Developing Prediction Equations for Soil Resilient Modulus Using Evolutionary Machine Learning

Abstract

The soil resilient modulus (M_R) is essential to pavement design. This parameter is determined through a costly and time-consuming repeated load triaxial test. Accordingly, prior research focused on implementing complex and interpretable machine learning (ML) models to predict M_R directly from soil parameters. However, existing models rely on either black-box machine learning, sacrificing interpretability, or traditional genetic programming (GP) approaches with underfitting issues. This study introduces an innovative approach using the Adaptive Layered Population Structure Genetic Algorithm (ALPS-GA) to develop accurate and fully interpretable M_R prediction models for cohesive soils. For this purpose, a soil dataset was adopted from the literature with 891 data points for the A-4, A-6, and A-7-6 soil classes. Three M_R prediction equations were developed for each soil class, and the performance of each equation was evaluated using the coefficient of determination (R²), the root mean squared error (RMSE), and the mean absolute error (MAE). The R² for the developed models ranged from 0.91 to 0.93 for the testing set; the RMSE ranged from 7.10 to 8.63 MPa, and the MAE ranged from 5.10 to 7.2 MPa, reflecting high-accuracy models. A comparative bias-variance analysis was done for the proposed models, and it was concluded that they do not tend to overfit or underfit the data, unlike previous models. Finally, a sensitivity analysis was implemented to investigate the impact of each soil parameter on M_R for each soil type.