Measurement Based FET Modeling using Artifical Neural ... works for both common-source FET and...

Welcome

1 February 7, 2012

Copyright © Agilent Technologies 2012

David Root

Jianjun Xu

Modeling and Measurement Technology Program (MMTP)

Measurement Research Laboratories

Agilent Technologies, Inc.

Agilent Restricted

Measurement-Based FET Modeling

using Artificial Neural Networks (ANNs)

Jianjun Xu and David E. Root

Modeling and Measurement Technology Program (MMTP)

Measurement Research Laboratories

Agilent Technologies, Inc.


Agilent Restricted

Presentation Outline

• Introduction • NeuroFET Modeling Flow Details • Examples and Model Validation

• Summary


Agilent Restricted

Introduction

“All models are wrong, but some are useful.”

- statistician George Box

“All models are approximations.

Some models are useful.”

- attributed to Mike Golio and others


Agilent Restricted

Purpose of Webinar

Introduce Agilent’s first measurement-based modeling solution

based on artificial neural networks (ANN) – NeuroFET

Three interoperable components:

(1) Data

Acquisition

(2) Advanced

ANN Training

(3) Compiled

Nonlinear Model

Agilent NeuroFET

•Agilent developed each component; designed them to work together

•Deployed for Agilent proprietary III-V MMIC technology at HFTC

•Commercialized in collaboration with Agilent EEsof EDA Division


Agilent Restricted

NeuroFET components

VNA

DC

Data Acquisition:

Agilent IC-CAP

(1) Automated, adaptive characterization system for FETs

Supports common gate and common source layouts

Minimizes impact of device degradation during characterization

Benefits:


Agilent Restricted

NeuroFET components

Agilent IC-CAP

(2) Unique and powerful Agilent ANN Training Technology

Model

Generation

Benefits:

•Advanced ANN training creates accurate & general model functions

•Same flow can fit very different looking device characteristics


Agilent Restricted

NeuroFET components

5th

NeuroFET model: Line

Measured data : x

Fund

4th

3rd

2nd

Pin (dBm)

Po

ut

(dB

m)

Nonlinear

Simulation

And Design

Agilent ADS

(3) Built-in (compiled) nonlinear transistor model in ADS

Benefits:

•Fast & accurate nonlinear simulation with robust convergence

•Model usable in all bias conditions (even for switch / mixers)

•Model works for HEMT, MESFET and other types of FETs 0DSV


Agilent Restricted

Conventional Transistor Modeling Flow

Device

Characterization DC + S-parameter

CGD vs bias

Drawbacks:

• Expert intensive (Ph.D.)

• Time consuming: years to develop, days to extract

• Technology dependent; hard to maintain

• Optimization not always well-posed; constrained

• May never get good results!

Equation formulation

Physical or Empirical

Parameter Extraction

Optimization

3 32 2

2

3

1

tan

)

h

2 2(

3

DD

DS GD

n

D n GS DS

D

n

GS

W qN aI

L

V V Va

I V

N

V

q

A

0 0

( ( )

1

2

)

GSG gs bi gd GD

b

G

i

G D

S GD

VQ C

Q WL q

V C

N

V

V

V

V

Good

Fit?

modify

values

Initialize

values

Simulate

Compare

to Data

Done

no

yes

VNA

DC


Agilent Restricted

Alternative: Measurement-based models

Directly construct nonlinear model functions from data

(DC + S-parameters)

– Table-based models (e.g. HP/Agilent FET (Root) Model)

– Artificial Neural Network (ANN) models (e.g. NeuroFET)

Models can be both general and accurate

Conventional I-V, Q-V model development and

parameter extraction replaced by ANN training

“The Device knows best!”


Agilent Restricted

NeuroFET: Measurement-Based FET Modeling using ANNs

Device


CGD vs bias

Advantages:

• Model is general and accurate

• Easy to extract (train)

• Technology independent; easy to maintain

• The model computation is fast

• Infinitely differentiable and smooth


using ANN

VNA

DC

)( w,V,VfI dsgsANNd

w

dI

ANNf

dsgs V V

ANN Training

Adjust w

dsgs V V

w

dI

ANNf

Measured

DC


Agilent Restricted

NeuroFET: Measurement-Based FET Modeling using ANNs

Device


CGD vs bias

Advantages:

• Model is general and accurate

• Easy to extract (train)

• Technology independent; easy to maintain

• The model computation is fast

• Infinitely differentiable and smooth


using ANN

VNA

DC

)( w,V,VfI dsgsANNd

w

dI

ANNf

dsgs V V

ANN Training

Adjust w

dsgs V V

w

dI

ANNf

Measured

DC

NeuroFET model (___ )

Measured data ( o )


Agilent Restricted

IC-CAP: NeuroFET Modeling Flow

Data Acquisition

Parasitic Extraction/De-embedding

Model Generation (ANN Training)

Verification

VNA

DC


Agilent Restricted

NeuroFET: Data Acquisition

For measurement-based model,

(1) it is important to get data everywhere.

(2) the acquisition routine should be intelligent and flexible to preserve

the device for integrity.

Why is a good Data Acquisition Routine important?

Vds

Vgs

Power dissipation compliance

Forward current compliance

Breakdown

compliance

Vds

Id

(1)

(2)


Agilent Restricted

G

D

S

GND

It works for both common-source FET and common-gate FET

G

D S

GND


Common-Source FET

(for Amplifier Application)

Common-Gate FET

(for Mixer/Switch Applications)

G D

S G

D S


Agilent Restricted


Id

Ig

Vds

Vgs


Agilent Restricted

NeuroFET Model

Lg Ld Rg Rd

Ls

Rs Cgs Cds

Cgd

Intrinsic FET G D

S


Agilent Restricted

Lg Ld Rg Rd

Ls

Rs Cgs Cds

Cgd

NeuroFET: Parasitic Extraction & De-embedding

G D

S

DC and SP of Intrinsic FET

Intrinsic FET


Agilent Restricted

Vds_EXT

Id

Vds_INT

Id

Before and After De-embedding

Vds_EXT Vds_INT

Vgs_EXT Vgs_INT

Extrinsic Intrinsic


Agilent Restricted

G D

S

Intrinsic FET

NeuroFET: Intrinsic Model

NeuroFET (Intrinsic) is a non-quasi-static model that accounts for

(1) Static I-V characteristics

(2) Terminal charges

(3) Frequency dispersion


Agilent Restricted

Frequency dispersion of small-signal characteristics

Vgs (V)

Vgs (V)

gm gds Vds=6V Vds=6V


Agilent Restricted

NeuroFET: Intrinsic Model Formulation

2

dsgsd

2

dsgsddsgs

ac

d

dsgs

dc

dd

ddt

tV,tVQdτ

dt

tV,tVdQ

dt

tV,tVdIτtV,tVI

dt

tdIτtI

))()(())()(())()(())()((

)()(

dt

tV,tVdQtV,tVItI

dsgsg

dsgs

dc

gg

))()(())()(()(

dc

gIgQ

+

_ 1

dc

dI ac

dI

G D

S

gIdI

1g

dQ


Agilent Restricted


2

dsgsd

2

dsgsddsgs

ac

d

dsgs

dc

dd

ddt

tV,tVQdτ

dt

tV,tVdQ

dt

tV,tVdIτtV,tVI

dt

tdIτtI

))()(())()(())()(())()((

)()(

dt

tV,tVdQtV,tVItI

dsgsg

dsgs

dc

gg

))()(())()(()(

dc

gIgQ

+

_ 1

dc

dI ac

dI

G D

S

gIdI

1g

dQ

dc

dII

Low Frequency


Agilent Restricted


2

dsgsd

2

dsgsddsgs

ac

d

dsgs

dc

dd

ddt

tV,tVQdτ

dt

tV,tVdQ

dt

tV,tVdIτtV,tVI

dt

tdIτtI

))()(())()(())()(())()((

)()(

dt

tV,tVdQtV,tVItI

dsgsg

dsgs

dc

gg

))()(())()(()(

dc

gIgQ

+

_ 1

dc

dI ac

dI

G D

S

gIdI

1g

dQ

ac

dII

High Frequency


Agilent Restricted


- DC

)( dsgs

dc

dd V,VII

)( dsgs

dc

gg V,VII

dc

gI

G D

S

dc

dI

gIdI


Agilent Restricted


- RF

dt

tV,tVdQtV,tVItI

dsgsd

dsgs

ac

dd

))()(())()(()(

dt

tV,tVdQtV,tVItI

dsgsg

dsgs

dc

gg

))()(())()(()(

dc

gI

G D

S

gIdI

ac

dIgQdQ


Agilent Restricted


- RF (small signal)

dc dc

g g

gs gs ds ds 11 12

ac ac21 22d d

gs gs ds ds

I I

V V V V

I I

V V V V

g g

d d

Q Qj j

Y Y

Y YQ Qj j

dc

gI

G D

S

gIdI

ac

dIgQdQ


Agilent Restricted


- RF (small signal)

)(

)(

12

11

YImag

V

Q

YImag

V

Q

ds

g

gs

g

)(

)(

22

21

YImag

V

Q

YImag

V

Q

ds

d

gs

d

)(

)(

22

21

YRealV

I

YRealV

I

ds

ac

d

gs

ac

d

),( dsgsg VVQ ),( dsgsd VVQ ),( dsgs

ac

d VVI

dsgs V ,Vdsgs V ,V

dsgs V ,V

dc dc

g g

gs gs ds ds 11 12

ac ac21 22d d

gs gs ds ds

I I

V V V V

I I

V V V V

g g

d d

Q Qj j

Y Y

Y YQ Qj j


Agilent Restricted

)()( ww ,V,Vf Q ,V,Vf Q dsgsdQ

ANNddsgs

gQ

ANNg

)()()( www ,V,Vf I ,V,Vf I ,V,Vf I dsgs

acd

I

ANN

ac

ddsgs

dcd

I

ANN

dc

ddsgs

dcgI

ANN

dc

g

Vgs Vds

w

dc

gI

dcgI

ANNf

Vgs Vds

w

dc

dI

dcd

I

ANNf

Vgs Vds

w

ac

dI

acd

I

ANNf

Vgs Vds

w

gQ

gQ

ANNf

Vgs Vds

w

dQ

dQ

ANNf

Constitutive Relations are Modeled by ANNs


Agilent Restricted

Z1 Z2 Z3 Z4

y1 y2

x1 x2 x3

Vjk

Wki

S Wki

Inputs

xi

Outputs

yj = S Vjk Zk k

Artificial Neural Networks (ANNs)

Parameters w = [Wki, Vjk]

• Universal Approx. Thm: Can fit any nonlinear function of many variables

• The model computation is very fast.

• Infinitely differentiable.

• Can be trained on non-gridded data in any number of dimensions.

y1 = f1(x1, x2, x3)

Hidden Neuron Output

Zk = tanh( )

y2 = f2(x1, x2, x3)

Hidden

Neuron


Agilent Restricted

Model Training ( )

- Standard Neural Network Training

Adjust w

E

dc

gI dc

dI

)( w,V,Vf I dsgs

I

ANN

dc

d

dcd

dsgs V V

w

dc

dI

dcdI

ANNf

Id

Measured Data


Agilent Restricted

gQModel Training ( ) - Adjoint Neural Network Training

)( w,V,Vf Q dsgs

gQ

ANNg

Qg

Adjust w

dsgs V ,V

)(

)(

12

11

YImag

V

Q

YImag

V

Q

ds

g

gs

g

Qg

gQ

ANNf

w

Vgs Vds


Agilent Restricted

Qg

gQ

ANNf

Vgs Vds

w )( w,V,Vf Q dsgs

gQ

ANNg

dsgs V ,V

)(

)(

12

11

YImag

V

Q

YImag

V

Q

ds

g

gs

g


)(Imag 11Y

)(Imag 12Y

1

Adjoint Neural

Network

w

Adjust w

gQE

Jianjun Xu, M.C.E. Yagoub, Runtao Ding

and Q.J. Zhang,

“Exact adjoint sensitivity analysis for neural

based microwave modeling and design,”

IEEE Transactions on Microwave Theory and

Techniques, vol. 51, pp.226-237, 2003.

gs

g

V

Q

ds

g

V

Q

Original Neural

Network


Agilent Restricted

Qg

gQ

ANNf

Vgs Vds

w )( w,V,Vf Q dsgs

gQ

ANNg

dsgs V ,V

)(

)(

12

11

YImag

V

Q

YImag

V

Q

ds

g

gs

g



Agilent Restricted

Vgs Vds

dc

gI

Vgs Vds

dc

dI

Vgs Vds

ac

dI

Vgs Vds

gQ

Vgs Vds

dQ

ANN ANN ANN ANN ANN

2

dsgsd

2

dsgsddsgs

ac

d

dsgs

dc

dd

ddt

tV,tVQdτ

dt

tV,tVdQ

dt

tV,tVdIτtV,tVI

dt

tdIτtI

))()(())()(())()(())()((

)()(

dt

tV,tVdQtV,tVItI

dsgsg

dsgs

dc

gg

))()(())()(()(



Agilent Restricted

VDS (V)

IDS (mA)

VDS (V)

IDS (mA)

Advanced Model Extrapolation Routine

- for robust convergence

Without Extrapolation With Extrapolation

Measured Data (training data)


Agilent Restricted

NeuroFET Model is compiled into ADS

The model, together with extrapolation routine,

is compiled into ADS.

• Same use model as any other transistor model.

• Model has simple scaling rules with gate width and number of fingers.


Agilent Restricted

Example 1

A 0.25um GaAs pHEMT device (Width=150um) was extracted:

• DC IV

• Qg, Qd, and

• S-parameters versus bias and frequency

• One-tone Harmonic Distortion

• Two-tone Intermodulation

ac

dI


Agilent Restricted

NeuroFET: Training dc

dI


Agilent Restricted

Model Validations

- DC


Measured device test data ( o )


Agilent Restricted

NeuroFET: Training gQ


Agilent Restricted

0 2 4 6 8 10 12 0

0.05

0.1

0.15

0 2 4 6 8 10 12

0.1

0.15

0.2

0.25

Im(Y11)/ and Qg/Vgs x 10-12

Vgs

-Im(Y12)/ and -Qg/Vds x 10-12

Vds (V)

Vgs

(F)

(F)

Model Validations -Charge Qg

NeuroFET model (__)

Measured device data ( o )

Qg (pC)

-2 0 2 4 6 8 10 12-0.8

-0.7

-0.6

-0.5

-0.4

-0.3

-0.2

-0.1

0

0.1

0.2

Vds (V)

NeuroFET model (__)



Agilent Restricted

-2 0 2 4 6 8 10 12-0.1

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

Qd (pC)

Vds (V)

Model Validations -Charge Qd

-2 0 2 4 6 8 10 12 0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

-2 0 2 4 6 8 10 12 -0.16

-0.14

-0.12

-0.1

-0.08

-0.06

-0.04

-0.02

0 Im(Y21)/ and Qd/Vgs x 10-12

Im(Y22)/ and Qd/Vds x 10-12

Vds (V)

(F)

(F)

NeuroFET model (__)


NeuroFET model (__)



Agilent Restricted

Model Validations -

ac

dI

ac

dIdc

dI

From DC

measurements

From S-parameter

measurements


Agilent Restricted

Vgs (V) Vgs (V)

gm gds Vds=5V Vds=5V

Frequency dispersion of small-signal characteristics

DC gm

NeuroFET (___ )

Measured data ( o )

RF gm

DC gds

RF gds


Agilent Restricted

Frequency : 1 GHz to 50 GHz

Model Validations - S parameters

NeuroFET model (___)

Measured device test data ( x ) Vds (V)

I d (

mA

)

1 2 3 4 5 6 7 8 9 10 11 0 12

10

20

30

40

50

60

70

80

0

90

Bias

Vgs = -0.5V

Vds = 5.0V


Agilent Restricted

Vds (V)

I d (

mA

)

1 2 3 4 5 6 7 8 9 10 11 0 12

10

20

30

40

50

60

70

80

0

90

Bias

Vgs = -1.5V

Vds = 5.0V



Measured device test data ( x )

Frequency : 1 GHz to 50 GHz


Agilent Restricted

Model Validations - One-tone Harmonic Distortion

Input Power (dBm)

Vds (V)

I d (

mA

)

1 2 3 4 5 6 7 8 9 10 11 0 12

10

20

30

40

50

60

70

80

0

90

Bias

Vgs = -0.8V

Vds = 3.0V

NeuroFET vs Measured

-200

-175

-150

-125

-100

-75

-50

-25

0

25

-50 -40 -30 -20 -10 0 10

Ou

tpu

t P

ow

er

(dB

m)

Input Power (dBm)

Fund

2nd

3rd

4th

5th

Fund

2nd

3rd

4th

5th

Physically

correct behavior 10 dBm

50 dBm


Agilent Restricted

Physically

correct behavior

NeuroFET vs Measured

-200

-175

-150

-125

-100

-75

-50

-25

0

25

-50 -40 -30 -20 -10 0 10

Input Power (dBm)

Ou

tpu

t P

ow

er

(dB

m)

Fund

IM3

IM5

IM7

Fund

IM3

IM5

IM7

Model Validations - Two-tone Intermodulation

Vds (V)

I d (

mA

)

1 2 3 4 5 6 7 8 9 10 11 0 12

10

20

30

40

50

60

70

80

0

90

Bias

Vgs = -0.8V

Vds = 3.0V

10 dBm

50 dBm

10 dBm

70 dBm

TOI


Agilent Restricted

A 0.25um GaN HEMT device (Width=400um) was extracted:

• DC IV

• S-parameters versus bias and frequency

• One-tone Harmonic Distortion

• Two-tone Intermodulation

Example 2


Agilent Restricted

Model Validations

- DC


Measured device test data ( o )


Agilent Restricted




Bias

Vgs = -3.0V

Vds = 12.0V

Frequency : 0.5 GHz to 50 GHz


Agilent Restricted





Bias

Vgs = -1.0V

Vds = 6.0V


Agilent Restricted

S22 is slightly off when pinched off





Bias

Vgs = -4.5V

Vds = 12.0V


Agilent Restricted

Bias

Vg = -2.5V

Vd = 12.0V

5th


Measured data : x

Fund

4th 3rd

2nd

Pin (dBm)

(dBm)

Model Validations - One-tone Harmonic Distortion (Freq=2GHz)


Agilent Restricted

Bias

Vg = -2.5V

Vd = 12.0V


Measured data : x

IM7 IM5

IM3

Fund

Pin (dBm)

(dBm)

Model Validations - Two-tone Intermodulation (Freq1=2GHz, Freq2=2.005GHz)


Agilent Restricted

Summary: NeuroFET Easy to use, fast to simulate, accurate

• Flexible, automated, data acquisition system takes data where needed

– Minimizes impact of device degradation on resulting model

• Advanced Agilent ANN-training creates accurate & general model functions

• Compiled ADS model works for HEMT, MESFET and other types of FETs

– Robust DC and RF convergence, compared to table-based models; fast to simulate

– Improved distortion simulation compared to table-based models

– Better power-added efficiency (PAE) and S-parameters versus bias over the entire range of

device operation, compared to many compact models for wide range of technologies

– Model can be used in all bias conditions (including for mixers and switches)

• Instrument control and powerful ANN training available in

Agilent’s ICCAP Modeling Software

• FET simulation model available in Agilent's Advanced Design System

(ADS), the industry standard RF simulator

0DSV


Agilent Restricted

• J. Xu, D. Gunyan, M. Iwamoto, A. Cognata, and D. E. Root, “Measurement-Based Non-Quasi-Static Large-Signal FET

Model Using Artificial Neural Networks,” 2006 IEEE MTT-S Int. Microwave Symp. Dig., San Francisco, CA, USA, June

2006.

• J. Xu, D. Gunyan, M. Iwamoto, J. Horn, A. Cognata, and D. E. Root, “Drain-Source Symmetric Artificial Neural Network-

Based FET Model with Robust Extrapolation Beyond Training Data,” 2007 IEEE MTT-S Int. Microwave Symp. Dig.,

Honolulu, HI, USA, June 2007.

• J. Xu, J. Horn, M. Iwamoto, and D. E. Root, “Large-signal FET model with multiple time scale dynamics from nonlinear

vector network analyzer data,” IEEE MTT-S International Microwave Symposium Digest, May, 2010.

• J. Xu, M.C.E. Yagoub, R. Ding, and Q.J. Zhang, “Exact adjoint sensitivity analysis for neural based microwave modeling

and design,” IEEE Transactions on Microwave Theory and Techniques, vol. 51, pp.226-237, 2003.

• D. E. Root, J. Xu, J. Horn, and M. Iwamoto, “The Large-Signal Model: theoretical foundations, practical considerations,

and recent trends,” in Nonlinear Transistor Model Parameter Extraction Techniques, Cambridge University Press,

2011, eds. Rudolph, Fager, & Root

• D.J.McGinty, D.E.Root, and J.Perdomo, “A Production FET Modeling and Library Generation System,” in IEEE GaAs

MANTECH Conference Technical Digest, San Francisco, CA, July, 1997 pp. 145-148.

• D.E.Root, D.McGinty, B.Hughes, “Statistical Circuit Simulation with Measurement-Based Active Device models:

Implications for Process Control and IC Manufacturability,” 1995 GaAs IC Symposium Technical Digest, San Diego, pp.

124-127.

• Root, D.E., Fan, S., Meyer, J. “Technology Independent Non Quasi-Static FET Models by Direct Construction from

Automatically Characterized Device Data” 21st European Microwave Conf. Proceedings, Stuttgart, Germany, Sept 1991,

pp 927-932.

References


Agilent Restricted

Agilent measurement-based modeling and

design solutions

Customer

Applications Modeling

Simulation

&

Design

Electronic design

automation software

Measurements

Agilent Nonlinear Vector

Network Analyzer

Examples: •HP/Agilent (Root) models

•AgilentHBT model and extraction module

•X-parameters / NVNA

•NeuroFET


Agilent Restricted

Where to find Information about NeuroFET

• IC-CAP NeuroFET Webpage:

http://www.agilent.com/find/eesof-neurofet

• IC-CAP Device Modeling Software:

http://www.agilent.com/find/eesof-iccap








Agilent Restricted

Questions and Answers


You are invited

63

You can find more webcasts

www.agilent.com/find/eesof-innovations-in-eda

www.agilent.com/find/eesof-webcasts-recorded

Dr. Hongxiao Shao

Skyworks

Dr. Peter Zampardi

Skyworks

http://www.agilent.com/find/eesof-innovations-in-eda







http://www.agilent.com/find/eesof-webcasts-recorded





Measurement Based FET Modeling using Artifical Neural ... works for both common-source FET and...

Documents

Transcript of Measurement Based FET Modeling using Artifical Neural ... works for both common-source FET and...