Physics-informed temporal convolutional networks for composite constitutive modeling

We propose a physics-informed temporal convolutional network (PITCN) framework for efficient constitutive modelling of composite materials under random loading paths. The core contributions lie in the effective integration of a temporal convolutional network to encode the historical stress-strain sequence into a material state vector, which is subsequently mapped by a fully connected network to approximate the free energy function. The constitutive relations are then derived via automatic differentiation. To ensure thermodynamic consistency under irreversible deformation, the non-negativity of both the free energy and the dissipation energy is enforced through penalty terms in the loss function. A key advantage of the PITCN framework is that it achieves high predictive accuracy without requiring experimentally inaccessible free energy or internal state variables during training. The method is efficiently integrated into ABAQUS through a user material subroutine, enabling seamless coupling with finite element solvers. The performance of the proposed PITCN is systematically evaluated at the material point level and in structural simulations under arbitrary loading/unloading conditions against conventional multiscale techniques, and is further validated against experimental measurements. Compared with conventional multiscale techniques, the PITCN achieves a speed-up of 19 times and a memory reduction of 107 times in structural applications, while maintaining comparable accuracy.