A predictive deep learning framework for path-dependent mechanical behavior of granular materials

As we transition into an era of data generation and collection, empirical summaries in the classical continuum modeling of granular materials cannot take full advantage of the increasingly larger data sets. This work presents a data-driven model for modeling granular materials, with the material data being extracted from discrete element method (DEM) simulations. A long short-term memory (LSTM) network is then employed to learn the mechanical behaviors of granular materials from the material dataset. Particular emphasis is placed on three elements: modification of LSTM unit cell, phase space sampling, and material history parameterization. The LSTM unit cell is modified so that the initial hidden state can be specified as the initial states of granular materials. Massive DEM simulations are performed to consider the effects of particle size distribution, initial density, confining pressure, and loading path on the mechanical behaviors of granular materials. The history-dependency of the granular materials is well represented by the architecture of the LSTM network and internal variable-based history parameterization. We compare the model predictions against DEM simulations to assess the performance of the proposed data-driven model. The results demonstrate that the model can predict the material behaviors of granular materials with different microstructures and initial states and reproduce the material responses under complex nonmonotonic loading paths. This data-driven model exhibits good generalization ability and high prediction accuracy in various situations.