This forward-looking article examines how foundation models (large transformer-based LLMs) combined with retrieval-augmented generation (RAG) can transform unstructured process safety "dark data"—particularly incident investigation narratives, shift logs, and reports—into structured, evidence-linked outputs. A highlighted use case is LLM-RAG workflows for root cause analysis: extracting causal factors, performing taxonomy-aligned classification, generating auditable reasoning chains, and producing summaries from real incident reports (illustrated with a BSEE offshore example).
Relevance to PSM incident investigation: It provides a practical blueprint for GenAI to augment expert investigators by improving speed, consistency, completeness, and traceability of RCA outputs. This supports deeper organizational learning, better regulatory defensibility of investigations, and scalable analysis of historical incidents/near-misses, while emphasizing the need for domain grounding, human oversight, and responsible deployment in high-stakes chemical process environments.
Note: This is a very readable overview of LLMs in PSM. The technical depth is limited, when compared to other psm.ai papers; however, the readability alone makes this a valuable resource.
