Constitutive modelling of natural sands using a deep learning approach accounting for particle shape effects

Particle shape plays a vital role in the macroscopic mechanical behaviour of a quasi-statically sheared granular material. However, most of the previous studies neglected the particle shape effects or characterized the particles with simplified geometries when predicting the constitutive behaviour of granular materials. This paper proposes a paradigm-shifting methodology for the constitutive modelling of granular soils subject to triaxial shearing which integrates the techniques of X-ray micro computed tomography (micro-CT), three-dimensional discrete element modelling and deep learning. Firstly, the micro-CT data of Leighton Buzzard sand particles used to reconstruct particles in discrete element method simulations are obtained separately. Secondly, the tomography data of a series of representative sand particles is used to construct the three-dimensional discrete element model of the sand sample which is then used to generate the numerical datasets for the subsequent deep learning task. Thirdly, a deep learning model called the long short-term memory network is developed to capture the combined effects of particle shape, confining pressure, and initial sample density on the constitutive behaviour of sands. Lastly, the effectiveness of the deep learning model is shown by comparing the model prediction on the testing datasets with the numerical simulation results. Furthermore, the capability of the model on predicting the real sand behaviour is demonstrated by an excellent agreement between the model prediction based on the input of tomography data and the soil response measured from the triaxial test.