The Classification of Wafer Defects: An Evaluation of Different Feature-Based ResNet Transfer Learning Models with Support Vector Machine

Abstract

Wafer defect detection is a non-trivial issue in the semiconductor industry. Conventional means of defect detection are often labour-intensive based that is prone to error owing to a myriad of issue. Hence, there is push towards automatic defect detection in the industry. This work shall investigate the efficacy of a transfer learning pipeline that consists of a different pre-trained ResNet convolutional neural network models in which its fully connected layer is swapped with different support vector machine (SVM) models in classifying the defect state of a wafer whether it pass or fail. The optimal hyperparameters are identified via the grid-search technique. It was shown from the present investigation that the features extracted via the ResNet101v2 transfer learning model with a linear-based SVM model with a C and gamma parameter of 0.01, respectively, could yield a validation and test classification accuracy of 96% and 94%, respectively, on a stratified 60:20:20 data split ratio. The result from the present study demonstrates that the proposed pipeline is able to classify the defect level of the wafer well.