The chemical process safety community is entering a new era driven by foundation models, shifting from task-specific, label-intensive deep learning to adaptable pretrained reasoning frameworks. This Commentary surveys how large language models (LLMs) and vision foundation models (VFMs) can address persistent bottlenecks in process safety, including data scarcity, rare-event imbalance, and limited transferability across facilities. From a language perspective, retrieval-augmented generation (RAG) enables models to operationalize unstructured industrial “dark matter”, such as incident narratives, maintenance logs, management-of-change records, and standard operating procedures, into evidence-linked outputs that support auditable root cause analysis (RCA) and scalable, semiautomated hazard and operability study (HAZOP) workflows. From a vision perspective, promptable VFMs and multimodal systems advance safety monitoring beyond binary detection toward semantic segmentation and contextual discrimination, enabling interpretable, morphology-aware characterization of dynamic threats. Despite this promise, deployment in safety-critical environments demands rigorous scrutiny. We identify key barriers including hallucination risks, ontology drift, privacy, and governance constraints, as well as the inadequacy of generic benchmarks for engineering-grade requirements in reliability, traceability, and failure transparency. We also propose a roadmap centered on “data as infrastructure”, emphasizing expert-supervised synthetic supervision, domain-shaped evaluation and alignment protocols, and hybrid workflows that integrate physical knowledge and validation layers. Ultimately, foundation models should be conceptualized not as autonomous decision-makers but as transparent, context-aware reasoning layers that empower human experts to convert fragmented industrial data into actionable safety intelligence.
Yang et al. (Tue,) studied this question.