On the application of PINN-style physics-regularised neural networks to high-temperature creep rupture life prediction

Le résumé fourni par la source

This study explores the potential of physics-informed neural networks (PINNs) to improve long-term creep life predictions for 2.25Cr-1Mo (Grade 22) steel using only short-term experimental data. Three modelling approaches were evaluated: a purely physics-based semi-empirical model based on the Stress-Modified Ductility Exhaustion (SMDE) formulation, a purely data-driven neural network (NNN), and a family of PINN models combining empirical learning with physics-based regularisation via a dual-loss approach. Model performance was assessed in both interpolation (within the short-term training domain) and extrapolation (on unseen long-term data). The SMDE model served as a reliable physics-based baseline, exhibiting stable interpolation and extrapolation behaviour. In contrast, the NNN model overfitted the short-term data and failed to generalise to long-term conditions. Through systematic exploration of physics–data weighting, two PINN configurations with physics weighting of 0.70 and 0.75 were identified as optimal, based solely on interpolation performance. These models subsequently outperformed both the SMDE and NNN baselines in extrapolation, demonstrating stable, conservative predictions beyond the training range. It should be noted, however, that unlike classical PINNs that embed partial differential equation (PDE) or ordinary differential equation (ODE) residuals via automatic differentiation, the present framework employs the semi-empirical SMDE creep damage formulation as a constitutive physics-based model. We therefore describe it as a PINN-style, physics-regularised neural network, which balances empirical fidelity with mechanistic regularisation. The novelty of this work lies in applying such a PINN-style dual-loss framework to creep rupture life prediction for the first time, integrating a mechanistic creep damage model directly into neural network training, and demonstrating improved extrapolation capability when only short-term data are available. These findings highlight the value of the PINN framework in enhancing model generalisation when only limited experimental data are available, particularly in contexts where physical mechanisms are well understood.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.