A thermodynamics-informed neural network for elastoplastic constitutive modeling of granular materials

Data-driven methods have emerged as a promising framework for material constitutive modeling. However, traditional data-driven models are hindered by limitations arising from a scarcity of datasets and the absence of explicit physical principles. These limitations lead to inadequate generalization capabilities and predictions that may contradict established physical laws. To overcome these challenges, this study introduces a thermodynamics-informed neural network (TINN) for elastoplastic constitutive modeling of granular materials. Following the thermodynamics-based elastoplastic theory, the TINN model incorporates elastic free energy, stored plastic work, and dissipation. It achieves this by integrating a path-dependent recurrent neural network (RNN) and three sub-fully connected neural networks to capture the mechanical response and energy evolution in sheared granular materials. The total loss function of TINN combines data-driven and physics-informed components. The effectiveness and generalization capabilities of TINN are evaluated by testing it on diverse datasets, including both simulated and experimental data. The simulated virtual data are derived either from an available elastoplastic model or via discrete element method (DEM) simulations employing stress probing techniques. The results highlight the superior generalization ability and robustness of the TINN model, surpassing purely data-driven models in performance and reliability.