This paper develops a domain-augmented LLM workflow for automated HAZOP worksheet generation. It combines multimodal interpretation of P&ID nodes with retrieval-augmented generation so that GPT-4o, GPT-4o-mini, and Llama 3.2 can draw on external HAZOP and incident knowledge instead of relying only on model parameters. The retrieval corpus includes 6,120 historical HAZOP records and 1,140 incident reports indexed for similarity search. The authors report substantial improvements in semantic agreement and in the validity and balance of generated HAZOP content compared with non-grounded LLM approaches. For PSM, this is especially important because it addresses a central weakness of generative AI in PHA: hallucination and weak domain grounding. The approach shows how an organization’s accumulated PHA and incident experience can be reused to make AI-assisted hazard identification more consistent, traceable, and technically credible, while still requiring expert validation.
The paper describes the complete HAZOP process, from dealing with drawings to generating safeguards and recommendations. The authors discovered that the LLM's tend to emphasize administrative safeguards--which are described in WORDS, something that comes very natural to LLMs--and de-emphasize engineered safeguards. We could have guessed that, but it's interesting that it was confirmed by the authors.
