Abstract
Abstract
This study presents a novel data generation framework that generates a large database for machine learning (ML)-based soil model predictions. The dataset comprised 216 sets of triaxial tests on morphologically mutated and gene-decayed granular samples. This database was then estimated using five widely utilized ML algorithms to predict the stress-strain relationship of granular soils. They include the support vector machine (SVM), bagged trees, Gaussian process regression (GPR), and back propagation neural network (BPNN) algorithms. Following the hyperparameter settlement, model training, and testing, all the trained models captured the effects of the multiscale particle morphology, initial packing state, and confining stress. The excellent training and testing performances indicate the superior quality of the generated dataset. The fine tree, exponential GPR, and BPNN outperformed the Gaussian SVM and bagged trees in terms of the predictive performance. Among them, the exponential GPR exhibits the best model performance in reflecting the particle morphology effect, whereas the fine tree and BPNN generally exhibit good predictive performance for missing local information. Furthermore, all the models are tested by the micro-tomography (μCT) experimental data. The findings of this study were validated through a comparison between the DEM and model prediction results.