Explainable AI for automated quality inspection in semiconductor wafer fabrication

Abstract

This paper presents an explainable AI system for automated defect classification in semiconductor wafer fabrication. A Vision Transformer model classifies 38 defect types from scanning electron microscope images, while Grad-CAM and SHAP-based explanation modules highlight defect regions for process engineers. The system achieves 97.2% classification accuracy on a production dataset of 280,000 wafer images and has been deployed in a 300mm fab reducing manual inspection by 65%.

Topics

Full Text Preview

The Vision Transformer (ViT-B/16) was fine-tuned on 280,000 SEM images from a 300mm semiconductor fabrication facility, classified into 38 defect categories following SEMI E10 standards. Images were preprocessed to 224x224 patches with contrast-limited adaptive histogram equalization. Grad-CAM attention maps overlay defect regions on original images, while SHAP values provide feature-level explanations. The system integrates with the factory MES (Manufacturing Execution System) via SECS/GEM protocol.

Published Through