Accurately predicting manufacturing defects is critical for semiconductor companies to reduce costs and get new chips to market faster. This paper presents two key advancements to an existing AI/ML-based defect prediction platform that significantly improve its accuracy.
First, the researchers incorporated physical failure analysis (PFA) data to fine-tune the feature vectors used in the AI model. This allowed the system to better capture real-world defect patterns that may not be apparent from design layout data alone. Second, they developed a reinforcement learning approach to reduce false positive defect predictions, further enhancing the platform's reliability.
These enhancements make the AI/ML defect prediction system more robust and valuable for semiconductor companies. By catching more actual defects while reducing false alarms, chipmakers can accelerate their design-to-manufacturing cycles and improve overall yield. The process-aware, self-learning capabilities showcased here represent an important step forward in making AI-powered defect detection a cornerstone of advanced chip design and manufacturing.
This paper was presented at the 2025 SPIE Advanced Lithography + Patterning symposium.