Microelectronic Engineering 86 (2009) 140–144

Contents lists available at ScienceDirect

Microelectronic Engineering

journal homepage: www.elsevier .com/locate /mee

Impact of particles in ultra pure water on random yield loss in IC production

Faisal Wali a, D. Martin Knotter b,*, Auke Mud b, Fred G. Kuper a,b

a University of Twente, Hogekamp 32-46, 7500 AE Enschede, The Netherlandsb NXP Semiconductors, Gerstweg 2 (FB 0.049D), 6534 AE Nijmegen, The Netherlands

a r t i c l e i n f o

Article history:Received 26 March 2008Received in revised form 1 August 2008Accepted 29 September 2008Available online 18 October 2008

Keywords:Defect densityLinear regressionParticle contaminationUltra pure waterYield2-Proportion test

0167-9317/$ - see front matter � 2008 Elsevier B.V. Adoi:10.1016/j.mee.2008.09.046

* Corresponding author. Tel.: +31 (0) 24 353 2225;E-mail address: Martin.Knotter@nxp.com (D.M. Kn

a b s t r a c t

The influence of environmental particle contamination on offline measured defects and manufacturingyield in integrated circuits is discussed. One of the sources of particle contamination is ultra pure waterused in different production tools at different stages of processing. Particle count data measured in ultrapure water is compared with the offline defects caused by process tools and the relation has been statis-tically confirmed. Particle count data is also compared with the defect density of large size products. Animpact of particle contamination on yield of 4–6% has been found. In this study, fundamentals are pro-vided to define the meaningful specifications of ultra pure water for wafer fabrication.

� 2008 Elsevier B.V. All rights reserved.

1. Introduction

The semiconductor industry is trying to increase the yield bycontrolling the contamination in the environment. Yield is definedas the average ratio of the number of usable devices that pass dif-ferent tests to the number of maximum potentially usable devicesat process start. By determining the probability of defects locatedin the critical areas, it is possible to predict yield in the integratedcircuits (IC) [1,2].

Yield is divided into two components: systematic yield and ran-dom yield. Systematic yield represents the deviations in deviceand material parameters. Random defect yield is often associatedwith contamination problem. It was observed that a large part ofthe random defects is due to particle contamination coming onthe wafer during different process steps and caused random yieldloss [3].

As a result of the shrinkage of technological features into nano-scales, it is becoming more necessary to control the nano particlecontamination [4]. Most of these contaminations have been gener-ated or/and coming from the environment around the Fab. We con-sidered ultra pure water (UPW) as an important environmentalsource of particle contamination. UPW is used in many processsteps like wet etch, cleaning steps, and lithography. To define real-istic specifications for the wafer environment, it is essential todetermine the impact of contamination coming from the UPW.

ll rights reserved.

fax: +31 (0) 24 353 3323.otter).

For detection and monitoring of particulate defects in the pro-cess line dedicated optical inspection tools are used that are basedon light scattering principles [5]. A laser beam scans the surface ofthe wafer and is scattered by the surface defects. The detectionmethod is used in two types of monitor procedures.

� In the offline monitoring, bare wafers are processed in differenttools and the defects added are measured.

� In inline monitoring, production wafers are inspected after allcritical process steps.

In this study, first we statistically investigate the possible rela-tion between the particle concentration in UPW and defects mea-surements with offline monitoring. Secondly, the relationbetween the particle contamination in UPW and defectivity dataof a mature product in the Fab is analyzed.

2. Materials and methods

2.1. Method to analyze particle counts in UPW

Two water treatment plants provide UPW required for IC pro-duction to two Fabs of NXP semiconductors. The amount of particlecounts in UPW coming out of two installations is measured byusing particle-counting tools. In Fab-I particle measuring systems‘‘optical particle counter HSLIS M50e” and in Fab-II ‘‘laser particlecounter: ultra DI50” are installed. Both of these tools are capableto detect particles down to 50 nm (Latex Sphere Equivalents). Tool

F. Wali et al. / Microelectronic Engineering 86 (2009) 140–144 141

M50e gives daily average value of the particles per liter of UPW.While in the DI50, every 40 min data is collected. After particledetection, UPW is supplied directly to different tools without anyfurther filtration.

All the data has been separated into four different particle sizeclassifications, i.e., equal and bigger than 50 nm, 100 nm, 150 nm,and 200 nm. Table 1 shows the typical average value of particlecounts and standard deviations for each size measured by bothtools [6]. The problem with measuring small size particles is inter-ference of background noise by the tool. For this reason inline de-fects have only be correlated with particles >200 nm. It should benoted that given size does not mean actual size. This size can devi-ate of the reported size by a factor of two [7].

2.2. Method to analyze offline defects in Fab-I

Process steps considered in this study are litho and cleaning. Inboth Fabs, 15–30 different mask steps are involved in typical sili-con technology. Cleaning tools increase the yield by reducing thecontamination level on the surface of the wafer. A typical silicontechnology involves 50–60 cleaning steps. The performance oflitho tools and cleaning tools is monitored offline on a weekly ba-sis. Particles on the surface of the wafers are measured using theKLA Tencor 6200.

1.E+04

1.E+05

W

200

250

r

Particles in UPW

Defects in cleaning tool

2.3. Method to analyze yield in Fab-II

A mature product (product-X) manufactured in the Fab-II with adie area of 50.1 mm2 (with minimum feature size 0.35 lm) is se-lected. A manufacturing database has been used to collect informa-tion about lot identification, duration and date the product wasprocessed in each step, and yield of the lot. In this case, particlesare monitored on the surface of the wafers using KLA SP1 Classic.

Four process steps have been selected for this study. The pre-gate oxidation cleaning step (according to the specifications ofinternational technology road-map of semiconductor) is criticallysensitive to particle deposition [8]. Additionally a photo poly gatestep and a photo metal-1 step are also sensitive to particle depo-sition. Cleaning steps before anneal are considered to be a non-critical step for particle contamination and one was selected aswell.

1.E+00

1.E+01

1.E+02

1.E+03

0Analysis days

Part

icle

s / L

in U

P

0

50

100

150

Def

ects

/ w

afe

10080604020

Fig. 1. UPW particle monitored data and defects in cleaning tool sorted by date inFab-I.

2.4. Statistical methods

Linear regression is considered to be the easiest way to deter-mine possible relation between two different data sets. In thisstudy, linear regression with confidence interval of 95% is used todetermine a relation between the particle counts in UPW and de-fect density. Furthermore, the values of slope (unity 103 cm) andintercept (unity defects/cm2) are considered significantly differentfrom zero only if the values are larger than two times the standarddeviation. This means that the confidence level for such hypothesisis larger than 97.7% [9].

However, linear regression between particle counts in UPW andoffline defects in process tools is not possible because

Table 1Typical particle performance of Fab, measured with different tools.

Fab Particle measuring tool Sampling time

Fab-I HSLIS M50e DailyFab-II DI 50 40 min

� Particle count data in Fab-I is an average value per day so it ispossible that peaks of particles appeared in UPW at a differenttime than the monitor process is performed.

� Defects generated in process tools can be due to other processproblems than particles in UPW. This means that the set of waferdefects is larger than the (partially) overlapping set of particledata.

Therefore a ‘‘2-proportion test” has been used. This compares aproportion from a single sample of data against a known propor-tion in order to evaluate the relation between two data. In a ‘‘2-proportion test”, the p-value with the confidence interval (C.I.) of95% is determined. The hypothesis test (sample proportion is largerthan known proportion) was considered significant if its p-value isless than 0.05.

3. Results

3.1. Relation between particle counts in UPW and offline defects inFab-I

In Fig. 1, the amount of particle counts in UPW and defects gen-erated by a cleaning tool (Clean-1) are plotted against the dates ofmonitoring over the year. Data of UPW was measured every dayover a year but the cleaning tools are checked only once or twicein a week. Data is only plotted if both measurements on the sameday are available. This figure indicates that in some cases when theparticles level in the UPW increases the defects added in the clean-ing tools are also higher.

Similarly in Fig. 2, the amount of particle counts in UPW and de-fects generated by a litho tool (Litho-1) are plotted against the datesof monitoring over year. This figure shows that there are fluctua-tions both in the particles present in UPW and in litho defects. Com-pared to the cleaning tool, the relation in litho tool is more obvious.Eleven times large peaks coincide in both data. This suggests that

Typical readings (particles/L)

50 nm 100 nm 150 nm 200 nm

1600 (300) 900 (80) 200 (100) 50 (20)300 (50) 150 (30) 100 (20) 50 (20)

1.E+00

1.E+01

1.E+02

1.E+03

1.E+04

1.E+05

0Analysis date

Part

icle

s / L

in U

PW

0

30

60

90

120

150

180

210

240

Def

ects

/ w

afer

30025020015010050

Water analysisLitho-1

Fig. 2. UPW particle monitored data and defects in litho tool sorted by date in Fab-I.

Period A

0

200

400

600

0

Part

icle

s/L

in U

PW

Period A

Period B

Period C

36030024018012060Analysis day

Fig. 4. Time trend of particle (>200 nm) contamination level of UPW in Fab-II.

142 F. Wali et al. / Microelectronic Engineering 86 (2009) 140–144

particle contamination increase in the UPW leads to an increase innumber of defects added by the litho tool.

To make the relation more measurable, data was plotted in a X–Y graph to measure the significance with ‘‘2-proportion test”. InFig. 3, particles present in UPW are plotted against defects causedby litho tools. The out-of-control limit is defined on X-axis and Y-axis by calculating the mean value of both axes (excluding theexcursions with known root causes other than particles in UPW)and added three times the standard deviations (3r). The dashedlines indicate the out-of-control limit and divide the graph intofour regions A, B, C, and D.

Region D represents the area in which both particle counts inUPW and defects in the litho tool are under the control limits. Inregion B, defects in litho are higher than out-of-control limit butthe number of the particle counts in UPW was under the controllimit. In this case the defects are caused by the tool and are not re-lated to UPW excursions. Region C represents the area with higheramount of particle counts in UPW, but there is no increase in de-fects in the litho tools. This could be either caused by non-killingparticles or the peak in the UPW was short and did not coincidewith the measurement time in the process tools. In region A, thedefects caused by litho tools and particles present in UPW are bothhigher than the out-of-control limits. This region represents thearea where defects in litho tools and the particles present in theUPW are possibly related.

To determine whether the number of occurrences in area A islarger than statistically expected, the ‘‘2-proportion test” is per-formed. The known proportion is the number of UPW excursionsin respect to the total number of readings (A + C/A + B + C + D). Thisknown proportion is compared with data proportion. Data propor-

00

50

100

150

200

A

CD

B

+y

+x

Particles / L in UPW

Def

ects

/ w

afer

3σ

3σ

100008000600040002000

Fig. 3. UPW particle monitored data and defects in litho tool sorted by date without-of-control value in Fab-I.

tion is the number of out-of-control events determined in lithowhen there are excursions in UPW with respect to the total out-of-controls measured in litho tool (A/A + B). The significance ofthe difference for this litho case is expressed by the p-value of0.003. This statistical analysis has been performed on three otherlitho tools and three cleaning tools. In all cases the p-value is lowerthan 0.05, which points to a strong statistical relation between theparticle contamination and offline defects.

3.2. Relation between the particle counts in UPW and yield in Fab-II

Fab-II has a higher frequency of the particle monitor in UPWand a linear relation may be established with yield. Fig. 4 showsthe measured particle contamination in UPW. UPW was disturbedby an ‘‘incident” on day 65: during a filter exchange. Since the inci-dent can introduce other types of particles than normally present,the data is separated into three periods

� Period T1 is stable region before the incident in UPW with anaverage number of particles 40(20).

� Period T2 is unstable region during incident in UPW with anaverage number of particles 60(30).

� Period T3 is stable region after the incident in UPW with anaverage number of particles 20(10).

UPW data is compared with the defect density data of productX. The defect density of the product is measured after critical andnon-critical steps as described in Section 2.3.

3.2.1. Yield variations due to particles present in UPW during pre-gateoxidation clean

In Fig. 5, particles present in UPW during a cleaning step areplotted against the defectivity data of all lots processed in a year.Statistical analysis shows that during the cleaning step before gate

y = 20(6)·10-5 x + 0.12(0.01)

0

0.1

0.2

0.3

0

Particles/L in UPW

Def

ect d

ensi

ty (/

cm

2 )

600400200

Fig. 5. UPW particle monitored data and defect density of pre-gate cleaning processwith p = 0.001.

F. Wali et al. / Microelectronic Engineering 86 (2009) 140–144 143

oxidation UPW quality has a linear relation with the overall defec-tivity of the product. At a first glance, the slope (20(6) � 10�5)looks significantly larger than 0. However, during this period alarge excursion appeared in UPW data (Particles/L > 100) and itmay offset the statistical calculation. This is another reason fordividing the whole period into three periods.

In Fig. 5, the intercept of the line represents the defectivity ofthe process excluding the effect of particles in UPW during thisprocess. The total possible yield without particles can be calculatedusing the Poisson equation [10]

D ¼ �ðln YÞ=A ð1Þ

where Y is the yield, A is the area of the die and D is defect density.Calculated yield with the intercept value is 88.8%. The actual

average yield for these 117 lots is 87.8%. The 1.0% yield differencebetween the calculated yield and actual average yield is the yieldloss caused by the particle contamination in the UPW only in thisprocess step during the whole period.

3.2.1.1. Period T1. To exclude the impact of the excursion, the threedifferent periods are separately analyzed. Defectivity data of 25lots during the first period are shown in Fig. 6. These lots were pro-cessed before the incident in the UPW. During this time, there arefluctuations in both data, but overall there are no excursions. Theslope (10(4) � 10�4) is significantly larger than 0 and indicates thatdefect density of the lots increases with increasing particle con-tamination in UPW. The significance of the relation is shown withthe p-value of 0.04. The value of the y-intercept (0.10(0.02)) is usedto calculate the yield in the absence of particles. A 4.1% yield lossdue to particle contamination is found in this period.

3.2.1.2. Period T2. Fig. 7 presents the defectivity data of the 38 lotsthat were processed during the incident in the UPW. In this case

y = 10(4)·10-4 x + 0.10 (0.02)

0

0.1

0.2

0.3

0

Particles/L in UPW

Def

ect d

ensi

ty (/

cm

2 )

100755025

Fig. 6. UPW particle monitored data and defect density of pre-gate cleaning processin period T1 with p = 0.04.

y = 10(7)·10-5 x + 0.13 (0.01)

0

0.1

0.2

0.3

0

Particles/L in UPW

Def

ect d

ensi

ty (/

cm

2 )

600400200

Fig. 7. UPW particle monitored data and defect density of pre-gate cleaning processin period T2 with p = 0.08.

the slope is not significantly different from zero (10(7) � 10�5),which suggests no strong relation between the particle counts inUPW and defect density of the lots processed during period T2.

It is possible that during the incident different kind of particlesentered the Fab with UPW. The new particles may have lowerdeposition probability on the critical areas or/and lower kill ratio.

3.2.1.3. Period T3. Fig. 8 shows the defectivity data of the 54 lots thatwere processed after the incident in the UPW. This is a period inwhich there is no long term trend up or down in the particle data.The slope (15(3) � 10�4) is larger than 0, which again suggests arelation between defect density of the lots and particle counts inUPW. The impact of these particles on yield is calculated by usingthe intercept value of equation shown in Fig. 8. The yield losscaused by particles in UPW during this step in period T3 is 3.1%.

The above results show that in the stable periods (no excursionsin both data), a linear regression can be established between theparticles in UPW and defect density of the product with good sig-nificance (p < 0.05). In the next paragraphs, only data of the bestperiod T3 is shown.

3.2.2. Yield variations due to particles present in UPW during photopoly process

During the poly gate patterning, development of the photo polyis one of the steps in which UPW is used inside the litho tools. Ta-ble 2 shows that during the photo poly step UPW quality has aninfluence on the overall defectivity of the product with a slope(7(1) � 10�4) significantly larger than 0. In this case, the yield loss(due to the particle counts in UPW) calculated using the interceptis 1.4%.

3.2.3. Yield variations due to particles present in UPW during photometal-1 step

Photo metal-1 is also considered a critically sensitive to particledeposition. Particle data of UPW was compared with the defectiv-ity data of the product during metal-1 step and results are shownin Table 2. The slope (8(1) � 10�4) indicates a significant relationbetween both data. The intercept is used to calculate a yield lossof 1.5% due to the particles in UPW during this step.

y =15(3)·10-4 x + 0.08 (0.01)

0

0.1

0.2

0.3

Particles/L in UPW

Def

ect d

ensi

ty (/

cm

2 )

0 100755025

Fig. 8. UPW particle monitored data and defect density of pre-gate cleaning processin period T3 with p = 0.001.

Table 2Random yield loss in different process steps during period T3 due to particlecontamination in UPW.

Steps Slope(103 cm)

Intercept(defects/cm2)

Yield loss due toparticles (%)

Pre-gate oxidationclean

15(3) � 10�4 0.08 (0.01) 3.1 (0.9)

Photo poly 7(1) � 10�4 0.074 (0.003) 1.4 (0.3)Photo metal-1 8(1) � 10�4 0.078 (0.003) 1.5 (0.3)Pre-cleaning of anneal �7(4) � 10�4 0.13 (0.01) Not significant

144 F. Wali et al. / Microelectronic Engineering 86 (2009) 140–144

3.2.4. Yield variations due to particles present in UPW during thecleaning step before anneal

An anneal step is considered a non-critical step for particle con-tamination. To verify this statement, the particles data of UPWsupplied during cleaning before the anneal step was comparedwith yield. The slope is �7(4) � 10�4. It is statistically not signifi-cant from 0. This indicates that UPW quality has no influence onthe overall defectivity of the product during this non-critical step.

This last experiment excludes that observed relations in earlierthree critical processes are due to a coincidental time trend thatchange in the same direction.

4. Discussions

The impact of particle contamination in UPW is not well known.This makes it difficult to define the specification limits on the qual-ity of UPW. The presented statistical analysis indicate that the Fab-II specific particle contamination entering in Fab-II with UPWcauses a random yield loss (during period T3) up to a total of 6%in three critical process steps.

The defect density (D) due to particle present in UPW can be ex-pressed as shown in Eq. (2)

D ¼ NSKRPd ð2Þ

where N represents the number of particles/L in UPW, S (L/cm2) theamount of UPW that contact a product during the fabrication at thecritical process steps, KR represents the fraction of killing particles,and Pd the probability that particles deposit onto critical areas.

The value of N and S are known in many of the IC Fabs. The un-known value of KR strongly depends on the particle compositionand Pd varies with the process parameters, particle compositionand wafer surface composition. For Fab-II, where N = 200 parti-cles/L of size >50 nm, UPW causes 6% yield loss in three stepsand this gives the value of SKRPd = 0.00049. For this particular casekeeping SKRPd constant, yield loss less than 0.1% can be achievedadjusting the specification limits to N = 4 particles/L of UPW. If only1 ppm defects are allowed in Fab-II than this limit will be extre-mely challenging with N = 2 particles/1000 L of UPW!

To decrease the random yield loss in IC, it is important to reducethe contamination level in UPW as well as to reduce the depositionprobability. The latter can be understood if the deposition mecha-

nism of the particles is known. Our further study is focused on theprobability of deposition of particles. Initial results reveal thatdeposition of particles is not a random process as was describedin the Yield models [8].

5. Conclusions

A statistical approach identifying offline defects caused by par-ticle counts in UPW has been described. For the first time a directimpact of particle contamination on random yield loss is found. Ithas been shown that particle contamination in UPW in three crit-ical steps causes a yield loss up to 6%.

In IC fabrications, the specification limits on the quality of UPW isdefined with ‘‘rule of thumb” because the impact of particle contam-ination in UPW on production is not well known. This work indicatesa strong statistical relation between liquid-borne particle contami-nation entering the Fab and random yield loss. These results gavean ultimate opportunity to establish a ‘‘realistic specification” onthe UPW to control the random yield loss in IC’s fabrication.

Acknowledgements

The authors would like to acknowledge Mike Bolt and JeroenWildoer from NXP Semiconductor for practical support in sortingthe Fab data. For financial support we acknowledge the EuropeanMEDEA+ program (2T102; HYMNE).

References

[1] W. Maly, IEEE Trans. Comput.-Aided Design Integr. Syst. 10 (1985) 166.[2] C.H. Stapper, Microelectron. J. 20 (1989) 129.[3] T. Kim, W. Kuo, IEEE Trans. Semicond. Manuf. 12 (1999) 485.[4] D.K. de Vries, P.J. Volf, IEEE Trans. Semicond. Manuf. 17 (2004) 188.[5] F. Holsteyns, J. Roels, K. Kenis, Q.T. Le, P.W. Mertens, IEEE Proc.-Comput. Digit.

Technol. 3 (2003) 7803.[6] In this paper, all values written in between parentheses represent the standard

deviation on the numbers noted before it.[7] D.M. Knotter, S.A.M. Wolters, M.A. Kasanrokijat, Part. Sci. Technol. 25 (2007)

435.[8] <http://www.itrs.net/Links/2007Winter/2007_Winter_Presentations/

Presentations.html>.[9] D.L. Massart, B.G.M. Vandeginste, S.N. Deming, Y. Michotte, L. Kaufman (Eds.),

Chemometrics: A Textbook, Elsevier, Amsterdam, 1998 (Chapter 5).[10] B.T. Murphy, Proc. IEEE 52 (1964) 1537.