Granular soils are ubiquitous earth materials that play a crucial role in the geotechnical and geological engineering practice. Granular particles within the granular assemblages can translate, rotate, and break under external loadings, leading to complicated and nonlinear micro-macro mechanical connections of granular soils. Most constitutive theories proposed for the continuous medium do not take into account the granular soils’ micro-scale mechanics. Thus, they could not fully describe the micro-macro connections of granular soils, especially when the multiscale particle morphology was involved. In this context, this research aims to utilise X-ray micro-tomography (μCT), spherical harmonic-based principal component analysis (SH-PCA), discrete element method (DEM) simulations, and machine learning (ML) algorithms to advance the understanding of the micro-macro mechanical behaviours of granular soils under triaxial shear.

In the first phase of this research, a novel spherical harmonic-based morphology descriptor was proposed to describe the incremental morphological variance of the multiscale particle morphology. In the second phase, the macro-micro mechanical behaviours of Fujian River sand (FJS) were carefully investigated through the X-ray μCT-based in-situ triaxial test. In the third phase, an additive virtual particle reconstruction technique was proposed using SH-PCA, allowing for the implementation of morphological gene mutation and decay at different length scales. In the fourth phase, the capability of DEM to describe the micro-macro mechanics of granular soils under different stress conditions was discussed and verified through a proposed μCT-based DEM technique. In the fifth phase, the μCT-based DEM technique was employed to investigate further the multiscale particle morphology effects on different aspects of the micro-macro mechanics of granular soils. In the sixth phase, the μCT-based DEM technique was developed to generate a large informative database considering different degrees of morphological gene-mutation and stress conditions. Based on this database, different ML algorithms were utilised to model the micro-macro mechanics of granular soils. In the seventh phase, the deep transfer learning technique was utilised further to enhance the extrapolation ability of the pre-trained ML models.

It is concluded that the morphology descriptor proposed in the first phase can effectively eliminate inter-scale effects with a uniform definition format across all length scales. The micro-macro mechanical behaviours of FJS in the second phase were highly related to the phase transition point behaviour. The reconstruction technique presented in the third phase can generate particle morphology step by step from large scale to small scale. The morphological gene mutation and gene decay can be readily implemented by setting different scaling factors. By considering different particle morphologies, initial packing states, and confining stress conditions in the fourth phase, DEM with realistic particle morphology successfully reproduces the macro-micro mechanical behaviours of granular soils. By changing scaling factors at different length scales, particle morphology at different length scales was found to contribute the most to different macro-micro responses of granular soils in the fifth phase. A large informative database was generated in the sixth phase, using three different algorithms to model the granular soils’ deviatoric stress-volumetric strain-axial strain relationship. The excellent prediction and verification results reflect the capability and superiority of ML algorithms to model granular soils. Furthermore, the pre-trained ML models had poor extrapolation ability, and the proposed deep transfer learning technique can quickly and accurately transfer pre-trained models to the new out-of-range morphological database.