WSDL: AI-Enabled Wafer Surface Defect Localization for Cleaning Process Optimization
In advanced semiconductor manufacturing, wafer surface integrity is strongly influenced by cleaning efficiency, particle contamination, and process variability. Conventional Automated Optical Inspection (AOI) systems primarily focus on defect detection but provide limited actionable feedback for cleaning optimization and yield learning. This work introduces UWSDI, an AI-enabled, class-agnostic framework that unifies wafer surface defect detection and pixel-level localization while delivering process-relevant insights for cleaning recipe tuning.
UWSDI integrates a frozen CNN feature extractor with a trainable Denoising Transformer Decoder that employs deterministic adapter routing and a segmentation-oriented anomaly head. Through cross-attention and recurrent self-attention, the model captures long-range spatial dependencies while preserving the periodic textures and circular symmetries inherent in patterned wafers. Unlike reconstruction-only approaches, UWSDI simultaneously outputs image-level anomaly scores and high-fidelity pixel-accurate masks, providing interpretable visual evidence—such as saliency maps—to support rapid root-cause analysis and contamination control.
We benchmark UWSDI on the Texture-AD (wafer AOI) dataset and real microscopy imagery, ensuring rigorous evaluation of both detection and localization performance. To enhance process relevance, we further assess UWSDI under simulated cleaning scenarios that emulate post-clean wafer conditions, including particle redistribution, residue patterns, and texture drift caused by process variability. Results demonstrate consistent localization performance under realistic fab-like conditions without the need for per-class retraining.
Compared with state-of-the-art unified models such as UniAD, UWSDI improves image- and pixel-level AUROC by 2–5% while maintaining stable inference latency and predictable computational cost. From a manufacturing perspective, UWSDI provides actionable spatial feedback that enables engineers to identify contamination hotspots, evaluate particle removal effectiveness, and refine cleaning recipes in a data-driven manner. By bridging AI-based anomaly detection with wafer surface conditioning, UWSDI offers a scalable, explainable, and operations-friendly solution for high-volume semiconductor manufacturing.
BIOGRAPHY
Leo Lin is a PhD candidate in the Artificial Intelligence Program at National Tsing Hua University and a visiting scholar at the University of Washington (2025–2026). He is also an IT Program Manager at Nanya Technology, where he leads AI, smart manufacturing, and automation projects in the semiconductor industry. His research focuses on computer vision, anomaly detection, vision-language learning, and AI for industrial applications. Leo has published work at CVPR 2025 Workshop, ICIP 2025, ECCV 2024 Workshop, and BMVC 2023. He also holds PMP, PMI-ACP, and PMI-PBA certifications and has received multiple awards for AI-driven innovation, project leadership, and teaching excellence.