Predictive modeling of soaked and unsoaked California bearing ratio for coarse and fine-grained soils: a comparative study using XGBoost and ANN architectures

Abstract

The California Bearing Ratio is a vital parameter for pavement design, but traditional laboratory testing is expensive and time-consuming. This study addresses the limitations of simplified empirical models by developing advanced data-driven frameworks using Artificial Neural Networks and Extreme Gradient Boosting. Utilizing a comprehensive dataset of soil samples, the research predicts bearing capacity from standard soil index properties and compaction parameters for both fine-grained and coarse-grained soils under soaked and unsoaked conditions. Results demonstrate that Extreme Gradient Boosting significantly outperforms neural networks. For fine-grained soils, the boosting model achieved a coefficient of determination of 0.972 and a mean squared error of 1.78, compared to 0.600 and 21.82 for the neural network. In coarse-grained soils, the boosting model reached a high correlation of 0.976 with nearly half the error of the neural network. By analyzing soil groups separately and employing ensemble learning, this study provides a precise and computationally efficient predictive tool for preliminary subgrade strength estimation. These findings offer a robust alternative to labor-intensive testing, supporting more efficient and cost-effective road infrastructure design.