Reconstruction-based Anomaly Detection with Machine Learning for High Throughput Scanning Electron Microscope Defect Inspection

With rapid advancement of high throughput wafer inspection systems, there are growing demands on more efficient and accurate defect inspection algorithms. This work focuses on investigating the feasibility of applying learning based algorithms to the image processing pipeline of the next generation wafer inspection tools. An end-to-end reconstruction-based anomaly detection methodology is proposed with two types of reconstruction models based on Principal Component Analysis and Convolutional Autoencoder. Both methods have demonstrated competitive performance compared to the current defect detection algorithm on multiple datasets with different patterns and various defect types.