Can Large Language Models Automate the HAZOP Process without Human Intervention?

This study tests whether four multimodal LLMs (GPT-4o, GPT-4o-mini, LLAMA 3.2, and Gemini 2.0) can autonomously perform complete HAZOP studies from a P&ID using a standardized prompt. 120 HAZOP worksheets were generated and benchmarked against expert-prepared references. While all models achieved high textual similarity (F1 > 86%), the proportion of semantically valid accident scenarios was low (19-37%), and safeguards were heavily biased toward procedural measures. The authors conclude LLMs are useful as assistive tools but not replacements for expert-led HAZOP, and that advanced prompt engineering and domain-specific fine-tuning are needed. Highly relevant to PSM as it provides empirical evidence on capabilities and limitations of prompt-based LLM automation for Process Hazard Analysis.