Latent Risk Pattern Discovery in Hydrogen Incidents via Retrieval-Augmented Generation

The transition to a global hydrogen economy introduces safety challenges driven by situations such as flammability and risk of leakage and by the conditions necessary for its storage. In this high-risk context, rigorous analysis of past incidents is fundamental for mitigating risks and having good public confidence. This study proposes a methodology based on Retrieval-Augmented Generation (RAG) with a domain-oriented risk analysis framework to automate the extraction of lessons learned from hydrogen-related accident reports. While the HIAD 2.1 database already includes a specific field for "lessons learned", this automated model enhances the process by uncovering non-explicit insights and ensuring consistency across large datasets. The approach enables contextual inference, allowing the model not only to retrieve information explicitly stated in the reports but also to derive latent risk patterns, provide cross-incident comparisons, and identify gaps in causal investigation that are not readily observable. Rather than answering predefined high-level questions, the RAG model analyses individual event descriptions to generate structured technical interpretations and prevention alerts, which can subsequently support cross-incident comparisons and broader risk insights. The methodology is applied to the HIAD 2.1 database, which contains structured and unstructured descriptions of accidents and incidents involving hydrogen in vehicles, transport systems, processing plants, and storage facilities. The study compares the RAG-generated insights with the existing lessons in the database to evaluate the model's effectiveness. The results show that the approach reduces interpretive bias and increases analytical consistency by converting incomplete descriptions into actionable safety information. Overall, the framework supports a shift from manual, retrospective analyses to a data-driven decision support process, contributing to more consistent hydrogen safety analysis and supporting the safe deployment of hydrogen technologies.

Keywords: Retrieval-Augmented Generation, Hydrogen Safety, Risk Analysis.