Data-driven strain–stress modelling of granular materials via temporal convolution neural network

Abstract

Machine learning offers a new approach to predicting the path-dependent stress–strain response of granular materials. Recent studies show that temporal convolution neural (TCN) networks, a mutation of the 1D convolution neural network (CNN), have a powerful capability of addressing time-related prediction tasks. In this work, TCN networks are constructed to explore their potential in capturing the constitutive law of granular materials. To train and test the TCN network, three types of numerical experiments are implemented to generate datasets via discrete element modelling. The Bayesian optimisation method is employed to find the optimum architecture of the network. Furthermore, to improve the training accuracy and efficiency, a transfer learning (TL) scheme is innovatively leveraged, which utilises the trained network parameters from a set of shorter time steps and/or coarser data points of the training strain–stress loading curves, as the initial values, to train the network for a longer time step. The prediction ability of the trained TCN network is assessed and compared with a recurrent neural network which has been proved to perform well in predicting constitutive laws of the granular materials. In addition, training datasets with artificially added noise are also used to test and analyse the robustness of TCN networks.