AI-Powered Distribution-Free Safety Framework for Time-to-Critical-Failure Forecasting

Abstract Reliable forecasting of time-to-critical-failure (TTCF) in safety-critical systems demands more than accurate mean predictions; it must also satisfy stringent regulatory tolerance-interval mandates. Nuclear power plants present a rigorous and high-stakes application domain, where the U.S. Nuclear Regulatory Commission (NRC) requires strict one-sided 95/95 tolerance-interval compliance. Traditional approaches rely on deterministic physics-based simulators, such as TRACE and MELCOR, which rely on parametric model assumptions and scenario-specific analyses. We present an AI-powered, distribution-free two-stage safety framework that rigorously meets these regulatory mandates without imposing parametric assumptions on calibration data. The first stage employs BiFusionNet , a snapshot ensemble with dual-pathway temporal convolutions engineered to extract sharp yet uncalibrated predictive distributions from heterogeneous, multi-rate sensor streams. The second stage applies AI-guided distribution-free calibration, using the predictive 10th percentile as a dynamic risk proxy, machine-learning-informed stratification, and Wilks’ nonparametric order-statistic method to construct robust Lower Tolerance Limits (LTLs). Evaluated on the large-scale Nuclear Power Plant Accident Data (NPPAD) corpus across five forecast horizons (10-300 s), our framework achieved a perfect 60/60 compliance rate, decisively exceeding NRC’s minimum requirements. The framework naturally adapts to longer lead times by widening predictive intervals and producing appropriately conservative LTLs, preserving the 95/95 safety margin while avoiding unnecessary conservatism.