Standardizing Failure Analysis Data and Enabling AI-Driven Learning Through Structured FA Reporting and Legacy Document Conversion
Standardizing Failure Analysis Data and Enabling AI-Driven Learning Through Structured FA Reporting and Legacy Document Conversion
Tuesday, October 6, 2026: 11:10 AM
Summary:
Failure Analysis (FA) in semiconductor development generates high-value knowledge, but much of that knowledge remains difficult to reuse because reports are often captured in inconsistent or unstructured formats. This paper presents a practical framework for AI-driven learning in FA by combining a standardized, machine-readable FA report schema with a large language model (LLM)-based pipeline for converting legacy FA documents into the same normalized structure. The proposed representation separates report-level and candidate-level data while also preserving annotations, images, and supporting evidence. For active investigations, structured capture enables closed-loop learning between diagnosis and confirmed outcomes. For historical investigations, schema-guided reconstruction unlocks institutional knowledge stored in legacy reports and integrates it into a unified FA repository. Together, these capabilities support cross-case search, similar-case retrieval, recurring-pattern detection, diagnosis-to-FA correlation, and the creation of labeled datasets for future AI models.
Failure Analysis (FA) in semiconductor development generates high-value knowledge, but much of that knowledge remains difficult to reuse because reports are often captured in inconsistent or unstructured formats. This paper presents a practical framework for AI-driven learning in FA by combining a standardized, machine-readable FA report schema with a large language model (LLM)-based pipeline for converting legacy FA documents into the same normalized structure. The proposed representation separates report-level and candidate-level data while also preserving annotations, images, and supporting evidence. For active investigations, structured capture enables closed-loop learning between diagnosis and confirmed outcomes. For historical investigations, schema-guided reconstruction unlocks institutional knowledge stored in legacy reports and integrates it into a unified FA repository. Together, these capabilities support cross-case search, similar-case retrieval, recurring-pattern detection, diagnosis-to-FA correlation, and the creation of labeled datasets for future AI models.