A physics-informed deep neural network for surrogate modeling in classical elasto-plasticity

In this work, we present a deep neural network architecture that can efficiently surrogate classical elasto-plastic constitutive relations. The is enriched with crucial physics aspects of elasto-plasticity, including additive decomposition strains into elastic and plastic parts, nonlinear incremental elasticity. This leads to Physics-Informed Neural Network (PINN) model named here as Elasto-Plastic (EPNN). Detailed analyses show embedding these the facilitates more efficient training less data, while also enhancing extrapolation capability for loading regimes outside data. EPNN material-independent; it be adapted wide range material types, geomaterials; experimental data potentially directly used in network. To demonstrate robustness proposed architecture, adapt its general framework behavior sands. We use synthetic generated from point simulations based on relatively advanced dilatancy-based granular materials train superiority over regular architectures demonstrated through predicting unseen strain-controlled paths sands different initial densities.