Towards efficient wafer visual inspection: Exploring novel lightweight approaches for anomaly detection and defect segmentation

Abstract

The rapid advancement of both wafer manufacturing and AI technologies is reshaping the semiconductor industry. As chip features become smaller and more intricate, the variety and complexity of defects continue to grow, making defect detection increasingly challenging. Meanwhile, AI has made significant strides in unsupervised anomaly detection and supervised defect segmentation, yet its application to wafer inspection remains underexplored. This work bridges these fields by investigating cutting-edge lightweight AI techniques for automated inspection of current generation of silicon wafers. Our study leverages a newly curated dataset comprising 1,055 images of 300 mm wafers, annotated with 6,861 defect labels across seven distinct types, along with PASS/FAIL decisions. From a data-centric perspective, we introduce a novel unsupervised dataset-splitting approach to ensure balanced representation of defect classes and image features. Using the DINO-ViT-S/8 model for feature extraction, our method achieves 96% coverage while maintaining the target 20% test ratio for both individual defects and PASS/FAIL classification. From a model-centric perspective, we benchmark several recent methods for unsupervised anomaly detection and supervised defect segmentation. For unsupervised anomaly detection, EfficientAD obtains the best performance for both pixel-level and image-wise metrics, with F1-scores of 75.14% and 82.35%, respectively. For supervised defect segmentation, UPerNet-Swin achieves the highest performance, with a pixel-level mDice of 47.90 and a mask-level F1-score of 57.45. To facilitate deployment in high-throughput conditions, we conduct a comparative analysis of computational efficiency. Finally, we explore a dual-stage output fusion approach that integrates the best-performing unsupervised anomaly detection and supervised segmentation models to refine PASS/FAIL decisions by incorporating defect severity.