Physics-guided neural network with an asymmetric safety loss for fire resistance prediction of insulated FRP-strengthened RC beams

Abstract

Machine-learned surrogates predict fire resistance (FR) of insulated FRP-strengthened RC beams accurately but violate basic physics, extrapolate unpredictably, and treat unsafe over-predictions the same as conservative under-predictions. We develop a physics-guided neural network (PGNN) that enforces six monotonicity constraints via automatic differentiation, two bound constraints, and a limit-state collocation constraint beyond the data envelope, trained with an asymmetric loss that penalises over-prediction fourfold. On 21,386 simulated fire exposures, a 2×2 factorial ablation over five seeds shows physics constraints cost no accuracy (test R² = 0.877 ± 0.005 vs 0.870 ± 0.014 unconstrained) while a model-agnostic Physics Violation Rate—introduced here for fair comparison with tree ensembles—drops from 67.7% (unconstrained MLP) and 55.3% (XGBoost) to 6.6%, and to 0.2% out-of-distribution. The asymmetric loss converts an implicit design margin into a learned one, cutting unsafe-error rate by 40% in-distribution and over 60% out-of-distribution. Physics and asymmetric constraints also reduce seed-to-seed dispersion of out-of-distribution safety behaviour sevenfold—a property we term constraint-induced reliability. Blind validation on 50 real furnace tests, graded by extrapolation severity, confirms the conservative failure mode transfers to real beams (with deployment guard, worst-case unsafe error 31.7 min vs 210.2 min for unconstrained control) while revealing that narrow material-property ranges in the simulation campaign prevent accuracy validation of any surrogate—identifying training-envelope widening as the binding constraint.