Classification of Wafer Maps Defect Based on Deep Learning Methods With Small Amount of Data

Abstract

This paper attempts to solve the problem of defect classification for the purpose of automation of the processing of flaw detection results. Article proposes an algorithm based on deep convolutional neural networks (DCNN) for recognizing patterns of defects in semiconductor wafers. In order to train the model, a composite training data set was created and applied. Its basis consists of synthetic data and an extra small amount of experimental data including about 20 examples. Verification of the work was carried out on an open data set WM-811K. The resulting classification accuracy is about 87.8%. This is a satisfactory result from a practical point of view. The developed algorithms can be used both in software systems for data analysis in the production of semiconductor wafers, as well as part of separate software modules for electronic flaw detectors.