Deep Learning Predicts Stress–Strain Relations of Granular Materials Based on Triaxial Testing Data

This study presents an AI-based constitutive modelling framework wherein the prediction model directly learns from triaxial testing data by combining discrete element (DEM) and deep learning. A learning strategy is proposed based on generally accepted frame-indifference assumption in constructing material models. The low-dimensional principal stress-strain sequence pairs, measured of testing, are used to train recurrent neural networks, then predicted stress augmented other high-dimensional or general tensor via coordinate transformation. Through detailed hyperparameter investigations, it found that long short-term memory (LSTM) gated unit (GRU) networks have similar performance problems, both satisfactorily predict responses granular materials subjected a given unseen strain path. Furthermore, unique merits ongoing challenges data-driven models for discussed.