An Integrated AI Platform for Accelerating Semiconductor Failure Analysis: Combining Similarity Search, Deep-Learning Image Retrieval, Automated Labeling, and Ontology-Guided Workflow Assistance
An Integrated AI Platform for Accelerating Semiconductor Failure Analysis: Combining Similarity Search, Deep-Learning Image Retrieval, Automated Labeling, and Ontology-Guided Workflow Assistance
Tuesday, October 6, 2026: 11:30 AM
Summary:
This paper presents an integrated AI platform that accelerates the semiconductor failure analysis (FA) workflow end-to-end by orchestrating six complementary capabilities: multi-criteria similarity search with nine-factor hybrid reranking, ontology-based automatic label prediction using TF-IDF and GPU-accelerated neural networks, crowdsourced label verification through interactive quizzes, on-premise large language model inference using quantized GGUF models, image similarity retrieval via DINOv2 vision transformer embeddings, and SPARQL-backed process ontology guidance for method sequencing. Unlike point solutions addressing isolated tasks, the platform creates a self-improving feedback loop where every analyst interaction — confirming a label, rating a search result, completing a quiz — feeds back into the underlying models. Pair-specific user feedback is the single most influential reranking signal, ensuring domain expert judgment directly shapes future results. Transparent scoring exposes per-factor breakdowns, building analyst trust. Deployed in production across multiple FA laboratory sites with no cloud dependencies, the platform demonstrates sub-second case retrieval across 50,000+ jobs, steady growth of labeled training data without dedicated annotation campaigns, and LLM-generated summaries that accelerate cross-site knowledge sharing.
This paper presents an integrated AI platform that accelerates the semiconductor failure analysis (FA) workflow end-to-end by orchestrating six complementary capabilities: multi-criteria similarity search with nine-factor hybrid reranking, ontology-based automatic label prediction using TF-IDF and GPU-accelerated neural networks, crowdsourced label verification through interactive quizzes, on-premise large language model inference using quantized GGUF models, image similarity retrieval via DINOv2 vision transformer embeddings, and SPARQL-backed process ontology guidance for method sequencing. Unlike point solutions addressing isolated tasks, the platform creates a self-improving feedback loop where every analyst interaction — confirming a label, rating a search result, completing a quiz — feeds back into the underlying models. Pair-specific user feedback is the single most influential reranking signal, ensuring domain expert judgment directly shapes future results. Transparent scoring exposes per-factor breakdowns, building analyst trust. Deployed in production across multiple FA laboratory sites with no cloud dependencies, the platform demonstrates sub-second case retrieval across 50,000+ jobs, steady growth of labeled training data without dedicated annotation campaigns, and LLM-generated summaries that accelerate cross-site knowledge sharing.