Domain-augmented large language models for automated root cause classification of offshore process incidents

Process industry incidents can inflict fatalities, heavy economic losses, and reputational harm, making robust Process Safety Management (PSM) essential. Yet the volume and free-text nature of modern investigation records outstrip human analytic capacity. We present a domain-augmented large language model (LLM) framework that combines structured Chain-of-Thought prompting with retrieval-augmented generation (RAG) to classify offshore-platform incident reports into the 11 top-level root cause categories defined in the ABS Group Root Cause Map™. A total of 1182 Bureau of Safety and Environmental Enforcement (BSEE) district investigation reports were parsed into structured JSON format, while ABS definitions and illustrative examples were embedded into a FAISS index to support on-the-fly retrieval. Four prompting setups were benchmarked with GPT-4o-mini on 100 randomly selected reports. The Domain CoT + RAG model raised document-level set-based F1-score from 0.552 (zero-shot baseline) to 0.663, driven by a 17 % higher precision. Predicted categories per report fell from 3.92 to 2.78, close to the human average of 2.55, showing that domain context curbs over-classification without sacrificing coverage. Category-level analysis revealed high performance in major categories such as Equipment Reliability, Procedure , and Human Factors (F1 > 0.75), while challenges persisted in Design and Documentation , which require more implicit causal reasoning. Conditional-probability mapping across the full dataset reproduced expected clusters of human-related failures consistent with risk-based PSM theory. These findings demonstrate that combining domain-specific prompts with information retrieval significantly enhances the reasoning capacity of generative LLMs for multi-label safety analytics, offering a scalable, low-cost pathway toward digitizing incident investigations. • An LLM-based framework is proposed to perform automated classification of offshore incidents. • Domain-specific CoT and RAG proves effective in enhancing the overall performance. • Overclassification cut significantly by embedding RCA guidelines and examples. • LLM reproduces human-related root cause clusters predicted by RBP. • The framework could offer efficient and scalable route to digitized RCA in PSM.