Data-driven surrogate modeling for accelerated multiscale FE 2 simulation of nonlinear thermomechanical behavior in composite materials

Glass fiber reinforced composites exhibit highly nonlinear and path-dependent thermomechanical responses under coupled loading. Accurately predicting this behavior using traditional multiscale frameworks such as FE remains computationally prohibitive, particularly for complex geometries involving constituent-level interactions. To overcome these limitations, this study investigates the Thermomechanical Elasto-Plastic Artificial Neural Network (ThEP-ANN), a novel data-driven multiscale surrogate designed to accelerate fully coupled thermomechanical simulations of such materials. The ThEP-ANN framework circumvents the complexity of explicit constitutive modeling by directly predicting macroscopic quantities (e.g., stresses and energy rates), rather than relying on free-energy potentials. This strategy significantly reduces the computational cost associated with high-dimensional automatic differentiation. Crucially, the surrogate explicitly incorporates the evolution of internal variables, enabling an accurate representation of history-dependent and path-integral material responses. The ThEP-ANN is implemented non-intrusively into the commercial finite element code Abaqus via a Meta-UMAT subroutine, allowing seamless integration into macroscopic simulations. The framework is assessed through a three-stage, fully numerical validation strategy, comprising generalization to unseen loading paths, mesoscopic validation at the Representative Volume Element (RVE) level corresponding to a single macroscopic Gauss point, and macroscopic structural-level simulations on a finite element model. The results demonstrate excellent agreement with high-fidelity FE solutions while achieving orders-of-magnitude reductions in computational cost. This work establishes a practical pathway for efficient large-scale numerical analyses of nonlinear thermomechanical composite structures.