Ensemble Learning Methods for Shear Strength Prediction of Fly Ash-Amended Soils with Lignin Reinforcement

Abstract

For industrial applications in cold regions, a fly ash-amended soil with lignin reinforcement is introduced here. Predictive models were developed for soil shear strength based on two ensemble learning algorithms, to facilitate understanding soil properties with limited prior knowledge. First, over 270 unconsolidated undrained triaxial tests were conducted on the stabilized soil with varying soil moisture contents, percent lignin fiber, confining pressures, and curing conditions (uncured, 15 days curing, and 15 days curing along with freeze-thaw actions). Extreme gradient boosting (XGBoost) and random forest (RF) were then applied to develop the predictive models for the specimens. A parallel genetic algorithm optimized the hyperparameters of both the XGBoost- and RF-based models. Next, monotonicity and sensitivity analyses were carried out with these models, to compare and assess their generalization capabilities. Finally, the peak shear strengths of the specimens finally correlated across different curing conditions. Given different training sample sizes, the XGBoost-based model consistently outperformed the RF-based model, particularly when half of the samples were fed as input. The root-mean-squared error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R2) values of the XGBoost-based model on the test set are approximately 105−135, 5%−7%, and 0.85, respectively. Some essential features that are overlooked by conventional analysis have been identified with ensemble learning. This study demonstrates the potential of ensemble methods to tackle regression and prediction issues arising in civil engineering with less experimental data.