Improving the prediction accuracy of small-strain shear modulus of granular soils through PSD: An investigation enabled by DEM and machine learning technique

This study presents a numerical investigation into the effect of particle size distribution (PSD) on the small-strain shear modulus (G0) of granular soils using the discrete element method (DEM). Drained triaxial tests were conducted on a set of spherical particle assemblies covering a wide variety of PSDs. The tests demonstrated that the G0 decreases with increasing the coefficient of uniformity (Cu), assuming the confining pressure (p') and the void ratio (e) remain constant. A corrected void ratio (e4), which regards mechanically unstable particles with less than four contacts as void, was developed. Analyses performed afterward revealed that e4 is more strongly correlated with G0 than e. Consequently, a Hardin-based G0 model parameterized by e4 and p' was established and it is more accurate than two commonly used models, which are parameterized by e, p' and Cu. Furthermore, micromechanical and microstructural features were investigated to acquire a better understanding of why models based on e and Cu yield less accurate results than models based on e4. Finally, a hybrid model combining the high explanatory power of statistical models with the high predictive power of machine learning models were proposed. The hybrid model which only uses measurable macroscopic parameters performs on par with the microstructural model and is more accurate than conventional empirical models. The proposed model offers a simple and streamlined approach to predicting G0 with satisfactory accuracy, which is a feasible strategy for maximizing the utilization of copious experimental data.