An Efficient Physics-Constrained Data-Driven Surrogate Model for Anisotropic Damage

To enable efficient multiscale simulation of macroscale structures subject to anisotropic damage, we introduce a novel physics-constrained data-driven surrogate modeling technique. The surrogate model is calibrated to predict the homogenized response of heterogeneous microscale representative volume element models undergoing path-dependent anisotropic damage progression. By construction, the surrogate model guarantees positive strain energy density and nonnegative mechanical dissipation under arbitrarily complex loading and unloading histories, eliminating the risk of nonphysical extrapolation during online prediction and reducing the amount of required data during offline training. We introduce the model, discuss its domain of applicability, and propose a calibration procedure. We then demonstrate the predictive capability of the model calibrated on periodic representative volume element model simulations of the transverse cross section of a notional square-pack composite material with a gradient-enhanced damage constitutive model applied to each constituent at the microscale.