Third-party cyber evaluations involving OpenAI models
openai.com·14h ago
AI Research
arXiv CS.AI·Jun 9·5 min read
Failure-Aware Refinement of Vision-Language Model for Lithography Defect Detection
TL;DR
Semiconductor lithography inspection faces challenges in accurately detecting small pattern defects like bridges and contamination. A two-stage vision-language framework, utilizing Qwen3-VL fine-tuned with LoRA, was developed to enhance defect detection and prediction refinement.
✦ Why It Matters
Engineers can leverage this framework to enhance defect detection accuracy in semiconductor lithography processes.
Key Takeaways
- Implement Qwen3-VL for improved defect detection in lithography.
- Utilize LoRA for efficient model fine-tuning.
- Focus on both defect classification and localization.
- Adopt two-stage frameworks for enhanced prediction accuracy.
Full Summary
Semiconductor lithography is critical for manufacturing integrated circuits, and detecting small defects such as bridges, burrs, and contamination is essential for quality control. A two-stage vision-language framework was developed, where Qwen3-VL, a vision-language model, was fine-tuned using Low-Rank Adaptation (LoRA) to predict defect counts, categories, and normalized bounding boxes from lithography images.
The first stage focuses on initial defect detection, while the second stage refines these predictions for improved accuracy. Experimental results demonstrated significant improvements in defect classification and localization, although specific metrics were not disclosed.
This methodology allows for more reliable inspections, potentially reducing manufacturing errors and costs. The implications for engineers include enhanced defect detection capabilities and improved quality assurance processes in semiconductor manufacturing.