Digital-twin driven thermo-mechanical coupled modeling for temperature-prediction in reheating furnaces

Accurate furnace-roof temperature prediction is critical for energy-efficient reheating-furnace operation, refractory safety assessment, and cleaner production. However, nonlinear radiation–conduction–convection coupling, multilayer refractory interfaces, heterogeneous boundary conditions, and real-time inference requirements remain difficult to reconcile using conventional numerical or purely data-driven models. This study proposes a multidimensional digital-twin framework for thermo-mechanical coupled modelling of a 2250 walking-beam reheating furnace. The framework couples geometric, thermal, and stress surrogate models with an application-specific Generalized Physics-Informed Neural Network (G-PINN). G-PINN integrates finite-element shape-function-based derivative evaluation, nodal degree-of-freedom assignment for prescribed Dirichlet boundaries, and Gaussian-quadrature-based energy-functional loss within a unified furnace-roof modelling pipeline. Under synchronized quasi-steady validation, the fixed trained model independently maps each time-indexed boundary-condition vector to its corresponding temperature field without online retraining. In Heating Zone II, G-PINN achieves an MAE of 2.3–3.2 °C and an R² of 0.920–0.942, reducing RMSE by approximately 40–50% relative to the steady-state FEM reference. The full framework obtains a mean MAE of 2.75 °C, satisfies prescribed Dirichlet boundaries exactly, and completes end-to-end temperature prediction, gradient evaluation, stress computation, and von Mises post-processing within 8.2 ms. FEM-referenced stress assessment indicates that operational stresses remain below 13 MPa in structurally active regions, while first-order heat-loss sensitivity analysis suggests a 0.64–2.18% uncertainty-bounded potential reduction in redundant heat loss. These results support real-time quasi-steady furnace-roof temperature prediction, stress assessment, and cleaner-operation decision support, while transient thermal inertia remains outside the present validation scope.