A curated collection of research, papers, and published resources organized across 11 categories, 20 RBPS elements, and every major AI discipline.
As interest in artificial intelligence (AI) accelerates across safety-critical industries, organizations face a growing challenge: distinguishing meaningful, process safety–relevant applications from general-purpose automation and hype.
psm.ai addresses this gap by organizing research and insights across the elements of CCPS Risk-Based Process Safety (RBPS), providing a structured, vendor-neutral view of how AI is being applied—while progressively expanding coverage across all 20 RBPS elements as relevant research and practical applications emerge.
Explore how AI is being applied in process safety disciplines →
This posting begins a deep dive on using GenAI for HAZOPs.
Yang et al. (Texas A&M) focus on the first step of a HAZOP, namely using GenAI for identifying
hazard scenarios.
Lee at al. look at the validity of GenAI results. They tend to be unimpressive (19-37% technically valid). However, this work was done using GPT-4o and other 2025 frontier models, and this field is advancing so rapidly that current “technical validity” rates are likely to be higher. Nonetheless, their
conclusions are still meaningful. This paper is regrettable behind a paywall, but the public domain
abstract and introduction are very comprehensive.
Elhosary and Moselhi invested considerable effort in improving the results of 2025 frontier models (GPT-4o, etc.) by vary careful application of retrieval-augmented generation (RAG). Their results are better than Lee et al., and that may be due to the influence of RAG.