AI-Based Localization and Classification of Visual Anomalies on Semiconductor Devices

This paper presents an AI-based system for automated visual inspection of semiconductor components, aimed at improving the Zero-Defect strategy in their manufacturing process. The system leverages unsupervised learning using Variational Autoencoder to learn and compare images of undamaged components to identify anomalies. An anomaly score is devised to enable detection of even minor flaws on the edges of components and decision rules are evaluated using appropriate metrics. The proposed system surpasses the current tape machine in detecting anomalies, hence contributing to achieving the Zero-Defect strategy in semiconductor manufacturing.