Causal Machine Learning for Risk Informed Decision Making in Probabilistic Safety Assessment

Probabilistic Risk Assessment (PRA) provides a well-established framework for evaluating risk in safety-critical systems by modeling accident scenarios, estimating their likelihood, and assessing potential consequences. Central to PRA is the representation of causal relationships among system components, human actions, and external factors, typically captured through structured logical models and expert judgment. Traditional PRA approaches are largely static, while Dynamic PRA extends risk assessment capabilities by incorporating time dependent system behavior. However, neither approach explicitly focuses on causal discovery, intervention analysis, or counterfactual reasoning. Causal Machine Learning (ML) offers a complementary paradigm by enabling formal causal inference and supporting the estimation of causal effects from operational and simulated data. This paper explores the potential integration of causal ML within the PRA framework. Opportunities include improved modeling of complex dependencies, enhanced capability to evaluate the impact of operational or design changes, and the incorporation of data driven insights into risk assessments. A conceptual framework is proposed in which causal ML augments traditional PRA through causal discovery, causal inference, and expert validation while maintaining the central role of domain expertise. An illustrative use case is presented to demonstrate how causal intervention analysis can distinguish observational associations from causal effects when estimating dependencies used in PRA models. Significant challenges remain, including data limitations, causal assumptions, verification and validation requirements, interpretability, and alignment with established PRA practices. The paper concludes by outlining research directions for integrating causal methods into safety assessment workflows and emphasizes the importance of combining data driven approaches with expert judgment to ensure reliability and trust in high-risk applications.