. Taylor 6i FrancisINT. J. PROD. RES., 2002, VOL. 40, NO. 10, 2207-2223 V ^

Wafer bin map recognition using a neural network approach

S. F. LIUt. F. L. CHENt* and W. B. LU|

Although the fabrication of modern integrated circuits uses highly automatic andprecisely controlled operations, equipment malfunctions or process drifts are stillinevitable owing to the high complexity involved in the hundreds of processingsteps. To detect the existence of these problems at the earliest stage, some import-ant analytical tools must he applied. Among them is wafer bin map analysis.When the bin map exhibits specific patterns, it is usually a clue that equipmentproblems or process variations have occurred. The aim was to develop an intel-ligent system that could automatically recognize wafer bin map patterns and aidin the diagnosis of failure causes. A neural network architecture named AdaptiveResonance Theory Network I was adopted for the purpose. Actual data collectedfrom a semiconductor manufacturing company in Taiwan were used for systemverification. Experimental results show that with an adequate parameter, theneural network can successfully recognize and distinguish random and systematicwafer bin map patterns.

I. IntroductionIti matiy coutitries. such as the USA, Japati, South Korea and Taiwan, semicon-

ductor manufacturing has emerged as one of the most important industries.However, complexity and cost continue to increase with each new generation ofsemiconductor products. Semiconductor companies are being forced to improvetheir manufacturing capability and develop their own techniques to create processsteps with tighter tolerances and smaller geometries on larger dies. To compete inthis world-wide market, quality improvement and yield enhancement have receivedincreasing attention. For these reasons, every semiconductor manufacturing com-pany makes an effort to monitor and control the manufacturing processes to reducevariations and enhance yield (Mirza et al. 1995).

Typically, all dies on a wafer after the completion of the wafer manufacturingprocess, must go through the so-called circuit probe (CP) test. The purpose of the CPis to test the electrical functions of a die to determine the inferior dies on each wafer.CP consists of serial pass-or-fail functional tests. The CP results determine the gradeoi each die on a wafer. Each grade will be presented using a unique bin code. Ingeneral, for different product types and different companies, the number of diegrades differs. Five. 10 or even 75 die grades are possible (Cox and Reynolds1996). A good die can pass the functional tests thoroughly and be assigned withthe best grade. Other bin code assignments usually indicate different degrees ofquality inferiority. Table 1 illustrates an example of CP results. Here, all of the

Revision received July 2001.t Department of Industrial Engineering and Engineering Management, National Tsing

Hua University. Hsinchu. Taiwan, R.O.C.t Taiwan Semiconductor Manufacturing Company (TSMC). Hsinchu. Taiwan, R.O.C.*To whom correspondence should be addressed, e-mail: flchen(uIE.nthu.edu.tw

tmcniuli.iiial Journal iil I'roilmiiwi Rncarrh ISSN 0(120 7543 print/ISSN Llhfi- 58KX online i ' 2(H)2 Taylor & Francis Lidhitp: . www.tiindr.co.uk/'journals

DOI: ll),10!<O/O02O754O2l0122275

2208 S. F. Liu el al.

Bin Count Yield (%)

DB 1 GOODDB 5 OPEN/SHORTDB 6 I/P LEAKAGE HIGH

87542

7800

DB 16 GROSS FUNCTION (6.5V) 5 0DB 18 CKB FUNCTION (4V) 22 I

Table 1. Example of CP results.

dies are bad except for tbose with Bin Code I. After all of the dies in a wafer havebeen CP-tested, u so-called wafer bin map (WBM) can be generated to depict thedistribution of quality grades. Since the occurrence of any quality inferiority canusually be attributed to some specific causes, it is believed that WBM analysis canhelp determine the possible causes of failures and help to devise solutions to preventthe reoccurrence of these failures.

Patterns existing in WBMs may be classified into random and systematic types(Kaempf 1995) (figure 1). Typical examples of systematic WBMs include cluster,circle, arc, doughnut and Bull's-eye patterns. Traditionally, the analysis of thesepatterns relies heavily upon the engineers' experience and domain knowledge.Since this 'pocket' knowledge is difficult to accumulate, several advanced techniqueshave been developed to aid in the diagnosis of WBM patterns. These techniques canbe categorized into statistic and artificial intelligence approaches.

1.1. Statistic approachIt presented a join-count statistics based spatial clustering method to analyse

WBMs in two-dimensional space. Suppose that the /th die is the centre of aWBM. the relationship between this die and the four neighbouring dies could bedefined as follows:

y(GB) ^'•</

amraaBfflfflSfflfflH

'BfflfflfflEBfflfflfflB'

\BfflfflfflfflEffifflHy

Figure I. Systematic (left) and random (right) WBMs (Kaempf 1995).

IVBM reeognition using a neural network approaeh 2209

where. y(GG).y(GB) and /(BB) are the number of times thai two neighbouring diesare good, one is good while the other is bad, and all are bad. respectively. \^ Y — I,die / is good. If K — 0, die / is bad. 6^ is a neighbouring index, and = 1 if dies / and,/are neighbours. Otherwise, ti,j ^ 0. The index below evaluates whether the dies onWBMs are scattered randomly:

When 8 approaches 0, the dies are randomly scattered on the WBMs and no sys-tematic patterns exist. When 0 > 0, the dies are scattered systematically (Taam andHamada 1993). A similar approach was also adopted by ICaempf (1995) to determinethe distribution of dies on WBMs using a binomial test.

Mirza et al. (1995) divided the patterns on WBMs into gross (systematic) andlocal (random) types and proposed methods to distinguish these two types of pat-terns. Gross failure means that a large die area or even all of the dies on a wafer arebad. This kind of failure may be caused by temperature variations during processingor other serious reasons. For Local failure, small die areas are bad. Pinhoies fromparticles or spot defects are usually the main causes. Mirza et al. used a Gibbs/Markov random field (G/MRF) model to develop a Spatial Randomness Test meth-odology. Because the G/MRh model may be used to determine object independenceor dependence, it can recognize gross failure and local failure on WBMs.

Friedman et al. (1997) developed a two-stage spatial pattern analysis approach.In the first stage, good or bad die images (white and black; 0 and 1) were precededwith an averaging operation where each die was averaged by its 3 x 3 or 5 x 5neighbourhoods. The original die is then replaced by the obtained average value..\ smoothed grey-level WBM could be obtained after this stage. A threshold valuemust be selected in the second stage. If the grey level of a die exceeds the thresholdvalue, it is determined to be bad; otherwise, it is good.

It must be noted that although the statistical based approaches can detect theexistence of systematic patterns, they lack the capability to identify the correctpattern types. It is therefore diflicult to devise solutiotis for the detected failures.

1.2. Artificial intelligent approachLin (1998) used machine vision techniques to preprocess WBMs. A supervised

neural network architecture (Back Propagation) was then applied to recognize thesystematic patterns. In his research, three limitations were determined, which arediscussed as follows.

• The image-based approach requires much storage space. It also requires longprocessing time in the preprocessing stage.

• In the back propagation (BP) architecture, the number of output nodes mustbe defined in advance. In fact, a skilled manufacturing engineer cannot be surehow many systematic patterns actually exist. If new patterns occur, they willnot be detected and recognized.

• Sufficient symbolic training patterns must be provided. A tremendous amounto{ time will be consumed in training new patterns.

Similar to Lin's work, software named NeuralNet^*^ Engineering Data Analysis(NEDA), developed by Defect & Yield Management, Inc. (DYM), was applied to

2210 S. F. Liu et al.

neural network techniques to detect similar patterns. Enough templates must beprovided to train the knowledge base. Meanwhile, users cannot retrain new WBMpatterns and hence the flexibility is limited.

In view of the limitations in the above methods, this research was intended todevelop an intelligent algorithm for detecting a greater number of different WBMs.To speed up the detection process atid to provide flexibility for tiew patterti learnitig.a neural network architecture named Adaptive Resonance Theory (ARTl) wasadopted in this research. Details about the development of this algorithm are pre-sented below.

2. Recognition of patterned wafer bin maps2.1. Design the input vector

Before the collected wafer bin maps can be analysed, the input data formatdesign is crucial for sample pretraining and future recall procedures. The inputtraining sample vector is also named the characteristic vector. The number ofprocessing units depends upon the type of problems to be studied. A lineartransformation function is usually used to pass the input vector into the nextlayer. In solving a problem, in which the recognition performance may beinfluenced by too much noise (Lin 1998). a noise reduction procedure was firstapptied to all of the training samples. The input vector design differs for everyproduct type. The number of dies in a specific product type determines thenumber of nodes in the input layer. The detailed notations are explained belowand shown in figure 2:

A' number of dies per wafer,X, input vector of the /th sample data, andXij 7th element of the input vector.

•DIDDIDIDDDDDDDI I I I I I I IDDI

DniDDnDDDDDIDDDDIDDCIDIDIDDDIDDDDDDDIDIDDDDDDIDIDDDDIDIDIDDD

DDDDDDDDDB>/

Figure 2. Relationship between the input vector and dies.

WBM recognition using a neural network approach 2211

where

1, if the bin code is not equal to 0;

0, otherwise.

After the required sample training data for the neural network has been pro-vided, the number of nodes in the input layer and their corresponding values must bedefined to start the training process. In this research, the unsupervised neural net-work was trained according to different product types. The reasons are as follows.

• The number of input processing units is the total number of dies for a wafer.Different product types have different numbers of dies for each type of wafer.The collection of weights must be prepared by product type.

• Since the unsupervised neural network can also classify a single input pattern,data insufficiency is not a critical concern in the neural network architecture.After training with the available samples, any type of pattern can be recog-nized.

• Although the life cycles of certain products may not be long, a considerablenumber of wafers will be produced in fabrication. Pattern recognition trainingis also required for these types of products.

2.2. Seleetion of network modelBecause of the huge number of WBMs. it is difficult and time consuming to

decide how many types of patterns may appear. That is. the number of outputpatterns is unpredictable. For this reason, network learning can only be accom-plished using the input data alone. This is the reason why unsupervised learningwas adopted. Chen and Liu (2000) used ARTl and SOM architectures to comparethe pattern recognition performance on defect maps (under wafer level considera-tion). The results showed that ARTl performed better than SOM. Therefore, amongthe many kinds of unsupervised neural networks architectures, the AdaptiveResonance Theory (ART) was considered in this research. The ART networkaccepts input vectors that are classified according to the stored pattern they mostresemble. The ART network also provides a mechanism allowing adaptive expan-sion of the neuron output layers until an adequate size is reached based on thenumber of classes. The ART network can adaptively create a new neuron corre-sponding to an input pattern if that pattern is determined to be 'sufficiently' differentfrom the existing clusters. This determination, called the vigilance test, is incorpo-rated into the adaptive backward network. The ART network modifications areARTl and ART2 (Pandya and Macy 1996). The ARTl model clusters binaryinput patterns and ART2 clusters analogue input patterns. Considering the signifi-cant processing timesaving, ARTl was selected here. A good knowledge-basedsystem must satisfy two characteristics: stability and plasticity. ARTl uses a vigi-lance test to learn new patterns without forgetting old knowledge and thus can solvethe contradiction between stability and plasticity. The vigilance test concept isdescribed as follows.

• If the characteristic of a new pattern is quite similar to a previously storedpattern (vigilance test passed), only a slight modification of the knowledge

2212 S. F. Liu et al.

contained in the old pattern will be executed. The characteristics ofthe old andnew patterns can be satisfied and the old knowledge can be properly retained.The stability ofthe system can be maintained.

• If the characteristics of a new pattern are not similar to any ofthe previouslystored patterns (vigilance test failed), new knowledge for the new pattern willbe created. This implies quick learning of a new pattern, or the so-calledplasticity.

The ARTI architecture construction includes an input layer, network connectionand output layer (figure 3). It uses an output-processing unit to present a certaincluster. Every connection weight between the input layer and the output units indi-cates the characteristics of a specific cluster. The number of output processing unitspassing the vigilance test may exceed one. so the network utilizes a match value tocontrol the output processing units. The vigilance test is first applied to the outputprocessing units with the highest match value. In general, the higher the match valuepossessed by an output-processing unit, the higher its similarity.

The vigilance value plays an important role in the ARTI algorithm. This valuecan be used to distinguish the similar patterns. The higher the assigned vigilancevalue, the fewer the number of output units that will pass the vigilance test. On thecontrary, the lower the assigned vigilance value, the greater the number of outputunits that will pass the vigilance tests. In other words, this systetn can detect similarbut different types of clusters.

The detailed steps of the ARTI algorithm developed in this research are statedbelow and shown in figure 4.

(1) Set network parameterNout = 1

(2) Set initial weighting matrix. W

W'\i\[\\ - I (a similar value)

Tjr value)

(3) Input a training sample vector X

output VBCtCf output la/er: clusters

input vaiablesinput vector

Figure 3. ARTI network.

WBM recognition using a neural network approach

set Noul

set

sei Pvalue

tnput pattern vector X

compute match values

select the outputprocessing unit withmax. match value

set= /can * 1

select the outputprocessing unit withmax. match value

no new clusteris created ,

the learning cycleended ??

Figure 4. ART! network procedure.

(4) Calculate the match value

J

2213

set

kount = 0

2214 S. F. Liu et al.

(5) Determine the maximum match value

net[j*] = m-dx{net[l])J

(6) Calculate the similar values

^ \\x\\(7) Judge and test the vigilance value

ify* < ^(vigilance value)go to step (8) (not similar, test another output processing unit)or elsego to step (9) (similar enough, modify weights)

(8) Test any available output processing unitsif Icount < Nouttest another output processing units to find the output processing unitwith max. match value.set kount = Icount + 1set net[j*] ^ 0go to step (5)or else (no other output processing units require testing)(a) create a new cluster outputset Nout = Nout + 1set new weight

W'{i\[Nout] =

(b) set output unit value

else

(c) go to step (3)(9) Modify the weight

(a) modify weights

WBM recognition using a neural network approach

(b) set output unit value

'n/i = Ior else

2215

(c) if a learning cycle is finished and no new clusters are created, the finalresults are obtained and the network ceases to learn situations,or else go to step (3)

3. Network training iThere are two phases in the ARTl architecture. One is the training process and

the other is the recognition process (figure 5). After the intelligent patterned WBMsrecognition system was conceptually designed, a practical software system was devel-oped for system implementation and verification. This system was developed usingBorland C++, under a Microsoft Windows 98 operating platform. Actual data froma product with 387 dies for a wafer were provided by a semiconductor company andtested through this system. Before the ARTl network can be used to identify thesystematic pattern types on the wafers, the network must be trained. Since ARTl isan unsupervised network, it is not necessary to link the input and output vectors toattain good recognition. That is. we do not need to send the training output resultsback into the network.

Because the learning is unsupervised, the ARTl network convergent processcannot be evaluated using error rate or square error. Total distance is used in thiskind of network and it can be defined as:

Total distance — min d'-

where (/'' is the distance between the pth samples and the /th output processingunit = 1 - similarity between the/?th samples and the jth output processing unit.

During our training processes, the ARTl network converged in three cycles. Thetime utilized for training 20 samples on a PC with an Intel Pentium 166 and 64 MBRAM was approximately 3 s.

Phase I : ARTlTrainingProcess

Phasf II: ARTlRecognition

Process

ARTl Neural Ne(*ork

Training Process

ARTl Neural NetworkRecognition Process

PalierneijWBMs

S^slemalic orRandomWBM;

Figure 5. Two phases in the ARTl network.

2216 S. F. Liu et al.

4. Experimental results 'To evaluate the performance ofthe developed system, the ARTl network must

be trained and tested using actual samples. Owing to the business security considera-tion and the fact that ARTl network does not need many samples for training. Asemiconductor manufacturing company in Taiwan provided only 20 WBM samples.Each of these map samples consisted of 387 dies. For these 20 WBMs samples,skilled manufacturing engineers identified the following three types of systematicpatterns:

(1) Type 1: ring-type with small thickness (figure 6). '(2) Type II: ring-type with large thickness {figure 7). 'i(3) Type 111: Bull's eye (figure 8).

The ARTl network was trained and adjusted to recognize the above three systematicpatterns. It adjusted the number of output nodes (number of pattern types) accord-ing to the vigilance value. With an adequate vigilance value, the training WBMs willbe assigned to the right group. According to the pilot runs in this study, when thevigilancevalueisset in the range (0.330, 0.340), the classification of all of the trainingWBMs would be the same as the results from skilled manufacturing engineers. Theexact vigilance value used in the ARTl training was set to 0.335. The ARTl networkconverged after three cycles.

With the trained ARTl network, the weight parameters were used to test theother WBMs. Test sample classification was determined using the match and thresh-old values. The match value (ranging from 0 to 1) indicates the similarity between atest sample and a certain pattern type. The higher the match value, the more likely atest sample will pass the threshold and be allocated to a group. If the match value issmaller than the threshold value, the current test sample will not be allocated to anygroup. In this situation, the test sample was either a random pattern or an unrecog-nized pattern. If the match value is greater than the threshold value, the test samplewill be recognized as one of the three known systematic patterns or unknown pat-terns and further analysis will be required. Figure 9 illustrates the procedures fortesting WBMs and the possible test results are summarized below.

(1) Match value < threshold value. In this situation, the test sample is eitherrecognized as a random or new pattern that has not yet been recognized.

(2) Match value > threshold value. The test sample is classified as one of thethree pattern types if only one match value exceeds the threshold value. Iftwo or more match values are greater than the threshold value, the currenttesting sample will be identified as an unknown pattern.

Seventy-five WBMs were tested in this research and all were of the same producttype (DRAM) but from three different lots. Among these 75 WBMs. random pat-terns and systematic patterns were included. The experimental results are summar-ized in table 2. Match [I] to match [3] in table 2 represent the match value of a testsample to the three systematic patterns.

According to the results in table 2, it is clear that the match values of the 50WBMs from Lot_2 and Lot_3 are quiet small, which means that these maps do notmatch any of the known patterns. Only random patterns exist on the maps fromLot_2 and Lot_3. These results are consistent with the recognition results verified bythe skilled manufacturing engineers. The WBM test results from Lot_l are of more

WBM recognition using a neural network approach 2217

I I 1 I 9 I! 1 1 1 S ] 1 9 9 3 S 2

1 1 1 2 0 0 0 2 2 0 9 0 2 2 2SI 1 1 g g s 2 7 g 2 1 1 1 9 1s i l O j g g t g z g g B O O i

S I. ' I I S

l o g g z g g g g i B i o o o o s G i ti g g g i B o g o o t i i i g B S i

11

I J D D S B O D O S1 9 t D 1 0 « g

1 J i 1 1 S

-t - l-*-l-t-f-t-t-t-1

no.l

o ! D 9 E i o g g M M] 9 0 0 ; 0 S 0 g e D S 0 S

1E D D S I g 0 1 gS B I t 1 !

no 3

i t i t i l5 2 2 5 5 2 S S S : 0 B

2 I 2 i 0 D 0 5 0 D 0 2 Z Q 0S 2 2 5 D 0 D 0 D 0 Q 2 S II S 52 0 5 D M

a o o s s zZ Z Z Z 0 D 5 a 0 0 0 D 0 0 0 0 0 0 0 3 D 25 2 5 i ] l i l ) a i } 0 0 a 0 a Q O 0 0 0 ! l 2 EB 1 0 0 S a D D D 0 n 0 D 0 D 5 0 0 I B [) 2

z z a s a o O a a o

I 0 0 i 2 a 5J J D D O D Q I O D O I O S

1 2 0 0 0 0 0 2 0 1S 2 2 9 9 2

10 I

1 1 I12 I

9 S3 9 2 2 5 2

2 2 2 2 2 2 O O S J 9 O5 2 2 2 2 0 2 0 0 2 2 0 2 2 92 2 2 2 2 S O 2 T 2 0 2 2 0 2 22 2 0 3 0 0 0 0 0 2 0 0 0 2 2 2t l 2 0 O O 0 0 0 0 0 0 2 f i 0 0 2 O2 O O O O 0 0 0 0 0 0 D O 0 0 2 S 2

222 2 25 2 2

0 D 0 0 0 0 a o a D 2 ! E2 S O ( l 0 ( I O D O M O 0 0 0 D D Q 9 2 3 !2 2 5 M 2 D 0 0 0 U S 02 2 5 M 2 D 0 0 0 U S 0 2 0 a ( l 5 a 2 2 i2 2 Z S < l D A ; a 0 0 0 [ I O D D D I } 2 2 2 22 2 t S O 0 D 0 0 5 0 0 < ) O O O O O O I J ! 2S 2 O t 2 0 O 0 D 0 0 D D a D O O 5 0 2 2 22 S 2 ! 9 Z 0 0 n D D I ] S G O ] ]0 D D ] S G O I ] | ] O Q 0 2 5

2 2 : 2 0 0 0 0 D l l ] O O O 0 0 5 0 5 2 22 2 2 2 2 0 0 0 0 0 0 0 0 5 0 0 2 0 2 5

2 0 0 0 5 5 0 5 0 0 0 0 6 0 0 0 2 22 2 2 0 M O O ^ 0 0 G a 0 0 0 2 3

5 0 2 0 D 0 2 0 5 0 5 0 2 0 2 23 3 D 0 e 0 0 0 5 0 0 0 2 5

0 0 0 1 9 0 0 0 2 25 2 S S 9 5

no, 4 no.5ir

li222 29 2

2 2 2 2 2 2 2 2 2 2 2 52 2 , 2 1 2 222 2 2 1 2 1 )2 1 2 1 2 0 5 J 1 2 0 2 2 2 2 2

2 2 2 9 0 0 0 0 2 0 2 0 0 0 0 2 5 22e2tl20ltD020000002i223 2 2 I i 0 2 0 0 0 2 0 2

0 0 0 0 ( 1 9 ( 1 1 1 2 2 3D a D D D D } 2

2 2 2 M i 9 0 0 0 5 2 0 B2 2 2 2 Z 0 Z M O i O 5 5 5 5 0 0 2 53 2 0 2 0 0 D S 5 5 S B 5

2 2 0 0 0 2 0 0 0 0 0 0 0 0 0 ( 1 0 0 2 22 D &DC 0 0 02 0 0 0 0 0 T 2 2 aZ 2 O 0 0 0 0 0 i 0 0 0 0 0 Z 2 2 52 t f O : 0 0 2 0 9 0 0 5 0 0 2 B3 3 5 0 0 0 0 0 0 0 0 0 2 2

0 0 0 0 0 5 0 0 5 25 2 2 2 J 5

2 0 i » 9 0 0 5 5 b 5 & B 0 0 2 S2 5 2 0 0 2 2 9 0 0 0 0 0 5 5 5 5 5 5 2 2 S2 Z ! a D 0 9 S 9 0 0 0 0 0 5

:02 2 0 0 0 0 1 [ ) 9 0 02 S 9 0 2 2 2 0 2 0 0 O Z O 2 23 3 0 2 0 0 0 D D S 2 5 2 2

2

no.6 no. 7

«1I3*9971>

IIn121314ISIt1111I t2g212!23U»

T 1 R H ill

193 9 9 2 9 2

S 9 2 3 2 Z S 0 2 0 Z Z I B V2 2 9 2 0 g g O D B D 2 2 Z 2 I

2 2 g g D 0 B ( l Q D g D g p E g g ; i l 2S 2 2 0 5 0 0 g g D g g 2 t i 0 2 g D 5 6

t ' o Z D M S M g g g g S S g g g 2 2

£ 2 I d O D 0 2 0 0 D C l 0 0 5 Q Q f l 2 ^ }

' i ls lsi i isi iHssiH's g 2 o g g g g g g o g o g 2 2 2 2

2 g g g 2 9 g g B g g g 2 D 2 2a a g z o o g o o g o g n z

1 g 9 2 2 D B E Z 22 1 2 1 2 1

no, 9

"TS 1—(f1zi4t

•Iign1!1314 1IEItIT1611ag21zzn242t

5 a ii ai—» H t-t 1 • -h «M » ••-4't-* '*-*-+-t-1-

9 •

Z ^ S Q n o a o o o n S a n n n a A rS O Z F i l l f l O D D D P P r D D Q J l A ' !

; I I j J J J J J J s 0 6 M M g g 5

i g g g ! 0 ! g g 5 0 ! g s ' 0 ! " °2 2 ? 0 & D B Q D Q a o o 9 n ? n n n

9 1 2 2 B s s S g g 0 g 0 g g " ' " "

Z B n l g g g ' g o o B o o z B S S

2 9 g g 0 B 3 s 1 b21 IJ f iS

no, 13

_9Z

Figure 6. Type I defect — ring-type with a small thickness.

2218 S. F. Liu et al.

Class 2

I 2 0 0 1 I 0 0 I

1 0 < g 9 9 g 1 :

1 9 9 g :9 g g !

I Z B Z

D0.8

Z Z Z I S l Z i l l l E

3 Z S g !s i 1 s 1 g1 i 1 ! t 01 1 S t 0 9Z I 1 ! 2 1i t : ! z 9::: z s Bt ! s ? g gs I a g gs z z g gS ! Z SS I S 9

i 3

o u s t0 0

9Bggts09

9i

ggBgg0

9

9i

9 99 gD gg g

D gg s! 09 99 99 99 9

9 It Z

9 B gg g gg g g

i l i t m i l lo o g 1 9 a 1 1 10 9 9 S 9 g 1 I Z

g g o o 9 9 g g g t t l0 0 0g 9 99 9 09 i 99 T 9

I B 19 g g

9 9 9 0 0 0 0 2 59 9 B B 0 0 g Z Z

9 1 g 0 g ! s z !9 g 0 0 0 ) 1 i 3g n o g 1 z g I !

g 0 9 s D s !

9 1 9 9 I g 10 5 9 1 1 1

0 0 j 0 e ! B sits Z 9 1

no. 12

3 1 1 2 i 9 Z g

i e i s a g o o 9 g 9 [! D S g ! t 0 9 9 9 t !i 1 g g 0 n 9 I I 9 9 g I

1 :11!!i i t i za i i s i !I S ! S Z 3oiimS S I t I ig I i I i s

g : s 19 e ! ! '

no. 14i! JT i r

no,15

j 211 e z s! z z i i i H s 9 z z : : g : i ! |> z ) z g g i 9 91B g 0 9 3 ] 3 1 ;

i i t ! i t o z 9 g g o s t i g g < i : z z i

) Z i r z ! 3 Z B t g > S 2 ! l3 i 5 2 g 9 g s g t j s s z

i m z s s 3 1 s zs 1 1 1 3 s

i i l-t-'-l-l-l-f-

S Z Z Z Z I

z s z 1 1 1 :i i z z i o jZ S 2 1 1 2 9

+-(-(-•

I

i 1 1

1 1 0(

0 g g!: 9 a g9 9 i g

1-*

1i S

9 9

i aB g0 g

•-•

I,

. 1

) (0 90 9

• -•-+ • * - ! -

I 1 1 !SZ ! I Z Zs s 1 ng i ! 9 sS I ! 9 1

9 9 i

9 g i

•+-I

z21

Z i 1

I' 5

no. 16 no. 17

13

1

Ii

1 Z! i 1 1 1 2

i 1 Z Z S I 1

i j z z i i i i z z tt 1 1 2 2 SS Z

Z 2 iZI2i 1 !i ii

t z z zi Z Z Z1 1 1 1Z I S S1 1 1 i

B s 1 0 g 9 g1 1 ; 1 1 1 1I 0 z s 9 g gi t 0 s 9 g gg o z 1 1 B (t 1 9 u 1 S I

i 1

B I

g 'i

\i12IZ

: 0 g zg 0i 9I 91 9

9 Z9 9

i.

11i 2 ZZ 1 23 31ill

1

9 1 1 I t ZB gI

1 1 z 9 1 g g 0 1 g 9 1

no. 18 no 19

I I I 1 I ! !11 M H I Ii ! 2 i i ; J!

1 I ! 1 I I ' ' I?1!IID t 1 Ii H i I > 9 > 1? I z 11 z g 11

I I 9 I < 9 I 9 Ii iz i i i i t !

- 1 > 9 g > z 1 9 1

1 E 1 2 I > 9 i 9 IMil .; 1 9 2 ! > g <I I 1 I I I I I

no 20

Figure 7. Type !l - ring-type with a large thickness.

IVBM recognition using a neural network approach 2219

01231567i10111213It15IB17IBJO20212223U2S

01

z345g

8910IIil

14IS1817181920Zl

23

Zb

Tl 5 [B 1! iS E

5 5

2 2 2 2 5 5 0 2 2 5 3 25 0 2 5 2 0 5 5 5 5 D 2 2 D 50 5 I 1 5 5 5 5 5 5 5 D D D 1 2 2

5 2 5 5 5 5 5 5 5 5 5 5 0 0 0 0 0 5Z 5 0 5 S 5 9 9 5 5 5 5 5 5 0 5 Z D 0 D5 2 5 S 5 5 5 5 5 5 5 5 D 5 5 D 0 0 0 2 2

2 S 5 5 5 B 5 5 5 5 5 5 5 5 5 5 0 0 0 5 2 2

2 S 5 5 5 5 5 5 5 5 5 5 0 5 5 5 S 5 0 0 0 5

2 5 5 5 5 5 5 5 S 5 5 5 5 5 5 5 5 5 5 1 0 25 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 0 5 5 0 25 5 5 5 5 5 5 5 S 5 5 5 5 5 5 5 0 2 0 9 22 2 2 5 5 5 5 S 5 5 5 5 5 5 5 5 5 5 Z 22 5 3 5 0 5 5 S 0 5 0 5 5 5 0 5 0 BZ 5 5 5 D 0 0 2 S 5 0 0 2 0 0 0 0 Z2 0 2 2 2 O 1 D 0 5 0 0 D 0 0 23 3 0 5 0 D 0 5 D O B Z 0 9

2 5 Z 0 0 D 0 0 0 25 5 2 5 55

i i-i . t , , , ^ w iH » * 1 * t +

no, 2Tf 5 IB IS !! S5

S GZ 5 2 S S 2

S S Z 2 Z Z Z Z Z 2 2 G2 5 2 Z 5 2 S 5 5 Z 0 2 2 Z DZ Z 5 5 5 5 B 5 5 0 0 2 2 0 2 2

! 2 2 S 5 S 5 5 5 5 5 5 S 0 5 D 0 0 Z 2

2 2 2 9 5 5 5 Ji S 5 9 9 S S S 5 5 9 5 9 9 2

5 5 5 5 5 S S S 5 0 0 5 0 0 5 5 5 5 5 S S Z5 5 5 5 5 5 S 5 5 5 5 0 5 0 5 S 5 5 5 E 2 5

2 2 2 B 5 5 0 5 B B B 5 B 5 5 5 0 5 5 2

S 2 S 2 2 2 2 5 B B B 0 O D 2 2

S S 2 2 2 S

no. 13

! _ . , 5 , _ lb 15 "Si IS 10123i

GTt

101112laIt19IB17181020Zl2223

na

5 iS S S 5 5 Z

5 2 5 2 5 5 5 5 5 5 5 22 5 5 5 5 5 5 5 5 5 5 5 5 2 22 5 5 5 5 5 5 5 5 5 5 5 5 S 5 2

2 2 5 5 5 5 5 5 5 5 5 5 5 5 5 5 3 32 2 5 5 5 5 5 5 5 9 5 S b 5 5 5 5 9 5 53 2 2 5 5 5 5 5 5 5 5 S 5 5 5 5 5 5 5 2 5

2 2 Z 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 9 5 2 35 0 0 5 5 5 5 5 S 5 5 5 5 5 5 5 5 5 5 5 9 9

D 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 9 9 9Z 0 5 5 5 5 5 5 5 9 9 5 5 5 5 5 3 S 5 5 5 55 5 9 & B 5 5 5 5 9 9 9 S 5 5 5 5 5 5 S 5 55 2 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 9 5 52 5 5 5 5 5 5 5 5 5 S 5 5 5 5 5 5 5 5 5 55 5 5 S 5 5 5 5 5 5 5 S 5 5 5 5 5 5 9 5Z 5 5 5 5 5 5 5 5 5 S 5 5 5 5 5 5 52 5 0 5 5 5 5 5 5 5 5 5 5 5 5 5 5 55 3 5 0 5 9 5 5 5 5 5 5 5 9 2 53 3 0 B 5 9 0 5 0 5 5 5 9 S

9 0 9 0 5 0 5 9 9 95 5 5 5 5 9

no, 10

Figure 8. Type III defect —BullVeye WBMs.

Sample lex ling

Figure 9. Sample testing procedure.

2220

Map

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

Lot_l

match[l]-0.134200match[2]-0.110980match[3]-0.054600match!!]-0.638739match!2] = 0.510920matchp]-0.425! 00match[I]-O.33723Omatch[21 = 0.251740match!3] = 0.234000match[l] =0.246600match[2]-0.170900match[3]-0.! 20900match[!l-0.374500match!2]-0.264140match[3]-0.304200match[l] = 0.668520match[2]-0.504580malch!3] = 0.335400match!!] = 0.473960match[2] = 0.371940matchp]-0.249600match[!]-0.077870match!2] = 0.078360match!3]-0.054600match[l] = 0.035!00match[2]-0.039000match[3]-0.046800match!!]-O.58116Omatch!2] = O.45IOOOmatchi3J = 0.370500match[l]-0.321530match[2]-0.259!80match[3] = 0.! 75500match!!] = 0.194820match!2]-0.118780matchp] = 0.105300match[l]-0.203030malch[2]-0.110620malch!3] = 0.109200match[l] = 0.667099match[2] = 0.587499match!3] - 0.382200match!l] = O.23715Omatch[2]-0.160260matchp]-0.! 20900match!!]-0.094890match!2]-0.086160matchp] = 0.035100match[l]-0.866849match[2] = 0.691039match!3]-0.643499match!l]-0.3!7810match[2] = 0.238980matchp]-0.163800match!!] = 0.464690match!2]-0.32l240matchp]-0.323700

S. F. Liu et a].

Lol_2

match[l] = 0.!01400match[2]=0.! 20900match[3] = 0.070200match!!] = 0.128080match!2]-O.I436OOmatchP] = O.!O53OOmatch[l]-0.071000match[2]-0.049640match!3]-0,042900match!l]-0.131490match[2]-0,! 26580matchp] = 0.089700match[!] = O.!6!55Omatch!2] = 0.10!760malchp]-0.109200match[l]-0.098300match[2]-0.088640match!3] = 0,074100match!!]-0,168730match[2] = 0.!47500matchp] = 0,081900match[!] = 0.1!2970match[2]-0.!25I60match!3] = 0.081900match!l] = O.I47090match[2]-0.!01760matchp]-0.097500match[!] = 0.134820match!2] = 0.088640match[3] - 0,089700match[l]-0.047730match[2] = 0.040420match!3]-0.023400match[l] = O.O585OOniatdi[2]-0.074460match[3]-0,066300match!!] = 0,087660match!2] - 0.089000m:ilchi3] -- 0,042900match[l] = 0.205820match[2] = 0.!5[400matchpi-0,093600niatch!!]-0,143060match[2]-0.099640matchp] = 0.054600match[!] = O.! 55250match!2]-0,117720match!3]-O,113!OOmatch[l] = 0.093470match[2] = 0.084740match!31 = 0.085800matc!i!!l-0,124800malch[2) = O,!13!OOmatch[3] = 0.! 24800match[!] = 0,292420malch!2]-0,236500matchp]-0,! 44300

Lot_3

match!!] = 0.!2! 700matt;h[2]-0.086160match[3] = 0,070200match[!]-0,068!60match[2]-0,053540maEch!3] = 0.050700match!l] =0.047730match[2]= 0,036520matchp] = 0,023400match!!]-O,!78O3Omatch!2]-0.!!7720match!3] = 0.! 20900match[l] =0.066300match[2] = 0,081900match!3]-0.058500match!!]-0.085670match[2]-0.067720matchp] = 0.062400match[l]-0,097320match!2)-O,08758Omatch!3] - 0.050700match!l]-O.O5!63Omatch[2]-0.070560match[3]= 0,054600match[!]-0,2l8970match!2]-0.! 66660matchp]-0.!! 7000match[l] = 0,120640match!2] - 0,097860match!3] = 0.085800match[l] = 0.090630match[2]= 0,075520matchp] =0,081900match!!] = 0.079860match[2] = 0.073040matcht3] = 0,042900match!!]-O,!I718Omatch!2] = 0.060280iTialch!3] = 0.046«00match!l] = 0.077380matchi2] - 0.053540matchp]-0.031200match[!] =0.128520match!2] = 0.!!2400matchI3] = 0.093600match[l]=O.2!53OOmatch[2]-0,209540match[3] = 0.! 56000match!!] = 0.060360malch[2] = 0.051060matchp] = 0,027300match[!]-0,059430match!2] = 0.062760match!3] = 0.058500malch[l]-0.036030match[2] = 0,020920match!3]-0.03! 200

(continued opposite)

WBM reeognition using a neural network approach 2221

Map

20

21

22

23

24

25

Lot_l

match[l] = O.IO995Omatch[2] - 0.095740matchl3] = 0.050700match(ll = 0.168370match[2] = 0.138280match[3] - 0.097500maldijll-0.051630match[21=^0.0719K0matchi3i-0.042900malch[l]-0.595720match[2I-0.536480match(3] = 0.315900malchj 11 = 0.249420match|2]-O.2IO28Omatch[3]-0.097500malchil]-0.287950match[2]-0.23ll80iTiatch[31 = 0.187200

Lol_2

match[I] = O.I3872Omatchi2]-0.I08l40niatch[31-0.085800matchin-0.231080match[2] = 0.145020match[3I-O.I 17000match(l] = 0.219430match[2] = 0.152460malch|3] = 0.124800matchlll-0,103260malch[2] = 0.128360match[3]^ 0.078000match[l] = 0.390720match[2] = 0.256000matchl3]-0.198900tnalchll] = 0.130920match[2] - 0.063820malch[3]-0.085800

Loi_3

matchfl]-0.055530match[21 = 0.045740malch[3] = 0.046800miili;h[l]-0.076890match[2]-0.060280matchi?] =0.031200niatc-h[l] = 0.095020malch[2] ^ 0.062400match[3] - 0.058500malch[l]-0.220800maldi[2] = O.I 06020malch[3]-0.089700matdi[l]-O.I3965Omaldi[2] = 0,114880maidi[3] = 0-085800maldi[l]^O.I2418Omaldi[2] = O.I 13460nialch[3] = 0.085800

Table 2. Testing results of 75 WBMs.

l l I I I• I I I

• I I

map no

• • • • •

! • ! ! ! ! .

II11

II

-t-t-t-t-t-l-t-l

n n n IT

map no. 6

I!'!!

I I11III

• M l I• I l l l l l l l l l l l

map no. 23

Figure 10. Examples of ring-type patterns with a mouse bite.

2222 S. F. Uu et al.

interest. Two of the match values (match [1] and match [2]) from maps 2. 6. 14 and23 are greater than the threshold value. According to the predetermined rule, if twoor more match values are greater than the threshold value, the tested sample will beclassified as having an unknown pattern. This does not agree with the preknownanswers provided by the manufacturing engineers. The algorithm failed to correctlyrecognize the four maps. The recognition rate can therefore be claimed to be about95% (71/75) at this moment.

To improve the recognition rate, the aforementioned four maps (maps 2, 6, 14and 23) from Lot_I deserve further analysis. The systematic pattern of map 2 is aring type pattern with a mouse bite on the top of the ring, while the patterns of maps6 and 23 have a mouse bite on the left and right sections of the ring (figure 10). Aring-like pattern is a common characteristic of these maps and the known ring typepatterns. Consequently, the test result shows that these maps are similar to the tworing type patterns. To increase the recognition capability of these pattern types, thesethree maps should be trained into the network to become a new pattern.

In summary, the approach presented in this paper helps skilled manufacturingengineers to screen out the systematic patterns. The recognition rate can be up to95%, which can further be improved by training with more data in the future. Thisapproach also successfully replaces the human recognition and improves the pro-cessing speed. In real situations, there may be plenty of possible situations in whichsystematic patterns will occur. Although the developed system can only recognizethree types of patterns due to the limited number of available samples, the ARTlnetwork can be retrained whenever new patterns are presented to the system. Thisflexibility makes it a valuable tool in wafer bin map recognition.

5. ConclusionsIn view of increasing demands in the area of quality improvement and capability

enhancement in semiconductor failure analysis, this research developed an auto-mated recognition system for wafer bin maps collected from circuit probe tests.The reason this area was chosen was because a fast and correct recognition of binmaps is an essential step in maintaining or improving product yield. This systemfeatures a modular structure and incorporates the Adaptive Resonance TheoryNetworkl (ARTl) technique. It has accomplished the initial requirements of ourexpectations. It provides not only automatic recognition of the known patterns butcan also detect potential and unknown patterns. Actual data obtained from a semi-conductor manufacturing company were used to train and test this algorithm. Therecognition rate was about 95%. It is believe that this rate can be improved aftermore sample maps have been trained into the network. A possible extension of thisresearch is the incorporation of defect knowledge to determine the possible causes inthe manufacturing processes.

AcknowledgementThis work was supported by the National Science Council of the Republic of

China under Grant NSC90-8218-E-007-0I0.

ReferencesCHEN, F . L. and Liu, S. F., 2000. A neural-network approach to recognize defect spatial

patterns in semiconductor fabrication. IEEE Transactions on SemicomiuclorManufacturing. 13, 366-373.

WBM reeognition using a neural network approach

Co\ , 1, and REVNOLPS, A., 1996, Manufacturing Digital's Alpha AXP: Rapid Application.sDevelopment to Assure Critical Success Through Quality (SAS Co.).

FRILDMAN, D. J,,HANS1IN, M . H . , N A [ R , V. N. and JAMI:S, D . A., 1997, Model-free estimation

of defect clustering in integrated circuit fabrication. IEEE Transactions onSemiconductor Manufacturing. 10, 344-359.

KAIMI ' I , U , . 1995, The binominal test: a simple tool to identify process problems. IEEETransactions on Scmicomluctor Manujaclwing. 8. 160-166.

LIN, G . T . , 1998, Patterned recognition of wafer bin maps. Master thesis, Taiwan University., A. L, DoNOGHUK, G. O., DRAKE, A. W. and GRAVES, S. C , 1995. Spatial yield

modeling for semiconductor wafers. Proceedings of the IEEE/SEMI AdvancedSemiconductor Manufacturing Conference, pp. 276-281,VA, A. S. and MA( v. R, B,.1996, Pattern Reeognition with Neural Networks in C++(IEEE Press),. W. and HAMADA, M , , 1993, Detecting spatial eifects from factorial experiments: anapplication from integrated-circuit manufacturing, Technometries. 35. 149-160.