A Multimodal Vision–Language Framework for Yield-Oriented and Explainable Wafer Defect Reasoning

Wafer defect diagnosis is critical for semiconductor yield management, yet most artificial intelligence (AI) systems treat wafer-map analysis as a pattern-classification task and provide limited support for structured diagnosis.We study structured multimodal post-training for wafer defect diagnosis and introduce WaferMapVQA, which reformulates wafer-map analysis as a five-stage visual question answering (VQA) task: anomaly detection, defect classification, spatial localization, appearance description, and process-informed analysis. Each stage is framed as a four-option multiple-choice question (MCQ), enabling explicit supervision under a compact and auditable output schema.Our post-training follows a two-stage design: supervised fine-tuning (SFT) stabilizes the answer schema, while Group Relative Policy Optimization (GRPO) refines stage-wise reasoning via format, accuracy, and domain-logic rewards.Experiments on wafer maps derived from the public WM811K dataset show improved question-level accuracy and a strong balance between strict image-level correctness and deployment-oriented errors, including fatal misses and false alarms.These results demonstrate that structured post-training, rather than model scaling, is the key to transforming generic VLMs into reliable and deployable wafer-diagnosis systems.