AI assisted HAZOPs— How good can they be?

Abstract

This paper presents a comparative analysis of three HAZOP studies conducted on a small generic solvent purifier system: one using Generative AI (GenAI), one employing an Expert System tool (HAZOP Assistant), and one conducted as a traditional human workshop. The GenAI approach, leveraging probabilistic models, identified hazards effortlessly but introduced inconsistencies as well as proved non‐compliant with the European Union (EU) AI Act's (Regulation (EU) 2024/1689) transparency requirements for high‐risk applications. In contrast, the HAZOP Assistant, with its algorithm‐based logic, achieved a faster analysis than the human workshop. It only generated deviations related to flow, pressure, temperature and level, but managed to fill in all the results from the traditional workshop and highlighted some missed scenarios from it. This makes it possible to ensure that the outputs are aligned with standards and are auditable. The human workshop, while thorough, required more time and personnel, and missed some of the hazards identified by the HAZOP Assistant. This study demonstrates that even though GenAI demonstrates great promise, older, deterministic types of computing have the ability to enhance efficiency and quality in safety studies, offering a scalable solution for process safety while meeting stringent regulatory demands.