Natural soils’ shear strength prediction: A morphological data-centric approach

The deformation characteristics and constitutive behavior of granular materials under normal forces acting on particles are dependent on the geometry of the grain structure, fabrics and the inter-particle friction. In this study, the influence of particle morphology on the friction and dilatancy of five natural sands was investigated using deep learning (DL) techniques. A Three-dimensional (3D) imaging technique using computed tomography was utilized to compute the morphology (roundness and sphericity) of collected natural sands. Triaxial tests were conducted on the five different natural sands at different densities and confinement stresses (σ₃). From the triaxial results, peak friction angle (${\phi }_p)$, critical state friction angle (${\phi }_{cs}$), and dilatancy angle (ψ) were obtained and modeled using conventional multiple linear regression (MLR) models and DL techniques. A total of 100 deep artificial neural networks (DANN) models were trained at different sizes of first and second hidden layers. The use of MLR resulted in R² of 0.709, 0.565, and 0.795 for ${\phi }_p$, ${\phi }_{cs}$ and ψ, respectively, while the best performed DANN (30 and 50 neurons for the 1st and 2nd hidden layers, respectively) had R² of 0.956 for all outputs (${\phi }_p$, ${\phi }_{cs}$ and ψ) combined. Using the best-performed DANN model, the weight partitioning technique was used to compute an importance score for each parameter in predicting ${\phi }_p$, ${\phi }_{cs}$ and ψ. The σ₃ had the highest importance followed by relative density, roundness, and sphericity with a relative importance of more than 10%. In addition, sensitivity analysis was conducted to investigate the effect of each parameter on the shear parameters and ensure the robustness of the developed model.