Predicting the stress-strain behavior of gravels with a hybrid deep learning approach

Abstract

Large-scale datasets and efficient model algorithms are crucial foundations in Machine Learning. The prediction ability of previous approaches in determining the stress–strain characteristics of gravels is hindered by small datasets and shallow Machine Learning methods with significant limitations in model generalization and feature extraction. With this consideration, a large-scale dataset for the stress–strain–volume change curves from triaxial compression test was established. This extensive collection includes 1039 records for 312 gravel types, along with stress–strain–volume change curves and 13 influence factors related to particle properties, soil mass properties, and test conditions. Subsequently, drawing inspiration from the success of hybrid Deep Learning models in sequence prediction tasks, such as air quality prediction, a novel Deep Learning model named Res-LSTM-PiNet was proposed through ablation studies. This model integrates the capabilities of the Residual Neural Network (ResNet) for deep feature extraction and Long Short-Term Memory (LSTM) for sequence feature learning, while also incorporating prior information constraints into the loss function. The results demonstrate that the proposed model effectively captures and predicts the mechanical behaviors of softening/hardening and shrinkage/dilatancy of gravels. Compared with the traditional LSTM model, the Mean Absolute Percentage Error of Res-LSTM-PiNet in predicting the stress–strain curve is significantly reduced from 28.2% to 14.3%. This study offers effective support for predicting the stress–strain–volume change curves of gravels in the absence of experimental data.