Prediction of resilient modulus with pre-post experimental data of undisturbed subgrade soils using machine learning algorithms

The resilient modulus (M_R) of subgrade, which shows relationship between stress and unit deformation of a pavement systems under traffic loads, is a design parameter of the pavement structure. Although a cyclic triaxial test apparatus can be used to directly determine the M_R of the subgrade in the laboratory, utilizing prediction models based on easily obtainable soil parameters, is a more efficient method when taking time and cost considerations into account. A comprehensive laboratory testing program is designed to create M_R prediction models using machine learning (ML) algorithms. 70 undisturbed soil samples are subjected to M_R tests, as well as physical and engineering soil properties tests (water content, field density, specific gravity, gradation, consistency limits, unconfined compressive strength, swell pressure, swell percentage). Soil samples are drilled from a highway that has been in operation for over five years.

First, a linear model like MLR is used in the study. Next, nonlinear regression models like RF, GBM, LightGBM, CatBoost, and XGBoost algorithms are used. Research findings showed that nonlinear regression models outperformed linear regression models in predicting the M_R (R² > 0.85), with the XGBoost algorithm yielding the best accuracy (R² = 0.90). Apart from the primary effects such as confining pressure (σ₃) and deviatoric stress (σ_d), it was found that unconfined compressive strength (q_u), natural water content (w_n), and swelling percentage (SR) are significant parameters in the prediction of M_R among all parameters.