Causal Discovery for Root Cause Analysis and Risk Prediction

Project summary

Finding the causes you cannot see.

Nuclear engineered assets operate in uncertain, evolving environments. Failures in such systems are rare, but can be catastrophic. Failure is often the result of subtle interactions among multiple latent (and unknown) causes, not a single observable trigger. This project proposes the development of advanced causal inference and uncertainty-aware models to identify hidden root causes and predict long-term risk in nuclear infrastructure and associated assets.

The EngD candidate will develop novel methods to discover causal structure from sparse, time-varying data, particularly when external conditions or monitoring quality change over time. Using recent advances in causal representation learning, dynamic graph algorithms, and imprecise probability, the student will build interpretable models that can answer "what-if" and counterfactual queries and support engineering decisions on inspection, maintenance and lifecycle extension.

The goal is not just accurate forecasting, but explanation and trustworthy risk reasoning. In partnership with the industrial partner, the project will include a secondment focused on practical workflows for condition monitoring and safety cases. Outcomes include novel algorithms, open-source tools, and industry-relevant use cases, all contributing to safer, more sustainable infrastructure management.

Aims and objectives

Aim: develop a rigorous, explainable, and uncertainty-aware causal inference framework to identify latent risk factors and support root cause analysis and decision-making for structural integrity in safety-critical systems.

Objectives:

1. Causal discovery under latent confounders and time-varying data. Develop dynamic causal discovery algorithms to extract directed acyclic graphs (DAGs) from structural health monitoring (SHM) data, accounting for latent variables and changing sensor fidelity. Integrate domain knowledge (degradation physics, stress-strain relations) into causal discovery to increase interpretability and trust. Address epistemic uncertainty by combining sparse observational data with expert priors and uncertainty intervals (p-boxes, imprecise probabilities).
2. Counterfactual reasoning and root cause attribution. Implement models that support counterfactual analysis ("What if component X was inspected earlier?") and embed causal graphs in simulation-based inference loops for backwards root-cause tracking following anomalies or failures. Design interpretable interfaces for visualising pathways from observed failures to probable causes with confidence bounds.
3. Uncertainty quantification and safe prediction. Use interval-based uncertainty propagation rather than fixed statistical distributions to ensure predictions have defensible, guaranteed bounds. Explore information-theoretic measures of uncertainty and information gain to prioritise monitoring, inspection and modelling effort.
4. Scalable algorithms and real-time monitoring integration. Adapt efficient message-passing and update-time algorithms to enable real-time inference on large graphs with cycles and diamond substructures. Extend to time-varying DAGs and streaming SHM data with out-of-order or delayed updates. Investigate deployment on embedded systems or edge devices for field integration.
5. Engineering impact and secondment translation. Collaborate with the industrial partner on use cases drawn from structural integrity assessments, safety case documentation or root cause reports. Translate academic methods into practitioner-ready prototypes with validation on synthetic or historic SHM datasets.

Alignment to STAND-UP impact targets