism becomes possible if these grains can undergo reversal on the same time scale
that the free layer traces out its dynamic switching trajectory, which can result
in domain drag or the “slow relaxer” dissipation process [24]. This effect could
make a significant contribution to the greater than intrinsic damping that persists
to higher T .
Given that these nanopillar samples were fabricated with the standard method
of ion beam etching continuous films, it is reasonable to assume that this side-
wall oxide is a common affliction for all Py-based nanopillar samples. Even if ion
mill definition and deposition of a passivation layer occur without breaking vac-
uum, typical ambient oxygen coverage of the exposed nanopillar sidewall will occur
within seconds in vacuum systems with moderate base pressures of 10−7 − 10−6
Torr. To our knowledge there have been no reports of actively protecting the side-
walls from oxidation. I suggest that the native AF oxide layer that forms can have
substantial, previously under-appreciated consequences for the ST behavior, lead-
ing to a substantially enhanced damping parameter which directly increases the
critical currents for switching. The presence of this AF perimeter layer may also
alter the boundary conditions that should be employed in micromagnetic mod-
eling of the free layer nanomagnet behavior and affect the dynamical modes of
ST-driven precession. We are currently investigating whether this AF perimeter
layer can account, at least in part, for the narrower than predicted ST-induced
microwave oscillator linewidths that have been observed in similar nanomagnets
at low T [25].
312
initial HS1
HS1
4.2 K
HS2
initial HS2
2nd scan
Figure 6.9: Switching fields HS,i of the free layer nanomagnet (defined in theinset) measured as a function of T . HS1 () are AP to P and HS2 (•) are Pto AP switching fields. All 20 HS,i for each T are shown, although some areindistinguishable due to the size of the symbol. The solid and dashed lines in areguides to the eye for the maximal HS1 and HS2 values, respectively, as functions ofT . At 4.2 K, the free layer switched consistently at fields HS1 ≈ 650 Oe and HS2
≈ 240 Oe, shown as open and filled triangles, respectively, for the first six GMRscans. Subsequent scans, however, showed more stochastic switching behavior thatpersisted for the duration of the experiment.
313
4.2 K
7th scan
4.2 K
11th scan
4.2 K
17th scan
4.2 K
18th scan
(a) (b)
(c) (d)
Figure 6.10: Progression of the random exchange field from the AF perimeterlayer is observed in subsequent minor loop GMR scans (a) through (d). Thispreviously unmeasured device (sample 3, an 80×180 nm ellipse) was cooled to 4.2K in Happ = 0, whereupon a single, orientation-setting major loop scan, followedby 20 GMR minor loop scans were taken. The sample was then sequentiallywarmed back to 160 K in 20 K steps, with 20 GMR minor loops measured at eachT .
314
6.5 Summary
In summary, I have performed time-resolved measurements of the spin torque-
driven switching of a Py nanomagnet at T = 4.2 K to 160 K. LLG macrospin sim-
ulations are in close quantitative agreement with the ST switching events, yielding
values of the parameterized spin torque function g(θ) = A sin θ/(1 + B cos θ) and
the damping parameter. I find α0 to be high, > 0.03, and strongly T -dependent,
which I attribute to the AF pinning behavior of a thin Py oxide layer on the side-
wall of the nanomagnet. The values of A and B are in fair numerical agreement
with the spin torque calculated from the two-channel model using the measured
magnetoresistance values of the nanopillar spin-valve. There is, however, consider-
able device-to-device variation in the spin torque asymmetry parameter B, which
is tentatively attributed to the variable nature of the AF perimeter layer. The
presence of an AF oxide layer can have a major effect on the nanomagnet dy-
namics. Controlling this layer will be important in optimizing spin torque-driven
behavior.
6.6 Exchange biasing in other samples
Due to the somewhat unstable nature of the AF perimeter layer, the exchange
biasing shows large variability in its effect on the free layer magnetic field-induced
switching between different samples. It appears to be scan number dependent,
as indicated in Figure 6.10 for sample 3 above, and may also depend on the field
applied during cooling Hcoolapp . In Figures 6.11 I show the same experiment for
sample 3, but after the data has been taken for Figure 6.9 and also after cooling
the sample in Hcoolapp = -3.2 kOe. As before, there is stochastic behavior in the
315
Figure 6.11: Exchange biasing after a second cool down, in Hcoolapp = -3.2 kOe. As
before, the switching fields show a stochastic behavior indicative of the exchange
biasing from AF pinning layer. Aside from Hcoolapp 6= 0 during cooling and also for
each 20 K warming, the method for acquiring this data was the same as those forFigure 6.9.
switching fields. However, when looking at the individual scans in sequence dips
in the resistance are observed indicating that the magnet may be breaking up
into a more complex micromagnetic structure which thereby affects the measured
resistance. This may be because cooling in field strengthens portions of the AF
layer, allowing it to produce locally strong AF pinning sites that act to break up
the uniform magnetization.
A similar effect occurs when the sign of the cooling field is reversed. In Fig-
ure 6.13, I show a plot of the switching fields HS1 and HS2, for a sample cooled
316
4.2 K
3rd scan
4.2 K
2nd scan
4.2 K
4th scan
4.2 K
6th scan
4.2 K
7th scan
4.2 K
5th scan
(a) (b)
(c) (d)
(e) (f)
Figure 6.12: Exchange biasing after a second cool down, in Hcoolapp = -3.2 kOe.
The shifting of resistance during some of the scans points to local pinning sightsof the AF sidewall oxide, possibly strengthened, albeit temporarily, by the in-fieldcooling.
317
in Hcoolapp = +3.2 kOe. Although the temperature range is not as detailed, it is
clear that the trends of HS1 and HS2 with T are different from that of Figures 6.11
or 6.9, where the slope of HS1 versus T was nearly twice that of HS2 versus T . Al-
though these trends are almost reversed for the sample cooled in positive field, the
stochastic fluctuations still affect more the AP to P reversal (HS1). This sample
was previously unmeasured and, when investigating the GMR minor loop scans
in sequence, a distinctly different behavior from what is shown for the Hcoolapp = 0
kOe sample in Figure 6.10 is seen. Right away the GMR which shows the local
pinning effects of the perimeter layer. After only four scans, the AF pinning is
broken and sample then returns to “normal” behavior of stochastic field switching.
An example of a sample exhibiting no stochastic switching field behavior at all is
shown in Figure 6.15.
318
Figure 6.13: Exchange biasing after a second cool down, in Hcoolapp = +3.2 kOe.
As before, the switching fields show a stochastic behavior indicative of the exchange
biasing from AF pinning layer. Aside from Hcoolapp 6= 0 during cooling and also
during each 20 K warming, the method for acquiring this data was the same asthose for Figure 6.9.
319
4.2 K
3rd scan
4.2 K
2nd scan
4.2 K
4th scan
4.2 K
6th scan
4.2 K
1st scan
4.2 K
5th scan
(a) (b)
(c) (d)
(e) (f)
Figure 6.14: Exchange biasing after a second cool down, in Hcoolapp = +3.2 kOe.
The shifting of resistance during the initial scans points to local pinning of the freelayer moment from the AF sidewall oxide, possibly strengthened by the in-fieldcooling.
320
Figure 6.15: A sample with no stochastic field switching, illustrating that forsome samples the sidewall oxide layer is either so weak that it plays almost norole in the field-induced reversal, or that it is not present at all. Unlike previoussamples with very clear exchange biasing, this sample, for has nearly symmetrictrends of HS1 and HS2 with T for T > 40 K.
321
6.7 Further details of simulation
6.7.1 Stochastic distribution of initial angle
Randomization of the initial angle θi leads to stochastic distribution of trajectories
from one switching event to another. For devices where the initial angle fluctuates
about 0 or 180, relative to the spin polarization axis, oscillations in the switching
trajectory, which are highlighted in the simulated macrospin switching event, will
get averaged out. I show how this comes about from averaging over thousands of
individual macrospin simulations, shown in Figures 6.16(a) - (c). 2000 averages
over these simulated macrospin switching events yields a much smoother curve
where the phase of the oscillations in the individual simulated switches is lost,
reminiscent of the actual measurement and is shown in Figure 6.16(d).
6.7.2 Simulation parameters A, B, and α0
The effects of the simulation parameters A, B, and α0 are described here for sam-
ple 2, an 80×180 nm ellipse with E0(4.2 K) = 1.1 eV, and HK = 261 Oe. These
tests are for T = 160 K, so that Rx = 7.357 Ω, E0(160 K) = 0.93 eV, σθrand≈ 5,
and T ′ = 164 K at I = 1.5 mA. The best fit values of the fitting parameters are
A = 0.54, B = 0.30, and α0 = 0.035. The different effects of changing these param-
eters on the switching speed ≡ 1/tswitch is shown in Figures 6.17, 6.18, and 6.19.
Figure 6.20 shows how the asymmetry parameter B affects the spin torque curve
g(θ), where a higher B yields increased (decreased) spin torque strength for AP
to P (P to AP) switching. All simulations shown in this section are averages over
200 events.
322
Time
2000 averages
(a)
(b)
(c)
(d)
mz
mz
mz
mz
Figure 6.16: Averaging over many simulated switches each following a trajectorydetermined by the initial conditions set by the random contribution to the initialangle θrand(T ).
323
Measurement
AP to P
A = 0.54
A = 0.59
A = 0.49
0
B = 0.30
α = 0.035
(a)
P to AP
(b)
Figure 6.17: A = 0.59 (solid red line), 0.54 (solid black line & best fit value),and 0.49 (dashed blue line) for (a) AP to P and (b) P to AP switching at T = 160K, B = 0.30, and α0 = 0.035. For both switching directions, increasing A acts toincrease the switching speed, which is expected since A is directly related to themagnitude of the spin torque per unit current.
324
Measurement
AP to P
(a)
P to AP
(b)
B = 0.30
B = 0.35
B = 0.25
0
A = 0.54
α = 0.035
Figure 6.18: B = 0.35 (solid red line), 0.30 (solid black line & best fit value),and 0.25 (dashed blue line) for (a) AP to P and (b) P to AP switching at T =160 K, A = 0.54, and α0 = 0.035. For P to AP (AP to P) switching directions,increasing B acts to reduce (increase) the switching speed, which is expected sinceB is the asymmetrically affects the magnitude of the spin torque, as indicated inthe spin torque curve g(θ) shown in Figure 6.20.
325
Measurement
AP to P
α = 0.035
α = 0.040
α = 0.0300
0
0
A = 0.54
B = 0.30
(a)
P to AP
(b)
Figure 6.19: α0 = 0.040 (solid red line), 0.035 (solid black line & best fit value),and 0.030 (dashed blue line) for (a) AP to P and (b) P to AP switching at T =160 K, A = 0.54, and B = 0.30. For both switching directions, increasing α0 actsto decrease the switching speed, which is expected since α0 works against the spintorque for most of the dynamic switching event.
326
θ/π
g(θ)
10.5
0.5
00
Figure 6.20: Asymmetric spin torque function, g(θ) = A sin θ/(1 + B cos θ). Theblue curve is A = 0.5 and B = 0.2 while the red curve is A = 0.5 and B = 0.9.This illustrates that the larger B, or asymmetry value, the more spin torque perunit current is available for AP to P switching since θi ≈ 180.
327
6.7.3 Simulation constants
Here I show the effects of changing other constants in the simulation: anisotropy
energy E0(4.2 K) (Figure 6.21), in-plane anisotropy field HK (Figure 6.22), exter-
nal field Hext (Figure 6.8), misalignment angle θmis (Figures 6.23, 6.24, and 6.25),
lateral area (Figure 6.26), diffusive current heating (Figure 6.27), and angular de-
pendence of damping (Figure 6.29). These tests are performed on sample 2, an
80×180 nm ellipse with E0(4.2 K) = 1.1 eV, and HK = 261 Oe. Simulation T =
4.2 K1, so that Rx = 5.932 Ω, σθrand≈ 0.7, and T ′ = 29 K at I = 1.5 mA. For
all of these simulations, the best fit values of A = 0.68, B = 0.31, α0(P to AP) =
0.070, and α0(P to AP) = 0.063. All simulations shown in this section are averages
over 200 events.
1Except for an in depth study of how θmis affects A, B, and α0 at T = 4.2 Kto 160 K, shown in Figure 6.25.
328
Measurement
0.9
0.7
1.1
0.5
1.0
0.8
0.6
0E [eV]
P to AP
AP to P
(a)
(b)
Figure 6.21: Effects of changing the uniaxial anisotropy energy E0 from 0.5 to1.1 eV. The effects are fairly minimal.
329
450
650
261
50
350
550
150
KH [Oe]
Measurement
P to AP
AP to P
(a)
(b)
Figure 6.22: Effects of changing the in-plane anisotropy field HK 50 to 650 Oe.There is a noticeable and consistent effect of decreasing the switching speed forhigher HK. This is expected because the amount of spin torque necessary to beginprecessing the nanomagnet is larger is there is a larger “restoring” force comingfrom shape anisotropy.
330
Measurement
1
2
3
4
5
0
7
9
11
misθ [ ]o
P to AP
AP to P
(a)
(b)
Figure 6.23: Effects of changing the misalignment angle θmis =0, 1, 2, 3, 4, 5, 7, 9, and 11, which approximates a systematic misalignmentfrom the experimental setup. There is a significant increase in the switching speedeven for small θmis ≈ 3. This is expected at such low T since the standard devi-ation of the random angle θrand is < 1. Consequently, the initial torque on thenanomagnet is much larger than what normal thermal fluctuations could provide.The estimated alignment accuracy for this experiment is to within 5, but clearlythere is a significant difference from this and the θmis = 0 simulation from whichall of the A, B, and α values are estimated. More simulations are performed toestimate A, B, and α at θmis = 5 and 10, shown in Figure 6.24.
331
0
5
10
misθo
o
o
misθ0510
[ ]o
(a)
(b)
Figure 6.24: Effects of changing the fit parameters (a) A and (b) B when amisalignment angle θmis = 0, 5, or10 is included. A drops by nearly 50% atθmis = 10, but all three curves are non-monotonic, showing an upturn at lowertemperatures. The alignment accuracy is better than 10, but the trend in Tshows that, at least for low T , although the trends in the best fit parameters areagreeable between the different θmis their values can vary widely. Perhaps thisindicates that more effort be put in to determine the actual θmis for each sample.B is not significantly affected by θmis 6= 0.
332
misθ0510
[ ]o
misθ0510
[ ]o
(a)
(b)P to AP
AP to P
Figure 6.25: Effects of changing the fit parameter α0 for (a) AP to P and (b)P to AP switching when a misalignment angle θmis = 0, 5, or 10 is included.α0 for both switching directions show very similar behavior with increasing θmis.Regardless of the added misalignment, the simulations still show that the dampingis enhanced at low T , which is attributed to the exchange biasing effects of the AFsidewall oxide layer.
333
60x180
80x200
80x180
100x200
60x160
80x160
100x180
2area [nm ]
Measurement
P to AP
AP to P
(a)
(b)
80x200
80x180
60x160
2area [nm ]
Figure 6.26: Effects of changing the lateral area of the elliptical nanopillar for(a) AP to P and (b) P to AP switching. Lateral area for sample 2 is estimated atarea0 = 80×180 nm (minor × major axis), and we show the effects of increasingand decreasing area by setting area = 100×200 nm (1.39 · area0); 100×180 nm(1.25 · area0); 80×200 nm (1.11 · area0); 80×160 nm (0.89 · area0); 60×180 nm(0.75 · area0); 60×160 nm (0.67 · area0). Switching speed changes uniformly aboveand below the best fit value, which is expected since area re-scales the net spintorque amplitude.
334
Measurement
heating
no heating
P to AP
AP to P
(a)
(b)
Figure 6.27: Effects of turning off Ohmic heating from a current in a diffusivewire for (a) AP to P and (b) P to AP switching done at T = 4.2 K. Inclusion ofheating effects clearly impacts the switching speed, where T ′ ≈ 60 K at I = 3 mA.It is important to note that the range of T where heating becomes significant (T <60 K) coincides with the T range where the upturn in α0 is observed. However,the effect Ohmic heating has on the slope of 1/tswitch versus I is opposite thatof α0 (see Figure 6.19) and so reduced heating effects would only increase thebest fit damping value. If the heating per unit current were increased (i.e. ifthe coefficient in Equation 6.6 were larger) the best fit damping values wouldbe smaller. Exploration of a range of values of this heating coefficient will beimportant in broadening the understanding of Ohmic heating in spin torque-drivenreversal.
335
Measurement
ν = 0.33
ν = 0.0
P to AP
AP to P
(a)
(b)
Figure 6.28: Effects of turning off the angular dependence of damping for (a)AP to P and (b) P to AP switching. The angular dependence is α(θ) = α0[1 −ν sin2 θ/(1 − ν2 cos2 θ)] from ref. [16]. For the simulation, the angular parameterν is typically set to 0.33, which is what is prescribed for Py/Cu/Py nanopillars.By setting ν = 0, which turns off the angular dependence, the differences in thesimulated switching speeds are nearly imperceptible.
336
6.7.4 Simulating a real pulse shape
Finally, I show the results of simulating the switching if the macrospin were sub-
jected not to an ideal square step-edge, but rather to a pulse shape like the one
measured on the oscilloscope, which is not a perfectly squared edge. This is due to
limitations in the pulser in producing, as well as the cables and amplifier distorting
the higher frequency Fourier components of the square wave. The current pulse
shape for the simulation was taken from a measured background trace and nor-
malized to one so that any desired I value would be a simple multiplicative factor.
The value of I is updated for each of the 1 psec time steps in the simulation. The
real-pulse-shape results are shown sample 2, an 80×180 nm ellipse with E0(4.2 K)
= 1.1 eV, and HK = 261 Oe. These tests are for T = 120 K, so that Rx = 6.883
Ω, E0(120 K) = 1.0 eV, σθrand≈ 4, and T ′ = 124 K at I = 1.5 mA. I compare
the simulation and measurement out to high currents, where the ideally square I
pulsed simulations show begin to diverge from the measurement, while the results
of simulating a real pulse shape yields switching speeds much more in line with
what is measured. The simulations shown in this section are averages over 500
events.
337
Measurement
instantaneous
real
P to APAP to P
(a) (b)
Figure 6.29: Effects of simulating with a real pulse shape for (a) AP to P and(b) P to AP switching. The normalized pulse shape, which is then multipliedby the I values plotted on the I-axis, is shown in the inset to (a). For low I,similar to the ranges shown in Figure 6.3, there is very little difference betweenan ideal, or “instantaneous” square step-edge pulse shape (solid black lines) andthe real pulse shape (solid red lines). However, beyond this range, a divergencebetween the instantaneously pulsed simulation and data begin to form at ∼2.5 mAfor P to AP and ∼1.5 mA for AP to P switching, where the measurements showa slight upturn with increasing I. The real pulse shape simulations are in muchbetter accord with the measurement, indicating that the observed upturn in thedata is simply a consequence of the non-ideal pulse shape as opposed to a newmode of fundamentally different dynamic behavior for switching. I note, however,that for I > 2.1 mA, the observed AP to P switching speed shows a slight downturn which is not seen in the real pulse shape simulations, and only slightly in theinstantaneous pulse simulations. The reasons for this down turn in the AP to Pdata are not clear but such curvature has not been observed in other samples.
338
References for Chapter 6
[1] Katine J.A., Albert F.J., Buhrman R.A., Myers E.B., & Ralph D.C., Current-driven magnetization reversal and spin-wave excitations in Co/Cu/Co pillars,Phys. Rev. Lett. 84, 3149–3152 (2000).
[2] Urazhdin S., Birge N.O., Jr. W.P.P., & Bass J., Current-driven magneticexcitations in permalloy-based multilayer nanopillars .
[3] Kiselev S.I., Sankey J.C., Krivorotov I.N., Emley N.C., Schoelkopf R.J.,Buhrman R.A., & Ralph D.C., Microwave oscillations of a nanomagnet drivenby a spin-polarized current, Nature (London) 425, 380–383 (2003).
[4] Grollier J., Cros V., Hamzic A., George J.M., Jaffres H., Fert A., FainiG., Youssef J.B., & Legall H., Spin-polarized current induced switching inCo/Cu/Co nanopillars, Appl. Phys. Lett. 78, 3663–3665 (2001).
[5] Slonczewski J.C., Current-driven excitation of magnetic multilayers, J. Magn.Magn. Mater. 159, L1–L7 (1996).
[6] Berger L., Emission of spin waves by a magnetic multilayer traversed by acurrent, Phys. Rev. B 54, 9353–9358 (1996).
[7] W. H. Rippard M. R. Pufall S.K.T.J.S. & Russek S.E., Injection locking andphase control of spin transfer nano-oscillators, Phys. Rev. Lett. 95, 067203(2005).
[8] Li Z. & Zhang S., Thermally assisted magnetization reversal in the presenceof a spin-transfer torque, Phys. Rev. B 69, 134416 (2004).
[9] Krivorotov I.N., Emley N.C., Garcia A.G.F., Sankey J.C., Kiselev S.I., RalphD.C., & Buhrman R.A., Temperature dependence of spin-transfer-inducedswitching of nanomagnets, Phys. Rev. Lett. 93, 166603 (2004).
[10] Koch R.H., Katine J.A., & Sun J.Z., Time-resolved reversal of spin-transferswitching in a nanomagnet, Phys. Rev. Lett. 92, 088302 (2004).
[11] Braganca P.M., Krivorotov I.N., Ozatay O., Garcia A.G.F., Emley N.C.,Sankey J.C., Ralph D.C., & Buhrman R.A., Reducing the critical currentfor short-pulse spin-transfer switching of nanomagnets, Appl. Phys. Lett. 87,112507 (2005).
[12] Jitter noise in the triggering of the oscilloscope prevents the complete removalof distortions in the pulse shape coming from the pulser and amplifier, givinga slowly oscillating background and a shoulder in the current step. Theseintroduce a maximum error of only 7% in the measurements of tswitch and arenot significant. AP to P switching data (not shown) are very similar in bothtime and I scales.
339
[13] The difference between mz = cos θ and ∆R ∝ [1− cos2(θ/2)]/[1+χ cos2(θ/2)]as functions of time is nearly imperceptible since χ = 2B/(1 − B) < 1 asmeasured in similar samples [11]. Urazhdin S., Loloee R., Pratt W., Jr.,Noncollinear spin transport in magnetic multilayers, Phys. Rev. B 71, 100401(2005).
[14] Donahue M.J. & Porter D.G., OOMMF user’s guide, version 1.0, technicalreport no. NISTIR 6376, Technical report, National Institute of Standardsand Technology, Gaithersburg, MD (1999).
[15] Russek S.E., Kaka S., Rippard W.H., Pufall M.R., & Silva T.J., Finite-temperature modeling of nanoscale spin-transfer oscillators, Phys. Rev. B 71,104425 (2005).
[16] Tserkovnyak Y., Brataas A., & Bauer G.E.W., Dynamic stiffness of spinvalves, Phys. Rev. B 67, 140404 (2003).
[17] Slonczewski J.C., Currents and torques in metallic magnetic multilayers, J.Magn. Magn. Mater. 247, 324–338 (2002).
[18] Xiao J., Zangwill A., & Stiles M.D., Boltzmann test of Slonczewskis theoryof spin-transfer torque, Phys. Rev. B 70, 172405 (2004).
[19] Valet T. & Fert A., Theory of perpendicular magnetoresistance in magneticmultilayers, Phys. Rev. B 48, 7099–7113 (1993).
[20] Garcia A.G.F., unpublished.
[21] McMichael R.D., Stiles M.D., Chen P.J., & Egelhoff W.F., Ferromagneticresonance linewidth in thin films coupled to NiO, J. Appl. Phys. 83, 7037–7039 (1998).
[22] Rezende S.M., Azevedo A., Lucena M.A., & de Aguiar F.M., Anomalous spin-wave damping in exchange-biased films, Phys. Rev. B 63, 214418 (2001).
[23] Weber M.C., Nembach H., Hillebrands B., & Fassbender J., Modified Gilbertdamping due to exchange bias in NiFe/FeMn bilayers .
[24] McMichael R.D., Lee C.G., Stiles M.D., Serpa F.G., Chen P.J., & EgelhoffW.F., Exchange bias relaxation in CoO-biased films, J. Appl. Phys. 87, 6406(2000).
[25] Sankey J.C., Krivorotov I.N., Kiselev S.I., Braganca P.M., Emley N.C.,Buhrman R.A., & Ralph D.C., Linewidths for magnetic precession drivenby DC spin-polarized currents, cond-mat/ 0505733 (2005).
CHAPTER 7
NOVEL DEVICE STRUCTURE: 3-TERMINAL NANOPILLAR
7.1 Introduction
At the time of this writing, magnetic random access memory (MRAM) remains
a developing concept that has yet to experience industry-wide adoption as a sub-
stitutive technology for Si-based memory technologies such as SRAM, DRAM, or
FLASH. Nevertheless, the universal characteristics of non-volatility, high density,
endurance, and speed make it an appealing goal [1]. MRAM is a based on CPP
magnetic tunnel junctions (MTJs) whose impedance is tuneable both by the size of
the element and the thickness of the tunnel barrier material. The lower coercivity
“free” layer is switched by an Oersted field coming from a nearby current-carrying
line. Oersted field writing is done in a cross-point architecture, where all MRAM
elements that lie along two perpendicular current-carrying write lines (a “Bit” line
and a “Word” line) see a fraction of the writing field and only the single MRAM
element at the intersection of these two write lines sees the combined Oersted fields
which are together sufficient to write the element.
Oersted field writing does not damage the device and so MRAM enjoys essen-
tially infinite device writability. However, this cross-point writing scheme places
tight tolerances on the acceptable uniformity of each of the magnets [1] in order
to avoid the problem of “half selection”, where elements not directly under the
selected cross-point switch erroneously because of an abnormally low coercivity.
If, instead of Oersted field switching, the element was actuated by a spin transfer
current flowing through the device itself, the “half select” problem would be mit-
igated owing to smaller switching currents from the spin torque versus the field
340
341
torque. This spin torque-based MRAM would most likely not use the cross-point
writing architecture but rather would address each MRAM element individually.
This would relax somewhat the magnetic uniformity tolerances required for Oer-
sted field writing and it is easy to see why. For Oersted field switching, the ability
to reliably access a single device is contingent upon how similar one can engineer
the coercivities of all the switching layers in all of the MTJs, demanding high
magnetic uniformity as previously stated. For spin torque switching, however, re-
liably accessing a single element is determined by how distant one can engineer the
switching currents Iswitch from the breakdown currents Ibreakdown. Spin torque
and barrier destruction mechanisms are physically distinct and so the chances are
better for tuning both independently, thereby increasing the window for an accept-
able chip-wide actuation current Iswitch < Iactuate < Ibreakdown.
For a spin-switched MRAM element, however, the advantage of infinite writabil-
ity is lost because the writing current will inevitably wear out the barrier over time.
But if there was a way of switching the element with a spin torque without hav-
ing to send the charge current across the barrier, wear out would no longer be a
problem. In metallic spin transfer trilayer devices, the spin torque is imparted on
the free layer by spin-dependent scattering of the spin polarized electrons off of the
spacer/free layer interface. There is no stipulation that the electrons must transmit
through the magnetic layer. Slonczewski proposed a device where electrons could
impart their spin angular momentum onto the free layer by scattering off of the
interface, but then conduct away by some other channel. If typical spin transfer
devices work via transmission mode, where electrons propagate into the free layer,
then this proposed spin transfer device would operate via reflection mode.
Since, in reflection mode, the electrons would propagate only very little into
342
the nanomagnet, a tunnel barrier on the opposite side of the nanomagnet would
see very little of the actuation current, thereby minimizing barrier wear out. An
idea originally proposed by Slonczewski [2] describes how reflection mode spin
transfer may be incorporated into a device, which is shown in Figure 7.1. The
bottom magnetic layer polarizes the current which actuates the free layer of the
top MTJ by application of the DC current (IDC). Voltage (V) across the MTJ is
monitored and can be converted to a junction resistance by application of a small
sense current (isense). The spacer layer between the polarizer and the MTJ is
unpatterned, which makes the device fabrication highly complicated. The presence
of top, bottom, and middle leads give this device the name “3-terminal nanopillar”.
Since IDC is designed to not propagate through the tunnel barrier, there is no
restriction on the junction R·A and also the problem of barrier wear out is reduced,
greatly increasing the endurance of these devices.
Electrons that propagate through the polarizing layer into the middle lead
follow the path of the electronic potential drop, but because the spacer/free layer
interface is in proximity to the fixed layer, on the order of the ballistic mean free
path of the spacer material, the near-ballistic electrons strike the interface with
the free layer. This event imparts the spin angular momentum to free layer but
the electrons then conduct out the side through the infinite spacer or middle lead.
For electrons flowing in an opposite sense, from the middle lead to the polarizer
layer, free layer reversal relies on spins scattered back from the spacer/fixed layer
interface, much like the 2-terminal nanopillar, as shown in Figure 7.1.
If realized, this device could further the possibility of MRAM or other spin-
switched-based magnetoelectronics as viable alternative technologies in the com-
puter and communications industry. In the following sections, I highlight a possible
343
Figure 7.1: The 3-terminal device as described in [2] and show here in its simplestform, consists of a bottom polarizing layer separated from a MTJ by a plane ofnon-magnetic metal, which is labelled the middle lead. The free layer of the MTJ isadjacent to this middle lead. The top of the MTJ and the bottom of the polarizerare connected to top and bottom metallic leads, respectively. A DC current (IDC)is passed between the middle and bottom leads. The geometry is such that thespin-polarized electrons emanating from, or scattering off of the polarizer/middlelead interface may strike the free layer and transfer their spin angular momentumto the free layer. Although the electrons provide the spin torque, they do notpropagate through the MTJ but instead conduct either out the middle lead orthrough the polarizer to the bottom lead, depending on the direction of currentflow. The idea is that all spin-torque-induced excitations possible in 2-terminaldevices in transmission mode spin transfer should likewise be possible in reflectionmode spin transfer like that exhibited by the 3-terminal nanopillar.
344
fabrication scheme for such a device while also discussing some results of devices
fabricated at Hitachi Global Storage Technologies (Hitachi) by Jordan Katine.
7.2 Proposed fabrication process
This proposed process represents a good starting point for 3-terminal nanopillar
fabrication. The main idea is that one nanopillar (MTJ) is made first, and then
a second nanopillar (spin-valve) is made directly on top of the first one. I am
placing the MTJ as the bottom pillar because we already know how to deposit
smooth layers onto a blank substrate [3]. If the MTJ were on top, the layers would
be deposited onto a patterned surface which could impede good MTJ behavior.
Although the 3-terminal device, in principle, requires only a single polarizing layer
in the top pillar structure, experimental verification of spin-conserving conduction
throughout the two-pillar stack is simplified by engineering both top and bottom
pillar structures to have their own magnetoresistive response.
Since this process employs the HSQ nanopillar process discussed in Ch. 3, I
avoid redundancies by not discussing in too much detail problems and unknowns
associated with the HSQ-based 2-terminal pillar fabrication. Potential problems
and unknowns are discussed here only when they are unique to the 3-terminal
fabrication process. Due to its length, I have reserved all written information
about each step for the figure captions.
345
Figure 7.2: Step 1: Photolith 1, define global alignment marks (1/6). Since thisfabrication requires an aligned VB6 exposure, photolithography must be performedto define alignment marks. In this step, spin on photoresist layers: LOR10A(undercut resist)/S1813.
346
Figure 7.3: Step 1: Photolith 1, define global alignment marks (2/6). Top downview of the photolithography. This is the only step done with contact lithogra-phy (HTG). All others will be with projection lithography (5× stepper). In thisschematic, the global alignment marks themselves are not visible, but the largevisual indicator marks are. These visual indicators are present to identify thelocation of the global alignment marks. This exposure also defines some pho-tolithography alignment marks. The lines extending horizontally to the edges ofthe wafer are to help in finding these alignment marks when using the 5×.
347
Figure 7.4: Step 1: Photolith 1, define global alignment marks (3/6). A zoomin on the global alignment marks, which will be 20 µm holes in the multilayerfilm. Absolute position of the global mark and the visual indicator marks may bedefined later.
348
Figure 7.5: Step 1: Photolith 1, define global alignment marks (4/6). A zoomin on a single device. The photolithography protects squares in the vicinity ofthe device that will be used in later steps to define the fine alignment marks.Coordinates of the resist squares are defined relative to the nanopillar position.This is done to place the alignment marks beneath the top most surface of themultilayer if placed on top of all of the layers, the quality of the marks maydegrade when the wafer is planarized in the CMP (step 17).
349
Figure 7.6: Step 1: Photolith 1, define global alignment marks (5/6). The resistsquares are 40 µm on a side, so that all four squares fall within a 320 µm square,which lies entirely within the maximum exposure field of the VB6 (327.68 µm).
Figure 7.7: Step 1: Photolith 1, define global alignment marks (6/6). Side viewof the resist after development. Make sure that the wafer surface is sufficientlycleaned by an O2 plasma before sputter deposition, which is done in the descum(20 sec (minimum) at 100W, 30 sccm, 30 mTorr O2).
350
Figure 7.8: Step 2: Multilayer deposition. Deposit the layers for the bottompillar, the MTJ. To achieve maximum uniformity of the deposited metal, do notuse clips to hold the wafer to the stage. Doing so will affect CMP etch unifor-mity. Instead, use the custom machined 4” wafer clamp and use the clips to gripthe clamp. The clamp shadows a symmetric annular region right at the edge ofthe wafer. If Ti or Cr are not in the AJA sputter system, a Ti adhesion layercan be deposited in the IBD. Transfer from the IBD to the AJA post-depositioncan be done in < two minutes and adhesion should still be good. Underlayersshould be grown for smoothness and should also include an exchange biasing layer(antiferromagnetic layer such as FeMn or IrMn) to pin the reference CoFe layerin the MTJ. Adjust the barrier thickness so that the R·A is high enough to pre-vent spin-torque switching of the tunnel junction. The thickness of the free Pylayer should be thin enough so that it is superparamagnetic at room temperature.This allows for exploration of the current-dependent energy barrier by monitoringtelegraph switching [4] and low switching currents controllable with bath temper-ature. Suggested layer thicknesses are: adhesion, Ti 100 A; underlayers, [Ta 50A/Cu 20 A/CuOx (Cu grown in Ar + O2) 200 A/*/]2/Ta 250 A; MTJ layers IrMn100 A/CoFeB 50 A/(AlOx or MgO barrier)/Py 20 A; cap, Cu 200 A/Ta 50 A.The asterisk (*) indicates that all of the oxidizable targets should be run for 75sec with the shutters closed to burn off any oxide that formed during the CuOxdeposition. The top Cu will serve as part of the top spacer to the spin-valve. TheTa is recommended for the HSQ e-beam process. Thin CoFe dusting layers aroundthe Py free layer may be necessary. Perform lift off after deposition.
351
Figure 7.9: Step 3: Spin e-beam resist. The lower molecular weight 495k PMMAunderlayer should be spun on at least 300 nm thick (e.g. 5.5% 495k PMMA inAnisole spun at 2500 rpm (330 nm)). Top layer resist is the standard 950k PMMAin MIBK spun at 2000 rpm (106 nm). The clearing dose is not known for theseresist layers.
Figure 7.10: Step 4: E-beam expose alignment marks. Using global alignment,expose 10 µm squares in each of the 40 µm square holes defined in the film. These10 µm squares should all be exposed in a single exposure field and will define themarks to which subsequent pillar exposures will align to. Develop.
352
Figure 7.11: Step 5: Define fine alignment marks. Evaporate Ti 100 A foradhesion, then at least Au or Ta 1000 A for the alignment mark. Lift off.
Figure 7.12: Top view. The global alignment should be effective enough toplace the 10 µm e-beam-defined squares inside the photolithography-defined 40µm square holes.
353
Figure 7.13: Step 6: Spin e-beam resist 495k PMMA/HSQ. The optimum thick-ness of the HSQ (XR 1541) is not known at the time of this writing.
354
Figure 7.14: Step 7: E-beam define bottom pillar mask. With fine alignmentto the pre-defined Au (or Ta) alignment marks, expose the bottom pillar. Due tothe expense of doing three VB6 runs per wafer, it is best to only expose a smallfraction of the wafer. A 3×3 grid of die at the middle of the wafer is suggestedover the usual 9×9 grid of die. Perform PEB and HSQ development procedures.
Figure 7.15: Bottom exposure e-beam pattern. This is how the CAD looks forthis e-beam exposure. The 40 µm squares are meant to expose a region local to thefine alignment mark, which will protect it from further processing so that it can beused a second time for the top exposure. Note that there is a 90 offset betweenthe coordinate systems of the VB6 and that of the room in which it resides.
355
Figure 7.16: Step 8: O2 plasma etch the PMMA underlayer.
Figure 7.17: Side view of alignment marks. The fine alignment marks are nowcovered in the exposed HSQ, which should protect them from subsequent ionmilling steps.
356
Figure 7.18: This top view shows the exposed pillar device and the four alignmentsquares inside their photolithography-defined 40 µm holes.
Figure 7.19: Step 9: Ion mill define bottom pillar. Etching in Ar + O2 will helpprevent sidewall shorting around the tunnel barrier.
357
Figure 7.20: Step 10: Insulate nanopillar with AlOx. IBD deposit AlOx withstage rotation. Vacuum may be broken to perform AFM if necessary. Proper angleof deposition is not known at this time.
Figure 7.21: Step 11: Photolith 2, isolate devices (1/2). Spin P20/S1827. Sincethe AlOx will be removed in this photolithography pattern, incidental etching bythe 300MIF developer is acceptable. Develop and descum.
358
(a)
(b)
Figure 7.22: Step 11: Photolith 2, isolate devices (2/2). (a) Top down view afterdevelopment. The resist covers what will become the leads and the alignmentmarks. The black arrow indicates the direction of exchange biasing in the Hitachinanopillars. (b) The mask used for photolith 2. Black is Cr and white is glass.
359
Figure 7.23: Step 12: HF etch AlOx. HF etch the AlOx to expose the top of theunderlayers.
Figure 7.24: Step 13: Ion mill to isolate devices. Once the AlOx is gone, ion millwith underlying metallic layers to isolate the devices.
360
Figure 7.25: Step 14: Measure AlOx height. Do profilometry on the AlOx-substrate step edge to find h1 which will be the target deposition height for AlOxrefill.
Figure 7.26: Step 15: Photolith 3, AlOx refill (1/2). Spin photoresist: a pro-tective 495k PMMA underlayer/LOR10A liftoff resist/S1813. Expose with thesame mask as in photolith 2. For subsequent photolithography steps using thePMMA/LOR10A/S1813 recipe, the LOR10A and S1813 develop in 300MIF, butthe PMMA underlayer needs to be etched in an O2 plasma. This is the reasonwhy PMMA is used, because its job is to protect the underlying AlOx from beingetched by the 300MIF [5].
361
(a)
(b)
Figure 7.27: Step 15: Photolith 3, AlOx refill (2/2). (a) Top down view afterdevelopment. (b) The mask used for photolith 3 is the same as that for photolith2. Black is Cr and white is glass.
362
Figure 7.28: Step 16: AlOx refill. IBD-deposit AlOx to the height measuredin step 14 (h1) while rotating the stage. It is not crucial to get the AlOx heightexactly to h1 since height fluctuations across the wafer will prevent a uniform AlOxheight above the substrate anyway. How much flexibility is allowed for the AlOxdeposition height is not known. Liftoff.
Figure 7.29: Step 17: CMP planarization. CMP planarize the wafer. This shouldhave the combined effects of smoothing out the AlOx as well as removing the HSQcap [6]. This step is unconfirmed and should be worked out entirely with standard2-terminal nanopillars before undertaking 3-terminal nanopillar fabrication.
Figure 7.30: Step 18: Photolith 4, sputter top layers (1/2). Spin photoresist: aprotective 495k PMMA underlayer/LOR10A liftoff resist/S1813. This step opensup a rectangular region just above the already patterned nanopillar into whichlayers for the top nanopillar will be sputtered.
363
(a)
(b)
Figure 7.31: Step 18: Photolith 4, sputter top layers (2/2). (a) Top down viewafter development. (b) The mask used for photolith 4. Black is Cr and white isglass.
364
Figure 7.32: Step 19: Clean off Ta cap. Short ion mill at 0 incidence to etch offthe Ta cap. Removing all of the Ta is crucial since the underlying Cu layer willserve as a conduction channel to switch the top Py layer in the MTJ. Consequently,spin-scattering in this Cu layer must be minimal.
Figure 7.33: Step 20: Deposit top layers. After ion milling to clean off the Ta capin the IBD, it is necessary to transfer the sample through air into the AJA. Oncein the AJA, it is best to perform a brief RF backsputter to clean off any residualoxides on the Cu surface. Proper etch time is not known. Suggested layers forthe top nanopillar are: spin-valve, Cu 200 A/CoFe 80 A/Cu 100 A/CoFe 200 A;top cap, Cu 200 A/Ta 50 A. Since the spin valve is not meant to be switched bythe spin torque, it is advisable to make both layers quite thick. A top AF pinninglayer is acceptable. But make them magnetically distinct enough so that theircoercivities are different and their GMR is detectable.
365
Figure 7.34: Step 21: Photolith 5, open alignment marks (1/2). Spin photoresistP20/S1813. It is not known if CMP planarization will remove the 40 µm HSQsquare cap above the alignment marks. This photolith step is designed to clearthe HSQ and any residual AlOx on top of the alignment marks.
366
(a)
(b)
Figure 7.35: Step 21: Photolith 5, open alignment marks (2/2). (a) Top downview after development. (b) The mask used for photolith 5. Black is Cr and whiteis glass.
367
Figure 7.36: Step 22: Open alignment marks. HF etch AlOx/HSQ above thealignment marks, O2 plasma etch PMMA, strip resist.
Figure 7.37: Step 23: Spin e-beam resist. Spin and bake e-beam resist(PMMA/HSQ) the same as done in Figure 7.13.
368
Figure 7.38: Step 24: E-beam define top pillar mask. With fine alignment tothe alignment marks uncovered in the previous steps, expose the top spin-valvepillar directly on top of the bottom MTJ pillar. It is unknown how the quality ofthe alignment marks will degrade due to encapsulation under PMMA/HSQ/AlOxfollowed by uncovering with an HF acid etch of the oxide layers. Due to the 10 nmoverlay accuracy limit, it is best to engineer the top layer with dimensions smallerthan the bottom layer. This increases the likelihood that the top pillar will becompletely circumscribed within the bottom pillar. Perform the PEB and HSQdevelopment procedures.
369
Figure 7.39: Top exposure e-beam pattern. This is how the CAD looks for thetop nanopillar e-beam exposure. Since there are no further alignment steps, it isnot necessary to protect the region surrounding the fine alignment marks. However,for ease of getting the exposure to come out with the correct device position, theCAD has exactly the same exposure field as the bottom pillar e-beam pattern–a320 µm square. Because the VB6 calculates the geometric center of the .FRE filesand uses that as the center of the image, it is simplest to make all overlaid alignedexposures have the same geometric center. This is why I have included the 40 µmcrosses. The crosses also give crude feedback about the accuracy of the alignment.If they fall directly on top of the 40 µm squares, it means that the alignmentprocedure worked. If they are not present, or are at some obvious offset, then thealignment did not work and a choice can be made to try the alignment again orto scrap the wafer. Either way, it is a helpful bit of feedback. Note that there isa 90 offset between the coordinate systems of the VB6 and that of the room inwhich it resides.
370
Figure 7.40: Step 25: O2 plasma etch the PMMA underlayer.
Figure 7.41: This is a top view of a device at this point. The crosses over the 40µm squares show that the alignment procedure worked.
371
Figure 7.42: Step 26: Ion mill define top pillar. The insulating AlOx etches veryslowly in an ion mill, so it is expected that thinning of this layer will be minimal.The goal here is to just uncover the bottom-most Cu layer. It will be helpful toconfirm that the Cu was uncovered and not etched completely by measuring thestep height with AFM and also by visual inspection with optical microscopy.
Figure 7.43: Step 27: Insulate nanopillar with IBD AlOx. Proper angle ofdeposition is not known at this time.
372
Figure 7.44: Step 28: Measure AlOx height. Do profilometry or AFM on theAlOx step edge where the central Cu lead is to find h2 which will be the targetdeposition height for the second AlOx refill.
Figure 7.45: Step 29: Photolith 6, second AlOx refill (1/2). Spin photoresist: aprotective 495k PMMA underlayer/LOR10A liftoff resist/S1813.
373
(a)
(b)
Figure 7.46: Step 29: Photolith 6, second AlOx refill (2/2). (a) Top down viewafter development. This step basically tries to level the plane of the wafer as muchas possible before the second CMP planarization step. (b) The mask used forphotolith 6. Black is Cr and white is glass.
374
Figure 7.47: Step 30: Second AlOx refill. IBD-deposit AlOx to the heightmeasured in step 28 (h2) while rotating the stage. The optimum angle of depositionis not known. The height h2 may be so small that this step may be dropped. Liftoff.
Figure 7.48: Step 31: Second CMP planarization. CMP planarize the wafer aswas done in step 17.
Figure 7.49: Step 32: Photolith 7, open leads (1/2). Spin photoresist P20/S1827.
375
(a)
(b)
Figure 7.50: Step 32: Photolith 7, open leads (2/2). (a) Top down view afterdevelopment. This step opens all bonding pads and the middle leads which arecurrently buried under AlOx. (b) The mask used for photolith 7. Black is Cr andwhite is glass.
376
Figure 7.51: Step 33: HF etch leads. HF etch the AlOx above all of the bondingpads as well as the contact for the middle lead defined by ion milling in step 26.The Cu of the middle lead will corrode slightly in the HF which can be identifiedby a darkening of the exposed Cu.
Figure 7.52: Step 34: Make Au contacts. Without stripping the resist fromFigure 7.51, ion mill etch off the corroded Cu. This should be done at 0 ion beamincidence. Once cleaned, place in an evaporator and deposit Cr 100 A/Au 300 A.The Cr adhesion layer should provide good Ohmic contact between the Au andthe Cu lead contact. Lift off.
377
Figure 7.53: This is a top view of a device at step 34.
Figure 7.54: Step 35: Photolith 8, define middle lead (1/2). Spin photoresist: aprotective 495k PMMA underlayer/LOR10A liftoff resist/S1813.
378
(a)
(b)
Figure 7.55: Step 35: Photolith 8, define middle lead (2/2). (a) Top down viewafter development. This step defines the geometry of the middle lead. (b) Themask used for photolith 8. Black is Cr and white is glass.
379
(a)
(b)
Figure 7.56: Step 36: Deposit Cu middle lead. IBD-deposit Cu middle lead∼100 nm thick. This should follow a brief (∼10 sec) ion mill clean of Au contact.(a) Side view of all the layers up until just after definition of the middle lead. (b)Top down view at the same point.
380
Figure 7.57: Step 37: Photolith 9, protect shorts (1/2). Spin photoresist: aprotective 495k PMMA underlayer/LOR10A liftoff resist/S1813.
381
(a)
(b)
Figure 7.58: Step 37: Photolith 9, protect shorts (2/2). (a) Top down view afterdevelopment. This step defines region of the middle lead that will be covered inSiOx so as to prevent shorts with the top lead which will eventually lie directly ontop of the middle lead. (b) The mask used for photolith 9. Black is Cr and whiteis glass.
382
Figure 7.59: Step 38: Protect shorts. Evaporate or IBD-deposit ∼150 nm SiOxwhile rotating the stage at a finite tilt angle (φ ≈ 25) from incoming SiOx. TheCHA evaporator may be used with the tilting, rotatable stage, one wafer at a time.Lift off.
383
(a)
(b)
Figure 7.60: Step 39: Photolith 10, define top lead. (a) Spin photoresist: aprotective 495k PMMA underlayer/S1827. This step defines the geometry of thetop lead. (b) The mask used for photolith 10. Black is Cr and white is glass.
384
Figure 7.61: Step 40: Clean off Ta cap. Short ion mill at 0 incidence to etch offthe TA cap. There is no need to break vacuum after this step.
385
(a)
(b)
Figure 7.62: Step 41: Deposit Cu top lead. After ion mill cleaning the topcontact, deposit 300 nm of IBD Cu to define the top Cu lead. Lift off. (a) Sideview of the completed device. (b) Top down view of the completed device.
386
7.3 Measured devices from Hitachi Global Storage Tech-
nologies
7.3.1 Fabrication differences
In a collaboration with Jordan Katine at Hitachi, 3-terminal nanopillars were fab-
ricated with a process very similar to the one outlined in the previous section.
Here I briefly highlight two important fabrication differences between the samples
from Hitachi and the proposed process of section 6.2. First, the fine alignment was
done with photolithography marks and so the alignment accuracy is on the order
of 40 to 50 nm (see Figure 7.68).
Second, concerns the order of certain steps. The Hitachi process first isolates
the devices with ion milling followed by AlOx refill before e-beam lithography
definition of the nanopillar. This is deliberately avoided in the Cornell process
because it is feared that, because AlOx ion mills very slowly as compared to most
metals, significant redeposition of the etched metals on the inner walls of the AlOx
walls may cause shorting. A schematic of this is shown in Figures 7.63 through 7.65.
7.3.2 3-terminal devices
A schematic of the 3-terminal devices fabricated by Jordan Katine is shown in
Figure 7.66. The structures were inverted from what is described in the proposed
fabrication process of section 6.2, meaning that the MTJs were made on top of
the spin-valves and the free layer (bottom layer in the MTJ) lay on top of the
middle Cu lead. Two wafers were fabricated, one with all metallic devices, the
other that had an MgO-based MTJ as the top pillar. The substrate is unoxidized
Si. The bottom Ta in the underlayers serves as the adhesion layer. Layerings for
387
(a)
(b)
Figure 7.63: For the Hitachi process, the device isolation is done before theion mill definition of the nanopillar. (a) The multilayer stack with device leadsisolated the e-beam-defined ion mill mask. (b) Ion mill definition of the nanopillar.Because the AlOx insulator ion mills very slowly, the metals will likely redepositon the inner facing walls of the AlOx.
388
(a)
(b)
Figure 7.64: (a) More IBD-AlOx is deposited to insulate the nanopillar itself. (b)After CMP planarization the ion mill mask is removed and the AlOx is thinned.There is a possibility that the AlOx may be thinned too much near the edges ofthe stack, exposing the redeposition spikes coating the inner walls of the initialAlOx.
389
Figure 7.65: Subsequent metal deposition may short to the bottom lead throughthese spikes. This is the reason why, for the proposed Cornell 3-terminal nanopillarfabrication, the nanopillar is defined by ion milling before patterning the leads.
the underlayers are very similar to the Cornell process.
The bottom pillar is identified as a pinned SAF nanopillar, where the Pt50Mn50
provides pinning along a line perpendicular to the bonding pads, as shown in
Figure 7.22. The pinned SAF should give a distinct magnetic response to the top
pillar. The composition of the CoFe in the bottom SAF pillar was not given. A
Ru 40 A/Ta 20 A cap used to be on top of the Cu 40 A above the SAF pillar but
that has been ion milled off. The equivalent capping layer for the Cornell process
is Ta 50 A. The Cu 40 A is likely thinner due to slight over milling. The Cu 150
A middle lead is the bottom layer of the top stack of metals, which have been ion
milled for top pillar definition. Consequently, this lead likely thinner due to slight
over milling.
The free layer is a trilayer: Co84Fe16 10 A/Ni86Fe14 20 A/Co84Fe16 10 A. Despite
the high magnetization Co84Fe16 dusting layers, the free layer is still superparamag-
netic at room temperature because of the significant fraction of the Py (Ni86Fe14).
The top spacer material is either a metallic (Cu 40 A) or tunnelling (MgO 6 A/7 A)
390
layer. Processing details for the MgO layer were not given. The reference Co84Fe16
90 A layer is thick enough to not be excited by spin transfer currents. Similar to
the bottom pillar a Ru 50 A/Ta 25 A cap used to be on top of the Cu 100 A above
the top pillar but has been ion milled off. A bilayer of Cr 50 A/Au 1000 A forms
the top lead. All devices, both bottom and top nanopillars, were circular. The
bottom SAF pillar diameters (50nm < dbottom < 200nm) were bigger than or equal
to the diameters of the top pillars (40nm < dtop < 200nm). SEM images of four
3-terminal devices after HSQ e-beam patterning of the top nanopillar are shown
in Figure 7.68. An SEM image of a completed device is shown in Figure ??.
7.3.3 Experimental results
GMR for these devices was measured between (1) top and middle leads, (2) top
and bottom leads, and (3) middle and bottom leads. (1) measures only the top
nanopillar while (3) measures only the bottom pinned SAF nanopillar. (2) should
measure the two in series. All data were taken by first scanning the control pa-
rameter (in this experiment applied magnetic field H and DC current IDC) to
maximum positive value, then to maximum negative value, then return to zero.
IDC ramp rates are ∼0.1 mA/sec. H was applied parallel to the exchange bias
direction of the PtMn pinning layer (Figure 7.22). All data are at room tempera-
ture.
Spin-valve as the top nanopillar
Both top and bottom pillars are metallic for these devices and so resistance levels
should be comparable for both. Detection of a ∆R for both when measuring top-
to-bottom-lead resistance should also be possible. There is very little uniaxial
391
Cu
lea
d 1
50 Cu
40
Un
der
lay
ers:
[T
a 5
0
/C
u 3
00
]
/T
a 5
0
Si
subst
rate
Au
10
00
A
Cu
10
0A
AA
A
A
Pt
Mn
20
050
50
2
Co
F
e 90
Top
sp
acer
Fre
e la
yer
A84
16
Co
Fe
23
Ru
7
Co
Fe
20
A
A
Cr
50
A
A
A
A
Figure 7.66: Schematic of the Hitachi 3-terminal nanopillar. The bottomPtMn/CoFe/Ru/CoFe pillar is the pinned SAF. “Top spacer” = MgO (6 A/7A) or Cu 40 A. “Free layer” = Co84Fe16 10 A/Ni86Fe14 20 A/Co84Fe16 10 A. “Un-derlayers” are engineered for smoothness [6].
392
100 nm
100 nm 100 nm
20 nm
(a) (b)
(c)(d)
Figure 7.67: SEM images of the 3-terminal nanopillar after e-beam definitionof top pillar HSQ mask. Nominal device sizes are (dbottom/dtop, all in nm) (a)200/75, (b) 200/100, (c) 200/120, and (d) 200/150. The ∼200 nm dark ringcircumscribing each HSQ mask (bright disk) is a local topographical effect of thebottom pillar (each 200 nm in diameter) coming from shadowing of the IBD AlOx.Such shadowing is inevitable, but it looks as if the Hitachi IBD AlOx depositionsequence has been tuned to mitigate shadowing effects. Drift in the quality ofthe fine alignment is most visible in (b) and (d), where the top nanopillars aremisaligned with the center of the bottom pillar by 40 to 50nm.
393
Top Leads Bottom Leads
Middle Leads
Figure 7.68: SEM image of a completed 3-terminal device at a 37 viewing angle.The leads are indicated. The bottom leads are under two insulation layers, butstill appear brighter than the middle leads (which are correspondingly under onlyone insulation layer) because the bottom leads are much thicker as they includethe underlayers. The middle lead is only a thin Cu film.
394
shape anisotropy in these devices due to their circular shape. Figure 7.69 shows
the top-to-middle lead differential resistance (i.e lock-in) measurement for sample,
a 75 nm (top diameter) device. For fields below saturation (|H| ≈ 600 Oe for this
sample), the magnetic layers show hysteretic behavior. Comparing Figure 7.69(a)
and (b) the hysteresis seems to come from a metastable magnetic state since it is
not always repeatable, possibly a varying micromagnetic state.
When measuring the entire device structure between the top and bottom leads,
we find that there is no magnetic signal coming from the SAF nanopillar, as shown
in Figure 7.70, where we have taken the electromagnet out to a maximum of H
≈ 3400 Oe. Although the GMR response of the top spin valve is clear, there is
no obvious GMR from the SAF, which should be broken by H ≈ 2000 Oe [6]. A
∆R on the order of that for the top spin-valve nanopillar is expected (see ref [7]
and Ch. 4 of this thesis). The slight hysteresis when the two layers are saturated
at high H cannot be due to the SAF since the same behavior is present when
measuring top-to-middle lead GMR (top spin-valve only) at high H. There also
lacks an appreciable resistance difference between top-to-middle-lead and top-to-
bottom-lead measurements. The middle-to-bottom-lead measurements show no
magnetic behavior as well as an uncharacteristically small resistance. These data
indicate that there is a short between the middle and bottom leads. The measured
lock-in resistance dV/dI of the pinned SAF bottom pillar for six spin-valve top
pillar samples was measured at less than 1 Ω.
There is no significant shorting through the un-oxidized Si substrate. Resis-
tance between neighboring device bonding pads was measured at ∼23 kΩ at room
temperature. It would seem from these data that there is a short somewhere be-
tween the middle and bottom leads. Such a short is conceivable based upon the
395
order of the ion milling steps used in the Hitachi process and that are illustrated
in Figures 7.63 through 7.65.
MTJ as the top nanopillar
Despite being deposited on a pre-patterned surface, the top MTJ nanopillars
showed reasonable R·A (7 – 13 Ω ·µm2) and ∆R·A (1 – 4 Ω ·µm2) values. DC resis-
tance for the top, MTJ for sample 2 (50 nm diameter) is shown in Figure 7.72(a)
while the differential resistance of the bottom, pinned SAF (100 nm diameter) is
shown in (b). Filled circles (•) are H increasing and open circles () are H de-
creasing. The near perfect linearity of this tunnelling magnetoresistance (TMR)
indicates how well it would work as a magnetic field sensor although it is not as
useful as a spin transfer device. The problem of an immeasurable magnetic re-
sponse from the pinned SAF bottom pillar afflicts these samples as well, as shown
in 7.72(b). Pinned SAF bottom pillar lock-in resistance dV/dI was measured for
over 30 MTJ top pillar devices, spanning all of the device areas. Plotting Wheat-
stone bridge box resistance as a function of pinned SAF bottom pillar nominal
area (i.e. the area engineered into the e-beam exposure pattern) shows a “short-
ing resistance” which tends to be larger for smaller diameter bottom nanopillars,
as shown in Figure 7.71. The data have been offset by the intrinsic Wheatstone
bridge resistance of 0.6 Ω (i.e. when the bridge resistance is set to zero). The
data suggest that some devices are shorts (dV/dI = 0.6 Ω), but also that some
devices whose resistance scales with smaller area and may consequently not be
shorted. No bottom pillar device measured showed any magnetoresistive response,
even those with dV/dI ≈ 4 Ω. Although the SAF-pinning should be broken by
field values achievable with the available electromagnets, the possibility that pin-
396
(a)
(b)
top pillar
top pillar
Figure 7.69: Top-to-middle-lead differential resistance as a function of H forsample 1, a bottom/top pillar diameter of 150nm/75nm. Filled circles (•) are Hincreasing and open circles () are H decreasing, as indicated. (a) Full antiparallel(high resistance at H = 0) and parallel (low resistance at |H| > 600 Oe) alignmentof the two magnetic layers is clear. However, the lack of a uniaxial shape anisotropymakes interpretation of the magnetic states below saturation difficult. (b) Sameas (a) except that here the resistance is non-hysteretic at positive H.
397
(a)
(b)
top + bottom
pillar
bottom pillar
Figure 7.70: (a) Top-to-bottom-lead and (b) middle-to-bottom-lead GMR as afunction of H for sample 1, a bottom/top pillar diameter of 150nm/75nm. Filledcircles (•) are H increasing and open circles () are H decreasing. The resistancebetween top and bottom leads (a) shows no additional ∆R from the bottom pillarand the combined resistances are very similar to that of just the top pillar. Smallhysteresis is present when taken out to high H, but this is also seen in both top-to-bottom-lead GMR measurements so it is not due to the bottom SAF pillar.The middle-to-bottom-lead GMR (b) shows no magnetic behavior as well as auncharacteristically low resistance. These data indicate a possible short betweenthe middle and bottom leads.
398
ning strength exceeds the maximum applied fields cannot be ruled out because the
precise pinning strength is unknown.
Nevertheless, the 3-terminal experiment was performed, where a DC current is
passed through the middle and bottom leads while resistance is monitored across
the top and bottom leads. Sample 3, a 120 nm diameter top MTJ pillar (150 nm
pinned SAF bottom pillar) with TMR shown in Figure 7.73(a), was measured in
this way. The V- for the DC resistance measurement was connected to the middle
lead along with the I- from the DC current source, defining the middle lead as
the device ground. In turned out, however, that because of a finite lead resistance
∼50Ω, there would inevitably be a small voltage buildup in the middle lead because
the DC current sees some resistance to ground. For a 1 mA DC current separated
from ground by a 50Ω lead resistance, a 50 mV voltage build up in the middle
lead near the nanopillar is expected. However, since we are measured the DC
resistance of the MTJ with a sampling current of 10 µA, the voltage across the
barrier (R ≈ 500 – 5000Ω) is ∼5 – 50 mV. This means that the voltage swing from
just the DC current itself is at least as large as the signal from the MTJ. This was
found to be true, and scans taken in this way gave erroneous resistance values (see
Figure 7.73(b)).
In order to correct for this voltage offset, we measured DC resistance using the
Keithley 2400 with its “OFFSET COMPENSATION MODE” enabled. In this
mode, the Keithley measures the resistance as R = (V2 - V1)/(I2 - I1) where V2
is the measured voltage with the sampling current on, V1 is the measured voltage
with the sampling current off, I2 = 10 µA (sampling current on), and I1 = 0
(sampling current off). This subtracts the voltage offset background coming from
the DC current source. Such offset compensated scans are shown in Figure 7.74
399
50 100 150 2000.00.51.01.52.02.53.03.54.04.5
B
ridge
Res
ista
nce
[]
Nominal Diameter [nm]
Figure 7.71: Pinned SAF bottom pillar resistance (i.e. Wheatstone bridge boxresistance) versus the nominal area of the bottom pillar (i.e. the engineered e-beam exposure pattern). Although some devices indicate a shorting (resistance ≈0.6 Ω), others show a resistance that scales with decreasing area. These devicesindicate that the lack of magnetic response may not be a shorting problem butinstead could be due to SAF pinning strength that exceeds maximum achievablefields from the electromagnet.
400
at various field values. At H = -200 Oe (a) there is a smooth change in RDC of
approximately 5Ω as the current is swept between ±12 mA. At H = 0 Oe (b),
there is an abrupt and irreversible resistance change ∆R ≈ 3.5Ω at IDC ≈ +12
mA. However, there is no corresponding resistance change in the TMR at H =
0 (Figure 7.73(a)) and the fact that it is irreversible is evidence for this ∆R not
coming from a spin torque. A similar irreversible jump in resistance is seen in
(e) (H = 150 Oe). For all other fields in these scans (c), (d), and (f), there is a
consistent arching background in the data, although the reason for this is unclear.
It is unlikely to be an artifact from the setup since it seems to be field dependent.
For example, the slope of the curvature is negative slope at H = -200 Oe (a), but
is mostly positive for H = 200 Oe (b).
7.3.4 Summary
Two wafers with 3-terminal nanopillar devices were fabricated by Jordan Katine
at Hitachi with a structure shown schematically in Figure 7.66. For both wafers,
the bottom nanopillar was a pinned SAF while the top nanopillar was either a
spin-valve or an MTJ. All devices were circular, and so the magnetoresistance for
both top pillars consistently showed a nearly linear response to an applied field H.
Neither of these wafers showed any magnetic behavior in the bottom nanopillar,
a pinned SAF pillar. The reasons for this are unclear, but because a lack of any
measurable GMR as well as an uncommonly low resistance of the junction, it is
suspected that there is a short between the middle and bottom leads of the device.
This short prevents magnetoresistance measurements of the bottom pinned
SAF pillar, but it would also reduce the amount of spin polarized current striking
the free layer of the top nanopillar. Based on the Hitachi fabrication process, such
401
(a)
(b)
top pillar
bottom pillar
Figure 7.72: (a) Top-to-middle-lead DC resistance of the MTJ for sample 2, 50nm diameter top MTJ pillar (100 nm pinned SAF bottom pillar) circular device(R·A = 12.0 Ω · µm2, ∆R·A = 1.92 Ω · µm2). (b) The middle-to-bottom-leaddifferential resistance for the same device (100 nm diameter). Filled circles (•) areH increasing and open circles () are H decreasing. Same as for the all-metallicsamples, these 3-terminal devices show no magnetic behavior in their pinned SAFbottom nanopillar.
402
(a)
(b)
H = -200 Oe
Figure 7.73: (a) DC Resistance (RDC) of sample 3, a 120 nm diameter top MTJpillar (150 nm pinned SAF bottom pillar), which is sampled by a 10 µA currentwhile passing a current between the bottom to middle connection IDC (bottom– middle). Thin line is for H increasing, thick line is for H decreasing. It showssome stable hysteresis at low H. The cross × and open circles represent fieldvalues where a DC current, passed between bottom and middle leads, was appliedin an attempt to switch the free layer of the MTJ (see Figure 7.74). (b) A firstattempt to switch the free layer of the MTJ while sending DC current betweenthe bottom and middle leads at H = -200 Oe (the × in (a)). The linear responsecomes from a voltage buildup in the middle lead (∆V = IDC (bottom – middle) ×Rmiddle lead) because of its finite resistance (Rmiddle lead ≈ 50Ω). The interceptof the line at IDC (bottom – middle) = 0 is the actual resistance of the MTJ(∼600Ω) at H = -200 Oe from (a).
403
H = -200 Oe H = 0 Oe
H = 50 Oe H = 100 Oe
H = 150 Oe H = 200 Oe
(a) (b)
(c) (d)
(e) (f)
Figure 7.74: DC Resistance (RDC) of sample 3 sampled by a 10 µA current,the same MTJ as shown in Figure 7.73(a) plotted versus DC current IDC (bot-tom – middle) passed between the bottom and middle leads, with the “OFFSETCOMPENSATION MODE” enabled on the Keithley 2400. Open triangles are forpositive-swept currents while filled triangles are for negative-swept currents. Thefield values are indicated in Figure 7.73(a) as: (a) ×, H = -200 Oe; (b) , H = 0Oe; (c) , H = 50 Oe; (d) , H = 100 Oe; (e) , H = 150 Oe; (f) , H = 200 Oe.
404
a shorting problem is possible, as shown in Figures 7.63 to 7.65. The measurements
of MTJ resistance plotted versus current passed between the middle and bottom
leads are shown in Figure 7.74 and there is no clear experimental evidence that
such a current excited any magnetization dynamics in the free layer of the top
MTJ pillar, which is consistent with other evidence that the bottom pinned SAF
pillar is being shorted. These results suggest a fabrication problem, rather than a
physically unviable device, is responsible for the lack of interesting results. If the
problem in the fabrication can be solved, 3-terminal nanopillars may be realized.
405
References for Chapter 7
[1] Slaughter J.M., Dave R.W., DeHerrera M., Durlam M., Engel B.N., JaneskyJ., Rizzo N.D., & Tehrani S., Fundamentals of mram technology, Journal ofSuperconductivity: Incorporating Novel Magnetism 15, 19–25 (2002).
[2] Slonczewski J.C., Exchange-driven magnetic excitation and integrated magne-toelectronics, unpublished (1999).
[3] Fuchs G.D., Ph.D. thesis, Cornell University (in preparation).
[4] Krivorotov I.N., Emley N.C., Garcia A.G.F., Sankey J.C., Kiselev S.I., RalphD.C., & Buhrman R.A., Temperature dependence of spin-transfer-inducedswitching of nanomagnets, Phys. Rev. Lett. 93, 16603 (2004).
[5] Ozatay O., Ph.D. thesis, Cornell University (in preparation).
[6] Lacour D., Katine J.A., Smith N., Carey M.J., & Childress J.R., Thermaleffects on the magnetic-field dependence of spin-transfer-induced magnetizationreversal, Appl. Phys. Lett. 85, 4681–4683 (2004).
[7] Emley N., Albert F.J., Ryan E.M., Krivorotov I.N., Ralph D.C., BuhrmanR.A., Daughton J.M., & Jander A., Reduction of spin transfer by syntheticantiferromagnets, Appl. Phys. Lett. 84, 4257–4259 (2004).
CHAPTER 8
CONCLUSION
In this dissertation, I have discussed in great detail the current nanofabrication pro-
cess employed to make magnetic multilayer current-perpendicular-to-plane (CPP)
spin-valve nanopillar devices for the study of current-induced excitations of a free
layer magnetic moment. The current becomes spin-polarized by the natural spin-
filtering effects of the ferromagnetic layers within the nanopillar [1, 2]. When
impingent upon the free layer moment, the spin polarization can excite a dynamic
response from the magnetic moment through an exchange of spin angular mo-
mentum, known as a spin torque [3]. The potential for technological applications
based on these spin torque-induced excitations has inspired research in DC current-
induced microwave oscillations [4–7] and hysteretic switching [8–11] of the free layer
moment. I have presented my own outlook for possible ways in which spin torque-
actuated devices may be incorporated into real future technologies. The outlook
is given from a business perspective, rather scientific one, because financial and
business-incentive arguments are all-too-often absent from discussions concerning
the potential for and efficacy of devices for future technological application that
are still in the realm of academic exploration.
Spin torque-induced reversal of the free layer moment depends upon the strength
of spin torque per unit current, the magnetization itself, and the magnitude of the
magnetic energy dissipation, or damping. I present two studies of spin torque-
induced reversal in magnetic multilayer nanopillars which demonstrate that the
strength of the spin torque and damping can have be adversely affected by the
addition of magnetic material in proximity to the free layer. In the first study,
summarized in Chapter 4, a third, oppositely aligned magnetic layer (Cobottom)
406
407
has been incorporated into the CPP spin-valve nanopillar giving a layer composi-
tion [Cobottom/Ru/Cofixed]/Cu/Cofree, where square brackets indicate what is
known as a synthetic antiferromagnetic (SAF) trilayer. In the SAF, the Cobottom
and Cofixed layers are aligned antiparallel (AP) by strong indirect exchange cou-
pling through the Ru spacer [12]. All three magnetic layers are patterned, so this
AP alignment does reduce undesirable dipole fields on the Cofree layer. However,
addition of the Cobottom/Ru layers reduces the spin polarization of the electron
current passing through the nanopillar, due to the opposing spin-filtering effects
from the two Co layers of the SAF. This leads to a weakened spin torque per unit
current incident on the Cofree layer.
In the second study, I perform time-resolved measurements of spin torque-
induced reversal of a free magnetic layer in Py/Cu/Py (Py = Ni81Fe19) elliptical
nanopillars at temperatures T = 4.2 K to 160 K. We compare our experimental
results with detailed Landau-Lifshitz-Gilbert simulations of the free layer switching
in the macrospin approximation and find good agreement between simulation and
measurement. This both confirms the applicability of the macrospin approximation
in the spin torque-driven regime and facilitates the quantitative determination of T -
dependent spin torque and magnetic damping parameters, which are determined as
functions of T . At higher T we find that the strength of the spin torque exerted per
unit current is in reasonable numerical accord with recent model calculations, and
that the damping parameter for the nanomagnet excitations is both anomalously
high, as suggested by previous pulsed current measurements [11], and T -dependent.
The strong T variation of the damping, in conjunction with anomalous behavior
of the nanomagnet field-induced switching observed in some devices, points to the
presence of an adventitious antiferromagnetic oxide layer around the perimeter of
408
the nanomagnet that forms naturally during sample fabrication. The presence of
this antiferromagnetic Py oxide layer can have a major effect on the nanomagnet
dynamics. Controlling this layer will be important in optimizing spin torque-
actuated behavior.
409
References for Chapter 8
[1] Upadhyay S., Louie R.N., & Buhrman R.A., Spin filtering by ultrathin ferro-magnetic films, Appl. Phys. Lett. 74, 3881–3883 (1999).
[2] Stiles M.D., Spin-dependent interfacial transmission and reflection in magneticmultilayers (invited), J. Appl. Phys. 79, 5805–5810 (1996).
[3] Slonczewski J.C., Current-driven excitation of magnetic multilayers, J. Magn.Magn. Mater. 159, L1–L7 (1996).
[4] Kiselev S.I., Sankey J.C., Krivorotov I.N., Emley N.C., Schoelkopf R.J.,Buhrman R.A., & Ralph D.C., Microwave oscillations of a nanomagnet drivenby a spin-polarized current, Nature (London) 425, 380–383 (2003).
[5] Rippard W.H., Pufall M.R., Kaka S., Russek S.E., & Silva T.J., Direct-currentinduced dynamics in Co90Fe10/Ni80Fe20 point contacts, Phys. Rev. Lett. 92,027201 (2004).
[6] Kaka S., Pufall M.R., Rippard W.H., Silva T.J., Russek S.E., & Katine J.A.,Mutual phase-locking of microwave spin torque nano-oscillators, Nature (Lon-don) 437, 389–392 (2005).
[7] Mancoff F.B., Rizzo N.D., Engel B.N., & Tehrani S., Phase-locking in double-point-contact spin-transfer devices, Nature (London) 437, 393–395 (2005).
[8] Katine J.A., Albert F.J., Buhrman R.A., Myers E.B., & Ralph D.C., Current-driven magnetization reversal and spin-wave excitations in Co/Cu/Co pillars,Phys. Rev. Lett. 84, 3149–3152 (2000).
[9] Urazhdin S., Birge N.O., Jr. W.P.P., & Bass J., Current-driven magneticexcitations in permalloy-based multilayer nanopillars .
[10] Grollier J., Cros V., Hamzic A., George J.M., Jaffres H., Fert A., FainiG., Youssef J.B., & Legall H., Spin-polarized current induced switching inCo/Cu/Co nanopillars, Appl. Phys. Lett. 78, 3663–3665 (2001).
[11] Braganca P.M., Krivorotov I.N., Ozatay O., Garcia A.G.F., Emley N.C.,Sankey J.C., Ralph D.C., & Buhrman R.A., Reducing the critical currentfor short-pulse spin-transfer switching of nanomagnets, Appl. Phys. Lett. 87,112507 (2005).
[12] Parkin S.S.P. & Mauri D., Spin engineering: Direct determination of theRuderman-Kittel-Kasuya-Yosida far-field range function in ruthenium, Phys.Rev. B 44, 7131–7134 (1991).