Note on CapsNet-Based Wafer Map Defect Pattern Classification

Abstract

Classification of wafer map defect patterns is important to monitor occurrence and further to assist root cause analysis of manufacturing-process-induced systematic defects. In this study we develop CapsNet-based wafer map defect pattern classifier. CapsNet is a variant of convolutional neural network, which extract features of images as vectors, not as scalars, and is expected to extract features more accurately under fluctuations of locations, angles, and scales of features in input images. Experimental results indicate that, by combining 2-stage (detector and classifier) approach, the proposed scheme shows higher accuracy on WM-811K real wafer map dataset for 8 categories in comparison to the previous work, on average and especially on the categories “Donut” and “Scratch,” which are difficult to accurately categorize by the previous work.