Data-driven constitutive modelling of granular soils considering multiscale particle morphology

This paper adopts a micro-tomography (& mu;CT)-based discrete element method (DEM) technique to generate a database for the constitutive modelling of granular soils. A large DEM database comprising 217 DEM simulations of morphologically gene-mutated and gene-decayed samples was generated. Based on this database, three neural network models, i.e., the backpropagation neural networks (BPNN), long short-term memory neural networks (LSTM), and gate recurrent unit neural networks (GRU) were utilised to predict the constitutive behaviours of granular soils. After training and testing, all trained models can reasonably predict granular soils' deviatoric stress-volumetric strain-axial strain relationship. It is found that: 1) the effects of particle morphology at different length scales, sample initial packing state, and confining stress condition can be well captured by all these models; 2) LSTM and GRU outperform BPNN in predictive performance with more local information; 3) With fewer weights and biases, more efficient computation, and more stable and even error distributions for different stages of axial strains, GRU shows the best predictive performance, followed by LSTM and BPNN. Furthermore, all three models are tested by the & mu;CT experimental data. The excellent consistency between model prediction and experimental results reflects these algorithms' feasibility, capability and generalization for the constitutive modelling of granular soils.

Automated summary

This paper uses a micro-tomography-based discrete element method to generate a database and neural network models for predicting the constitutive behaviors of granular soils. The trained models can reasonably predict the stress-strain relationship, with GRU model showing the best performance.