Wafer Failure Pattern Analysis Roger Jang jangmirlab org

Wafer Failure Pattern Analysis 張智星 (Roger Jang) jang@mirlab. org http: //mirlab. org/jang 清華大學 資訊系 MIR實驗室 2021/5/21 1

Types of Analysis z. Clustering y. Given failure patterns, find their natural grouping z. Ranking y. Given a wafer failure pattern, find similar patterns (returned in a ranking list) z. Classification y. Given a failure pattern, find patterns by its cause by using a set of training data -2 -

A Typical Wafer Failure Pattern -3 -

Flowchart for Clustering z. Steps for clustering wafer failure patterns y. Import failure patterns from external files y. Normalize the pattern if necessary y. Define a distance metrics (with rotationinvariance or not) y. Cluster wafer failure patterns x. K-means clustering x. Hierarchical clustering x. Fuzzy c-means clustering -4 -

Original Patterns -5 -

Binary Patterns -6 -

Round Binary Patterns -7 -

K-means Clustering z 20 trials of k-means clustering, with no. of clusters = 5 -8 -

Results of K-means Clustering z Use background colors to indicate clusters. -9 -

Hierarchical Clustering z. A typical result of hierarchical clustering -10 -

Ranking of Wafer Patterns z. Data at a glance: -11 -

Flowchart for Ranking z. Steps for ranking wafer failure patterns y. Import wafer patterns from external files y. Define a distance metrics (with rotationinvariance or not) y. Rank all wafer patterns based on a given pattern y. Plot the ranking results y. Adapt parameters to approximate human’s ranking (relevancy feedback) -12 -

All Wafer Patterns -13 -

Distance with Rotation-Invariance -14 -

Ranked Results w/o Rotation Query pattern No. 1 No. 2 No. 3 -15 -

Ranking Results with Rotation Query pattern No. 1 No. 2 No. 3 -16 -

Enhancement of Failure Patterns z. To further improve the ranking, we can enhance the failure patterns in various ways y. Apply median filter to eliminate salt-and-pepper noise (The results are shown in the next few slides. ) y. Extract objects/regions from connected components y. Use statistics of extracted objects as features for further processing -17 -

All Wafer Patterns after Median Filter -18 -

Ranked Results w/o Rotation Query pattern No. 1 No. 2 No. 3 -19 -

Ranked Results with Rotation Query pattern No. 1 No. 2 No. 3 -20 -

Classification of Wafer Failure Patterns Ring Line Sector Center Localized Edge Automatic classification -21 -

Features based on Radon Transform z CRT Radon transform: Edge Center Ring -22 -

Features Based on Region Statistics Flowchart Raw data Image segmentation Red indicates failure dies Noise reduction Image segmentation Each region filled with different colors Region property measurement -23 -

Operators for Region Identification z. Erosion (浸蝕) and Dilation (膨脹) -24 -

Available Classifiers z K-nearest-neighbor z Quadratic classifier z Naïve Bayes classifier z Sparse-representation classifier z Gaussian-mixturemodel classifier z Neural networks z SVM: Support vector machine z CART: Classification and regression tree z LVQ: Learning vector quantization -25 -

Other Enhancements z Distance metrics y. Hamming distance y. L-1 norm y. L-2 norm y… z ROI extraction y. Median filter y. Hough transformation y. Projection y… z Invariant Operators y. Rotation y. Radial translation y. Expand or Contract y… z Adaptation y. Learning to rank y. Artificial neural networks y… -26 -