PHYSICS-INFORMED NEURAL NETWORK SURROGATE FOR THERMO-METALLURGICAL-MECHANICAL QUENCHING MODEL

Abstract

A physics-guided conditional neural surrogate is developed for the transient thermal, microstructural, and stress response of a quenched plain-carbon-steel cylinder. The Robin bound- ary condition is governed by a heat-transfer-coefficient (HTC) field h(T,z) whose boiling peaks vary along the axial coordinate. A family of 10,000 admissible HTC functions was generated for train- ing, with 500 validation and 500 test functions. The conditional physics-informed neural network (PINN) receives normalized space–time coordinates and a 40-component HTC descriptor; separate fully connected heads predict temperature, phase fractions, and mechanical quantities. The thermal and metallurgical fields are coupled through transformation kinetics and latent heat, whereas the me- chanical branch is driven by the thermal and phase histories without mechanical feedback to those fields. A conservative competing-hazard phase integrator refines the diffusional and martensitic out- puts without FEM labels. The general six-case screening gives a mean temperature RMSE of 56.17◦C and identifies the early boiling transient as the principal limitation. For the out-of-family HTC03 case, FEM-assisted adaptation and targeted early-time correction reduce the mean full-field temperature RMSE over twelve instants from 6.77 to 3.50 ◦C. A final boundary specialist, assessed at 45 held-out times, gives a mean full-field RMSE of 2.20◦C and a maximum sampled nodal error of 16.62◦C. The FEM-calibrated mechanical comparison gives von Mises RMSE values of 8.09 MPa at 10 s and 1.73–2.76 MPa at 30–600 s. The results distinguish general conditional screening from case-specific numerical refinement.