Semiconductor Wafer Map Defect Classification: Wu et al. Features with Random Forest / SVM vs Convolutional Neural Network

After wafer probe testing, the map of failed dies on a wafer forms spatial patterns - Center, Donut, Edge-Loc, Edge-Ring, Loc, Near-full, Random, Scratch - that point to different process faults. On the real WM-811K data set (Wu, Jang and Chen, IEEE TSM 2015: 811,457 wafer bin maps of 46,293 lots, 172,950 labelled, 9 classes, 'none' : Near-full = 990 : 1) the project compares the hand-crafted features of Wu et al. (13 region failure densities, Radon-transform profiles and the geometry of the largest failure region) with Random Forest, RBF-kernel SVM and gradient boosting against small convolutional neural networks trained on a laptop CPU. All splits are by lot and every choice is made on validation lots; on 35,246 maps of 2,152 held-out test lots gradient boosting reaches macro-F1 0.8835 / accuracy 97.69 % (deployed), Random Forest 0.8303 / 96.56 %, RBF-SVM 0.8248 / 96.45 %, a position-keeping CNN 0.7651 / 96.61 % and the specification CNN only 0.6465 / 93.83 %. GNU Octave reproduces the features to 3.6e-14; a complete MATLAB version (written, not run) and a web application with five databases are included.

Category: Electronics and Communication Engineering (Semiconductor Manufacturing and Yield Analysis, Image Processing, Radon Transform, Hand-Crafted Features, Random Forest, Support Vector Machines, Gradient Boosting, Convolutional Neural Networks, Class Imbalance, GNU Octave, MATLAB, Python)

Electronics & Communication Engineering Project E11 from ProTechSC Academy, with graphical abstract, video demonstration and project guidance.