TCN and LSTM neural network for predicting path-dependent constitutive behaviours of idealized granular soils

The constitutive response of granular soils, exhibiting complicated nonlinearity, is deemed to initial state and path-dependent in geotechnical engineering research. As a data-driven methodology, deep learning (DL) provides a higher possibility in time-related prediction. In this paper, Temporal Convolution Neural (TCN) Network, a type of 1D-Convolutional Neural Network (CNN) and Long Short-Term Memory Neural Network (LSTM-NN), are employed to establish the high-dimensional relationships of granular soils from micro to macro levels subjected to multiple stress paths. This work consists of following steps. CTC-model is firstly applied to calibrate the accuracy of DEM simulation against real experimental data. Secondly, multitudes of DEM simulations on idealized granular soils considering the effects of particle size distribution, initial void ratio, confining pressure and loading paths are performed to enlarge datasets. Afterwards, the prediction performance of two trained DL model is evaluated through the variables of principle stress, volumetric strain, deviatoric fabric and mechanical coordination number against the DEMbased datasets. Finally, A full-scale comparison on the generalization ability between the proposed TCN and LSTM-NN on account of extrapolated tests is presented, highlighting the differences in these two timedependent DL models. The results demonstrate that both DL-based models are capable of accurately predicting the high-dimensional constitutive behaviours of idealized granular soils with different initial states, as well as reproducing the mechanical response under complicated stress paths.