This dissertation proposes and evaluates two approaches, with-in a Process Risk and Analysis (PRIA) framework, that advance data-driven hazard discovery in complex cyber-physical systems. First, a transferability-based clustering method groups subspace-identified segments into operational modes, prioritizing cross-predictive performance over raw feature proximity. Second, a Neural Switching Linear Dynamical System (NeuralSLDS) model anchors learning in identified state-space models with Bayesian residuals, calibrated predictive uncertainty, and context-aware operating mode transitions. Together, these methods generate physics-informed, probabilistic world models that can support model-in-the-loop analysis of hazardous trajectories. The approach is validated on a simulated point-kinetics reactor and a centrifugal chiller dataset, demonstrating coherent mode discovery across regimes, low short-horizon prediction error with well calibrated predictive uncertainty. Overall, PRIA works toward bridging the gap between high-fidelity physics-based models and reinforcement learning algorithms for the purpose of novel hazard discovery that contributes to enhanced safety and security of cyber physical systems.
Raymond Fasano (Thu,) studied this question.