CSAM anomaly detection with AI

Abstract:
This paper presents an industrial application of unsupervised autoencoders for the task of anomaly detection within CSAM (C-type Scanning Acoustic Microscopy) images to detect potential defects in the semiconductor manufacturing process as part of the work done within the ECSEL project iRel4.0. The approach presented leverages the high-yield nature of the semiconductor industry to generate a functional unsupervised autoencoder model with significantly shorter training time than supervised approaches to achieve comparable results. This approach has been validated with both synthetic wafers with manufactured defects as well as real production wafers. The approach extends current CSAM inspection methods by adding die-level resolution to potential defects as well as surpasses current manual optical inspection methods contributing to achieving Zero-Defect strategy in semiconductor manufacturing.