Wear 305 (2013) 14–22

Contents lists available at ScienceDirect

Wear

0043-16http://d

n CorrE-m

imin.ka1 Pr

Ohio St

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

Experimental study and modeling of the effect of mixed size abrasivegrits on surface topology and removal rate in wafer lapping

Xiaohai Zhu a,1, Chunhui Chung b, Chad S. Korach a,n, Imin Kao a,n

a Department of Mechanical Engineering, Stony Brook University, Stony Brook 11794-2300, NY, USAb Department of Mechanical Engineering, National Taiwan University of Science and Technology, Taipei 106, Taiwan, ROC

a r t i c l e i n f o

Article history:Received 14 December 2012Received in revised form4 May 2013Accepted 20 May 2013Available online 30 May 2013

Keywords:LappingMixed size abrasivesSurface featuresImage analysis

48/$ - see front matter & 2013 Elsevier B.V. Ax.doi.org/10.1016/j.wear.2013.05.007

esponding authors. Tel.: +1 631 632 1752; faxail addresses: chad.korach@stonybrook.edu (Co@stonybrook.edu (I. Kao).esent address: Department of Mechanical andate University, Columbus 43210, OH, USA.

a b s t r a c t

The semiconductor industry has continued to increase the diameter of wafers in recent years, whichposes a challenge in the lapping of prime wafers as the processing time is proportional to the square ofthe diameter, and the surface quality is a function of the features generated in the lapping process. As afree abrasive machining (FAM) process, where the abrasive grits act as third-body particles, lapping isinfluenced by abrasive size distributions; however, past studies focus on a single abrasive or sizedistribution, where the effects of mixed size abrasive distributions on surface feature generation are stillunknown. In this study, lapping experiments are conducted on silicon by mixing two SiC abrasive grits,with different mean sizes and at various ratios, under two normal loadings. Lapped surfaces areexamined by optical microscopy, where the number and size of critical surface feature types arecharacterized quantitatively with image processing. The results are correlated with the material removalrate (MRR) by modeling a lapping quality index (LQI) to evaluate different mixed abrasive ratios, where itis shown that lapping performance can be improved by mixing abrasives at high loadings.

& 2013 Elsevier B.V. All rights reserved.

1. Introduction

The rapid development of the semiconductor industry presentsa need for increasing the diameter of wafers [1,2]. This trendmakes it more challenging to achieve good surface quality ofwafers with high manufacturing efficiency. In order to controldefects, it is important to analyze the microscopic surface topologyin manufacturing processes. Bullis [3] reviewed various kinds ofsilicon defects. Young et al. [4] investigated the surface features ofsilicon wafers under various parameters in lapping and grinding.Much research has been conducted on the methodology ofautomatic recognition based on feature characterization. Udupaet al. [5] studied surface topology for unpolished wafers andproposed to use shearography to detect several kinds of featuressuch as swirl shapes and groups of particles. Yuan et al. [6]employed Bayesian inference procedure for parametric patternrecognition. Hwang and Kuo [7] and Yuan and Kuo [8] proposedmodel-based clustering approaches to simultaneously identifydefect clusters and their spatial patterns on the wafer. Chen andLiu [9] used neural networks for spatial defects pattern

ll rights reserved.

: +1 631 632 8544..S. Korach),

Aerospace Engineering, The

recognition. As a popular method of characterization, fuzzy clus-tering algorithms were also studied [10–12]. However, there haslacked research analyzing the results statistically for differentsettings of manufacturing parameters.

Parameters of the lapping process have been widely studied[13–15], but the impact of abrasive size distributions was rarelyinvolved. Bhagavat et al. [16] first showed that the mixed abrasivesresult in higher material removal rate (MRR) than the single-sizedabrasives. Chung et al. [17] further identified that maximumremoval rate is achieved at 1:1 mixing ratio of two different sizesof abrasive grits. However, the papers only presented the impacton average surface roughness, without statistical results of micro-scopic topology.

In this study, two different sizes of silicon carbide abrasives, F-400 and F-600, were mixed at five different mixing ratios, with theratio of the abrasives to the carrier fluid (de-ionized water) keptthe same, and two different loadings, 2.3 kg and 4.1 kg, wereapplied separately for each mixing ratio. Still photos of lappedwafer surfaces at each mixing ratio and loading were taken usingan optical microscope. Image processing was employed to detectthe numbers and sizes of several types of surface features that mayresult in defects. A model integrating surface topology andremoval rate was utilized to evaluate the performance in differentsettings of mixing ratios and loadings. The result can provide agood reference for the parameter optimization of the free abrasivemachining (FAM) process.

Fig. 1. (a) Flow chart of the experimental procedure and (b) schematic of the lapping process used.

Table 1FEPA grading chart of F-400 and F-600 SiC powders (μm).

SiC powder D3% D50% D94%

FEPA F-400 32 17.371.5 8FEPA F-600 19 9.371 3

X. Zhu et al. / Wear 305 (2013) 14–22 15

2. Experimental setup of lapping and macroscopic results

2.1. Experimental setup

Lapping experiments were carried out on (111) silicon waferswith a diameter of 76 mm and a Logitech PM5 one-sided lappingmachine was employed (see Fig. 1). A grooved cast iron plate wasused as the lapping plate. The wafers were wax-mounted ontoglass plates, which in turn were vacuum-chucked to a LogitechPP6GT lapping jig. The jig also provided a constant normal load onthe wafers during lapping. The dial gauge mounted on the jigmeasured the material removal depth in real time.

Two different grades of silicon carbide abrasives, F-400 andF-600, as shown by Table 1, were used in the experiments. Themedian sizes of F-400 and F-600 are 17.3 μm and 9.3 μm, respec-tively. Five different ratios of the weight of F-400 powder, W400, tothe total weight of abrasives, Wtotal, were employed: W400/Wtotal¼0, 0.25, 0.5, 0.75 and 1. The ratio of the total weight ofabrasives to the weight of de-ionized carrier fluid, C, was kept at aconstant value of 0.154. Two different normal loads, 2.3 kg and4.1 kg, were applied on the jig. Overall, there were 10 differentsettings with five mixing ratios and two loads.

The experiment of each setting lasted for 30 min. The angularvelocity of the cast iron lapping plate was kept at 70 RPM. Thedepth of material removal was recorded without interrupting theprocess every 5 min. After lapping, the wafer was cleaned by de-ionized water, and the wax was melted to remove silicon waferfrom the glass plate. The average Root-Mean-Square (RMS) surfaceroughness after lapping was measured by an XP2 profilometerwith a diamond probe at eight randomly selected locations. Fig. 1

shows the flow chart of the experimental procedure and schematicof the lapping process.

2.2. Summary of macroscopic results

The material removal depth, as shown in Fig. 2(a), is nearlylinear with respect of time. The average MRR is defined as the totalremoval depth divided by the operation time of 30 min. Theaverage MRR and RMS surface roughness at different mixing ratiosand under different loadings are listed in Table 2. Fig. 2(b) showsthat higher MRR comes with the consequence of higher surfaceroughness under the same loading. W400/Wtotal¼0.5 with the half–half mixed abrasive slurry has the highest MRR and surfaceroughness under the same loading, and W400/Wtotal¼0 with onlythe small grits has the lowest. The result applies to both 2.3 kg and4.1 kg loadings [17].

3. Analysis of surface topology

In order to statistically study the microscopic surface topology,we present the methodology for the image capturing of lappedwafer surfaces and the characterization with image processingtechniques. Images of lapped surfaces (see Fig. 3) were taken usinga digital optical microscope (Keyence VHX-500) with magnifica-tion of 500� . Images were taken for each wafer along twoperpendicular lines from center to edge of the wafer as shown inFig. 3(b). Typical images before and after mixed-abrasive lappingare shown in Fig. 3(a) and (c), respectively. There were approxi-mately 175 images taken along each line. Only the first 150 imageswere analyzed because the images taken near the edge had poorerimage quality.

The topology of lapped wafers shows typical surface featuressuch as cracks, indentations, and scratches under the microscope.Based on characteristics of different feature types, the ones thatmay exacerbate into defects after lapping are identified as targetfeatures by image analysis.

0

20

40

60

80

100

120

Mat

eria

l R

emoval

Dep

th (�m

)

Time (minutes)

Material Removal Depth under 2.3 kg

Loading

W400/Wtotal=0.00W400/Wtotal=0.25W400/Wtotal=0.50W400/Wtotal=0.75W400/Wtotal=1.00

0

0.5

1

1.5

2

2.5

3

3.5

0 5 10 15 20 25 30 0 0.25 0.5 0.75 1

W400/Wtotal

Comparison of MRR and RMS Surface

Roughness under 2.3 kg Loading

RMS roughness (μm)

MRR (μm/min)

Fig. 2. (a) Material removal depth and (b) material removal rate (MRR) versus RMS surface roughness at different mixing ratios and under 2.3 kg loading [17].

Table 2The average material removal rate (MRR; μm/min) and RMS surface roughness (μm)at different mixing ratios and under different loadings [17].

W400/Wtotal¼0 0.25 0.5 0.75 1

2.3 kgMRR 1.072 2.544 3.317 2.256 2.300Surface roughness 0.6313 0.9688 1.0646 0.7688 0.8318

4.1 kgMRR 2.072 5.133 5.422 4.950 5.028Surface roughness 0.5771 0.9729 0.9792 0.8688 0.8458

X. Zhu et al. / Wear 305 (2013) 14–2216

3.1. Features and characterization

The original microscopy images, as shown in Fig. 4((a), top), havedifferent background illuminations. When the grayscale was used inthe classification of feature types, this caused inconsistency of thesame feature for different images. To circumvent this the imagecontrast was adjusted at first, as shown in Fig. 4(b), to provideuniformity of the features for analysis. Three typical feature typesare identified by image analysis and indicated in Fig. 4(b):

(I)

High negative contrast features in the images that are likely tocause defects.

(II)

High negative contrast features which include a positivecontrast at the periphery, typically on one edge. In this case,the negative and positive contrast regions are merged to asingle target feature.

(III)

Small, positive contrast features that are isolated from thenegative contrast features are irregularities and excludedfrom detection, and may be due to surface contaminations.

It is important to note that the wafers were cleaned prior tomicroscopy, though small surface contamination may be present.Though chemical characterization of the surfaces were not per-formed, the small, positive contrast features observed presentsimilar to SiC particles when imaged with light microscopy. Also,the formation of possible chemical compounds, which wouldlikely be SiO2, is unlikely due to the large particle sizes used inthe wafer lapping process, leading to high material removal ratesthat would inhibit oxide formation. Only feature types I and IIwere targeted during image analysis.

3.2. Methodology of image processing

Based on the characterization, image processing was used todetect the numbers and sizes of target features (Type I and II). Dueto the interest in large target features that may develop into defects,

small features were excluded by designating a size threshold duringthe image analysis. Utilizing a 6 pixel size (equivalent to 24 μm2), thesmall features were filtered from the images. All the steps wereaccomplished using the Image Processing Toolbox of Matlab [18],and are outlined in Fig. 5. Details of the methodology for imageprocessing can be found in the Appendices. The images with all thetarget features detected are shown in Fig. 4((a), bottom).

4. Results of surface features and discussion

In this section, statistical analysis of surface topology is con-ducted to gain an insight into the influence of mixing abrasive gritson surface quality. At the same time, lapping efficiency is con-sidered by involving MRR in the modeling of an evaluation index.All detected features 4 24 μm2 were categorized into size ranges.The image processing results indicated that most features wereo100 μm2; therefore, the size range of 24–100 μm2 was dividedinto 3 equal ranges to categorize the detected features. Table 3 liststhe numbers of target features in each size range at five mixingratios, and under two loadings. The results are plotted in Fig. 6 foreach load separately.

Observing the results in Table 3 and Fig. 6, there is nosignificant difference found in the feature counts between theloads of 2.3 and 4.1 kg, except there is consistently lower featurecounts for the 4.1 kg load for the W400¼0 case. Also, there is adecrease in the feature counts at W400/Wtotal¼0.75 for the 4.1 kgload, and is most pronounced at the large feature area sizes.Though surprising that there is only small differences in thefeature counts for the two load levels, due to nearly twice theapplied load, the trend is consistent with previously reportedsurface roughness data (see Fig. 2 and Table 2). The trend can beexplained by our mixed abrasive surface roughness model [17]where larger loads lead to more particles in contact and hence alower load per particle. The lower load results in a decreasedparticle penetration depth into the substrate, which is equated tofeature formation. The case of W400¼0 is actually lower for thehigher 4.1 kg load, because it does not include any large particlesthat cause deep penetrations into the surface and creates an evendistribution of F-600 particles on the surface. The penetration ofthe SiC grits may create brittle fracturing in the Si wafer, and withthe subsequent particle motion can cause feature formation andmaterial removal.

When the larger F-400 particles are included in the slurry, inany percentage, the feature count increases significantly for bothloads (see Fig. 7). Though, the trend demonstrates that fordecreasing percentages of large particles there is a decrease inthe feature counts. This is explained by the lower percentage of

Fig. 3. (a) Image of a (111) Si wafer surface prior to mixed abrasive lapping. Surfaces were lapped with alumina abrasives and etched by vendor; (b) images were taken alongtwo perpendicular lines from center to edge of the wafer and (c) a typical image of wafer surface after mixed-abrasive lapping, with scale bar indicating 100 pixels as 20 μm.

X. Zhu et al. / Wear 305 (2013) 14–22 17

large particles and in-fill of sufficient small particles to balance thenormal load, though with a resulting increase in particle contactswhich decreases the load/particle and thus reducing feature like-lihood. For the 25% case, there is an increase in the feature counts,which goes against this argument. The increase is attributed tosporadic large particles present (from the upper-end of the particledistribution), creating more features.

One trend to be noted from the data is for the higher load(4.1 kg) case there was a decrease in feature counts between themixed cases which became more pronounced for larger featuresizes (see Fig. 6). We can conclude that there are less large featurespresent for the higher load, thus the smaller F-600 particles aresensitive to the load. This trend is most significant for larger defectsizes and for higher loads.

Examining the two extreme cases, W400¼0 (W600/Wtotal¼1)and W600¼0 (W400/Wtotal¼1) for the smallest (Ao50 μm2) andlargest (A4100 μm2) feature sizes, shows significant differencesfor the two applied loads (see Fig. 7). Examining the feature countsbetween mixing ratio W400/Wtotal¼0 and 1 show a decrease infeatures for increased feature size. In the 2.3 kg case, for featureareas 24 μm2oAo 50 μm2 a 75% ratio exists between the mixingratios 0 and 1, but for A4100 μm2 a 66% ratio exists. Likewise inthe 4.1 kg case, 57% and 34% ratios exist, respectively. This trend isexpected, since the smaller particles should not generate as deepfeatures as the larger particles, and this will be indicated in thefeature counts as related to feature area size. For the higher loadcase, there is a larger difference between the small and largefeature counts. The ratio of counts between mixing ratios forA4100 μm2 to counts of 24 μm2oAo50 μm2 is only 60% (i.e.34:57), compared with a ratio of 86% in the low load case. This canbe explained by the fact that at the higher load, an increasednumber of active particles will be statistically in contact with thesubstrate, leading to a distribution of the load over more particlesand hence a decrease in particle penetration depth. The dataindicates that the smaller particles are more affected by thisphenomenon (see W400¼0 in Fig. 7), where the larger particle

sizes follow a trend of increased overall depth and featureformation. Thus, there are two contrary effects: (1) that whenload increases, the load per particle decreases, decreasing thecontact depth such that the feature number decreases, and(2) when the load increases, the depth increases and then thefeature number increases. The smaller F-600 particles are moreaffected by the former one than the large particles.

The phenomena that for mixing ratios withW400¼0 the featurenumbers in all size ranges of the low (2.3 kg) loading case arelarger from those of the higher loading case, while for othermixing ratios (40) the feature numbers are nearly the same, maybe due to the aforementioned particle size effect. Since the mixingratio with the W400¼0 case only has the small F-600 particles, thiscase is influenced by the first effect, that is, feature numberdecreases with increased load due to a lower per particle loadfrom more particle contacts. Modeling of the mixed abrasiveparticle contact indicates an increase in penetration depth forboth (F-400 and F-600) particle sizes, though the increase in depthfor the larger particles is nearly 50% more than the small particles[17]. The results of the model provide insight to this observedtrend in feature formation. Results from roughness measurementsby profilometry show that roughness data for the mixing ratiosbetween 0 and 1 have larger RMS roughness values. Compared tothe feature count measurements, feature numbers are the lowestfor the mixed cases. Similar to the roughness results, MRR washighest for the mixed cases. Previous modeling [17] showed thatfor the mixed cases, the roughness and MRR increase due to theconfluence of the F-400 and F-600 particle distributions. What isintriguing is that for the highest MRR and roughness cases, thefeature sizes are the lowest. To quantify this phenomena, a lappingquality index (LQI) is developed, which is the ratio of the totalvolume of material removed divided by the surface featurevolumes, and designated by L. A large LQI would provide assess-ment of a relatively higher quality lapping process.

The total volume of material removed is calculated from theproduct of the MRR for each case, and the lapping time and wafer

Fig. 4. (a) Original microscopy images of lapped wafer surface (top), and images with all the target features detected (bottom) and(b) grayscale images after contrastadjustment, in which I, II and III mark three typical features of the lapped wafer surface.

X. Zhu et al. / Wear 305 (2013) 14–2218

area. The lapping time for the average MRR values reported inTable 2 is 30 min; and the wafer area is a constant circular areavalue (¼πr2) for a 76 mm diameter wafer. The removal volume (VR)is given as

VR ¼MRR AWt ð1Þ

where AW is the total wafer area, and t is the lapping experimenttime. The surface feature volume (VF) is calculated from theproduct of the total feature area and the measured RMS surfaceroughness average (Rq). The total feature area (AF) is computed asthe sum of the total feature count for a case multiplied by thefeature area size (AFs). Since the feature area sizes were measured

Fig. 5. Flow chart of the target feature detection method.

Table 3Number of target features in each size range (μm2) at different mixing ratios andunder different loadings of (a) 2.3 kg and (b) 4.1 kg.

(μm2) W400/Wtotal¼0 0.25 0.50 0.75 1.00

(a) 3 kg Loading24≤Areao50 2567 3455 3302 3316 342350≤Areao75 462 721 731 701 72775≤Areao100 209 332 292 332 302Area≥100 241 379 374 371 367

(b) 4.1 kg Loading24≤Areao50 1994 3499 3299 3278 351250≤Areao75 320 699 729 618 74075≤Areao100 146 311 301 267 347Area≥100 132 406 283 273 389

Fig. 6. Number and size of target features at different abrasive mixing ratios for(a) 2.3 kg loading and (b) 4.1 kg loading.

X. Zhu et al. / Wear 305 (2013) 14–22 19

in ranges, the average of the range is taken for the feature area sizerange, e.g. for the feature area size range of 50 μm2 oAo75 μm2,the average size of 62.5 μm2 is used for the area; for the otherfeature area size ranges, this same protocol was followed. For thecase of A4100 μm2, an area of 100 μm2 was used. The total featurearea is determined as

AF ¼∑Feature Size RangeNFAFs ð2Þ

where NF and AFs are the feature count (from Table 3) and featurearea size, respectively, for a given range. The total feature area isplotted in Fig. 8 as a function of the 10 studied cases. The totalfeature area for both of the loading cases present similar results inmagnitude and trend, with a low total feature area measured forthe case of entirely small F-600 particles (W400¼0). The mixedcases had similar values for both the applied loads and the casewith entirely large F-400 particles (W400/Wtotal¼1). The results areconsistent with the observations in feature counts, where anyaddition of large F-400 particles increased the number of observedsurface features. Removal volume (VR), feature volume (VF), andthe LQI for each case are summarized in Table 4, where the lapping

Fig. 8. Computed total feature area for abrasive mixing ratios W400/Wtotal¼0–1 for2.3 kg loading and 4.1 kg loading.

Table 4Lapping quality index (LQI�106 μm2), removed volume (VR�1011 μm2), andfeature volume (VF�105 μm2) for different mixing ratios and under different loadlevels, 2.3 kg and 4.1 kg.

W400/Wtotal

Load

2.3 kg 4.1 kg

VR VF LQI VR VF LQI

1 3.15 1.96 1.61 6.88 2.08 3.310.75 3.09 1.79 1.73 6.77 1.83 3.700.5 4.54 2.46 1.85 7.42 2.18 3.410.25 3.48 2.32 1.50 7.02 2.34 3.000 1.47 1.05 1.40 2.83 0.69 4.10

Fig. 9. Lapping quality index (LQI) for abrasive mixing ratios W400/Wtotal¼0–1 for2.3 kg loading and 4.1 kg loading.

Fig. 7. Number of smallest and largest target features at abrasive mixing ratiosW400/Wtotal¼0 and 1 for 2.3 kg loading and 4.1 kg loading.

X. Zhu et al. / Wear 305 (2013) 14–2220

index is computed finally as

L¼ VR

VF¼ MRR AWt

AF Rqð3Þ

Though the surface feature area is an important predictor of thesurface quality, the LQI can provide a quantitative non-dimensionalparameter that addresses the performance of the lapping process bycomparing aspects of the material removal to the surface featurevolume. It is expected that processes that lead to different MRR mayhave quite different surface feature formation. This can lead to similarLQI values for two very different processes, e.g., a lowMRR/low featurevolume process versus a high MRR/high feature volume process. Thevalues of the LQI are plotted in Fig. 9 for each of the mixed abrasive

cases and the two load levels. Not surprisingly, the LQI for the lowload cases is significantly lower than the higher load cases. Thisindicates that a similar feature volume is generated for each of theload levels, and is consistent with the earlier discussion regardingsimilar feature counts for the two load levels. The result providesevidence that a high quality surface can be generated while main-taining a high MRR from the increased load. A significantly differenttrend exists between the two load levels, though, for the location ofmaximum LQI. For the low load level, a maximum of the LQI is foundfor the mixed abrasive case of W400/Wtotal¼0.5. For this mixedabrasive case an optimum balance between the MRR and the featurevolume is formed, and an improvement in performance is achievedby using the mixed abrasives. The high load level exhibits aninconsistent trend in the LQI, with maximums at both W400/Wtotal¼0and 0.75. The high LQI for the W400¼0 case is not surprising, sincethis lapping case exhibited the least number of feature counts (seeTable 3 and Fig. 6). The unique trend ultimately is determined by thefact that the feature counts for the high load level drop significantlyfor the W400/Wtotal¼0.5 and 0.75 cases (see Fig. 6(b)). These resultsdemonstrate that though the addition of large particles increases thefeature volume, the increased MRR balances the feature formation tocreate similar LQI values for the high load level cases of W400/Wtotal¼0 and 0.75. Thus, the trade-off between MRRs can be equatedby using a mixed abrasive ratio of 0.75 at the high load level. Animportant sensitivity to load has been identified for mixed abrasivegrit lapping that provides evidence that for higher loads the benefitsof mixed abrasives are less significant than for lower loads, but mixedabrasives can balance the feature formation with a high MRR, if thisis desired. The results presented here on free abrasive machining canbe extended to other processes such as wire saw cutting.

5. Conclusions

This paper addresses the surface feature count and size character-ization and analysis for free abrasive machining (FAM) using mixedabrasive distributions. A methodology for characterizing the featureson silicon wafers lapped with mixed abrasives of two sizes based onimage contrast was developed, and applied to characterization of fivedifferent abrasive mixing ratios and two load levels. Feature countswere found to be similar for the two load levels, though previousresults have observed higher material removal for higher loads. Thisresult demonstrates the feature formation was minimally dependenton the material removal rate (MRR), and hence the applied load. Alapping quality index (LQI) was generated from the feature count data,and data on the surface roughness and MRR for each case. The LQI wasfound to increase for the higher load due to the increased MRR. A cleartrend was observed for the 2.3 kg case, with the highest LQI at a 50%abrasive mixing ratio; though the 4.1 kg case had two maximums at

Fig. A.1. Grayscale images (a) converted from the original images and (b) after contrast adjustment, with their respective histograms of grayscales shown below.

Fig. B.1. Steps of feature detection after contrast adjustment. (a) Pixels of grayscale 0–20 or 230–255 are identified as white objects in binary images; (b) closely spacedobjects are merged; (c) only the objects containing pixels with maximum contrast gradient, i.e., feature (II), are retained and (d) regions with grayscale 0–20 are added tocompensate for the loss of feature (I).

X. Zhu et al. / Wear 305 (2013) 14–22 21

W400/Wtotal¼0 and 0.75. The result indicates that for higher loads, theintroduction of any larger abrasive particles to the slurry will decreasethe LQI by an increase in feature formation, where for the lower of thetwo loads smaller particles did not always equate to less featureformation, at least when compared with the roughness and MRRchanges. This is believed to be due to a more uniform particledistribution for high loads and smaller particles, compared with anincrease in spurious larger particle abrasion for mixed abrasive cases athigh loads. We conclude that lapping performance can be improvedwith mixed abrasives at high loads bymaintaining a surfacewithout anincrease in surface features which may lead to defects in the material.

Appendices

This section introduces in detail each step of the imageprocessing used in feature detection. MATLAB commands usedare given in italics.

A. Preprocessing

A.1. Conversion to grayscale imagesThe command “rgb2gray” converts the RGB values to grayscales

ranging from 0 to 255, with 0 as black, 255 as white and othervalues in between, as shown in Fig. A.1(a).

A.2. Contrast adjustmentThe command “imadjust” returns a linear mapping of the

middle subset of pixel values to the range throughout 0–255,leaving the lowest and highest 2% subsets clipped. The results areshown in Fig. A.1(b).

A.3. Shrinking of image sizesThe command “imresize” shrinks images by 0.1 with bicubic

interpoloation before following steps. This step promotes thespeed of batch processing, as it reduces an image to 120 by160 pixels.

B. Detection of target features

B.1. Detection of the negative and positive contrast regions in theimage

The command “bwareaopen” returns 0–20 and 230–255 aswhite objects (1's), and reduces noises by removing objectssmaller than 3 pixels. Results are shown in Fig. B.1(a).

B.2. Merging of adjoining negative and positive contrast regionsBoth using a disk-shaped element with 3-pixel radius, “imdi-

late” connects closely spaced objects, and then “imerode” smooths

X. Zhu et al. / Wear 305 (2013) 14–2222

the boundaries. Objects smaller than 24 mm2 (6 pixels) areremoved using “bwareaopen”. The results are shown in Fig. B.1(b).

B.3. Detection of edges between negative and positive contrast regionsUsing the Sobel operator [19,20], “edge” returns points with the

heuristically maximum contrast gradient.

B.4. Selection of target features using the detected edgesObjects not having the detected edges are filled black (0's), and

pixels within grayscale 0–20 are re-added by the “union” operation,as shown in Fig. B.1(c) and (d).

B.5. Calculation of feature number and sizeAfter images are enlarged to the original size, “bwlabel” returns

the number of target features while “regionprops” finds the size bypixels. The physical size is calculated on a scale of 25 pixels for1 mm2 (see the measurement reference bar in Fig. 3(c)).

References

[1] International Technology Roadmap for Semiconductors 2010 Update. ITRS(2010). Available from: ⟨http://www.itrs.net/Links/2010ITRS/Home2010.htm⟩.

[2] Intel, Samsung Electronics, TSMC Reach Agreement for 450 mm WaferManufacturing Transition, Intel Corp. News Release (2008). Available from:⟨http://www.intel.com/pressroom/archive/releases/2008/20080505corp.htm⟩.

[3] W.M. Bullis, Current trends in silicon defect technology, Materials Science andEngineering: B 72 (2–3) (2000) 93–98.

[4] H.T. Young, H.T. Liao, H.Y. Huang, Surface integrity of silicon wafers in ultraprecision machining, The International Journal of Advanced ManufacturingTechnology 29 (3) (2006) 372–378.

[5] G. Udupa, B.K.A. Ngoi, H.C.F. Goh, M.N. Yusoff, Defect detection in unpolishedSi wafers by digital shearography, Measurement Science and Technology 15(1) (2004) 35–43.

[6] T. Yuan, S. Bae, J. Park, Bayesian spatial defect pattern recognition insemiconductor fabrication using support vector clustering, The InternationalJournal of Advanced Manufacturing Technology 51 (5) (2010) 671–683.

[7] J.Y. Hwang, W. Kuo, Model-based clustering for integrated circuit yieldenhancement, European Journal of Operational Research 178 (1) (2007)143–153.

[8] T. Yuan, W. Kuo, Spatial defect pattern recognition on semiconductor wafersusing model-based clustering and Bayesian inference, European Journal ofOperational Research 190 (1) (2008) 228–240.

[9] F.L. Chen, S.F. Liu, A neural-network approach to recognize defect spatialpattern in semiconductor fabrication,, IEEE Transactions on SemiconductorManufacturing 13 (3) (2000) 366–373.

[10] Y.H. Man, I. Gath, Detection and separation of ring-shaped clusters using fuzzyclustering,, IEEE Transactions on Pattern Analysis and Machine Intelligence 16(8) (1994) 855–861.

[11] N.G. Shankar, Z.W. Zhong, Defect detection on semiconductor wafer surfaces,Microelectronic Engineering 77 (3–4) (2005) 337–346.

[12] C.H. Wang, W. Kuo, H. Bensmail, Detection and classification of defects patternon semiconductor wafers, IIE Transactions 38 (12) (2006) 1059–1068.

[13] M. Buijs, K.K.-V. Houten, Three-body abrasion of brittle materials as studied bylapping, Wear 166 (2) (1993) 237–245.

[14] U. Heisel, J. Avroutine, Process analysis for the evaluation of the surfaceformation and removal rate in lapping, CIRP Annals—Manufacturing Technology50 (1) (2001) 229–232.

[15] I.D. Marinescu, E. Uhlmann, T. Doi, Handbook of Lapping and Polishing, Taylor& Francis, Boca Raton, Florida, 2006.

[16] S. Bhagavat, J.C. Liberato, C. Chung, I. Kao, Effects of mixed abrasive grits inslurries on free abrasive machining (FAM) processes, International Journal ofMachine Tools and Manufacture 50 (9) (2010) 843–847.

[17] C. Chung, C.S. Korach, I. Kao, Experimental study and modeling of lappingusing abrasive grits with mixed sizes, Transactions of the ASME: Journal ofManufacturing Science and Engineering 133 (3) (2011) 031006-1-8, http://dx.doi.org/10.1115/1.4004137.

[18] The Mathworks Inc., Image Processing Toolbox User's Guide, Version 7.0,Natick, Massachusetts (2010).

[19] R.C. Gonzalez, R.E. Woods, Digital Image Processing, Addison Wesley, Boston,Massachusetts, 1992pp. 414–428.

[20] R.D. Boyle, R.C. Thomas, Computer Vision: A First Course, Blackwell ScientificPublications, Oxford, 1988pp. 48–50.