Prediction of soil water characteristic curve of unsaturated soil using machine learning

The soil water characteristic curve (SWCC) is pivotal in studying unsaturated soil, elucidating the relationship between soil suction and water content. Its engineering relevance lies in its ability to determine various characteristics of unsaturated soils, such as permeability, diffusivity, and shear strength. However, obtaining a comprehensive SWCC across a broad range of suctions through direct measurement, whether in situ or in the laboratory, presents challenges due to its inherent difficulty, expense, and time intensiveness. Given the inherent challenges associated with directly measuring the Soil Water Characteristic Curve across a wide range of soil suctions, which are often time consuming and costly, researchers have adopted indirect methodologies that utilize regression analysis to infer the SWCC. This study endeavors to employ machine learning algorithms, encompassing linear and non-linear regression techniques, alongside artificial neural networks, to predict the SWCC. These models utilize grain size distribution (GSD) and soil suction as input parameters. To fulfill this aim, a comparative analysis between different models was undertaken, evaluating their performance using statistical metrics. The outcomes reveal a robust association between SWCC and GSD, derived from an extensive dataset. Among the models examined, the Random Forest Regression model demonstrates superior predictive capability for the SWCC, exhibiting minimal mean absolute error (MAE = 0.007), root mean squared error (RMSE = 0.020), and an r2-score of 0.93.