LLM-based overlay issue classification and solution optimization in semiconductor manufacturing

As Moore’s Law decelerates and talent shortages threaten semiconductor innovation, the industry faces dual crises of technical scalability and institutional knowledge erosion. This paper presents a centralized AI framework integrating large language models (LLMs) and ontology-based machine learning to address yield optimization, process control, and knowledge fragmentation in high-volume manufacturing (HVM). By synthesizing virtual fabrication datasets (160M wafers) and deploying domain-specific AI agents (PRISM and INFER), our system achieves 0.2nm overlay prediction accuracy (R²=0.98) and reduces troubleshooting time from weeks to minutes. We demonstrate how semantic unification of structured/unstructured data enables dynamic adaptation to process variations while preserving IP security. Case studies in lithography overlay control highlight 95% diagnostic alignment with human experts, with ethical safeguards ensuring human oversight. This work provides a roadmap for democratizing semiconductor expertise through AI-augmented workflows while proposing open standards for industry-wide collaboration.