Physics-Informed Responsible AI for Predictive Process Safety and Asset Integrity: A Statistically Validated Framework for Catalyst-Degradation Monitoring in Hydrogen Production

Industrial process safety in high-consequence sectors such as hydrogen production increasingly looks to artificial intelligence for predictive, rather than reactive, risk management. A recurring weakness of this literature is the publication of physics-informed neural network (PINN) frameworks whose headline performance claims are not independently verifiable from the accompanying implementation. This paper presents a physics-informed, uncertainty-aware framework for predictive catalyst-degradation monitoring in a steam-methane-reforming (SMR) hydrogen production unit, implemented entirely from first principles, with every network operation and gradient computation derived analytically and verified against numerical differentiation to relative error below 10⁻⁸, so that every reported result is exactly reproducible from the accompanying implementation without dependence on opaque automatic-differentiation internals. Rather than assert a predetermined result, we report what a fair, matched-condition, multi-seed statistical comparison against classical Statistical Process Control (SPC) shows: detection accuracy (ROC AUC approximately 0.97) and false-positive rate are statistically indistinguishable between the two methods (p > 0.2, N = 30 independent noise realisations); SPC's fixed threshold is, in fact, significantly faster to first alarm in the matched stationary scenario (p < 0.0001). The genuine, statistically robust advantage of the physics-informed framework emerges under non-stationary conditions: under benign ambient sensor drift of realistic magnitude, the adaptive uncertainty-aware detector maintains a 0% false-alarm rate across all tested drift rates, while the fixed-threshold SPC baseline degrades to a 9.2% false-alarm rate (N = 5 seeds per condition, consistent effect across all replicates). Remaining-useful-life (RUL) prognostics show a similar pattern: point estimates are comparable, but the physics-informed estimate is materially more consistent across noise realisations (coefficient of variation 0.22 versus 0.46). A 48-run Sobol variance-based sensitivity analysis identifies sensor-noise level and the physics relaxation-rate prior as the dominant drivers of detector performance. We provide complete, gradient-verified, reproducible code; an explicit falsifiability protocol; a scientific and technical risk assessment; and an experimental-validation roadmap. This statistically qualified account, in which the framework does not uniformly dominate the classical baseline but offers a specific, mechanistically explained, and empirically verified robustness benefit, provides a rigorous, transparent basis for deployment decisions that a claim of uniform superiority alone would not.