Molecular dynamics enabled data-driven modeling of constitutive behavior and failure in composite materials

Modeling the constitutive and damage behavior of composite materials is challenging due to their complex multiscale composition and the difficulty of experimental characterization at small scales. Molecular dynamics (MD) simulations offer detailed insights into the mechanical behavior of materials at the atomic scale. However, existing challenges of MD simulation of composites include high computational costs and selection of effective force fields. In this work, we develop an MD-enabled surrogate modeling (metamodeling) framework that simultaneously captures rate-dependent and path-dependent constitutive behavior and detects failure at the atomic scale. We model each constituent and their interface separately to reflect their distinct behaviors. We discuss the selection of appropriate force fields in order to produce comparable results among constituents. Time step information is embedded in the data prior to model training, enabling the model to infer the rate-dependent response solely based on the given strain path. Multi-task Gated recurrent unit (GRU)-based neural networks are developed to capture rate- and path-dependent constitutive response and failure. We demonstrate the proposed methodology using glass fiber-reinforced epoxy composite as an example, and train the proposed GRU-based model on MD simulation data. The robustness and generalization capability of the models are demonstrated through comparisons with MD simulation results with unseen loading patterns.