Prediction of hydraulic conductivity of sand with multivariate-index properties using optimal machine-learning-based regression models

The determination of geotechnical correlations and coefficients relies on the assumption of error-free training data used for empirical models, but this assumption may not always hold true. Empirical-risk assessments based on noisy data cannot guarantee the accuracy of regression results. The statistical robustness of empirical-design parameters is influenced both by the soil properties and regression models used. This study proposed a non-stochastic regression method for predicting the hydraulic conductivity of sandy soils based on relevant soil parameters. No changes in content have been made. The approach involved the following steps: data preprocessing, regression-algorithm selection, model optimization, uncertainty estimation, and model selection. The study identified trends in hydraulic conductivity and pore-structure characteristics based on specific model parameters that were derived from empirical data. The paper presents a compilation of a regression model and methods for fine-tuning parametric representations. The prediction results highlighted the best-fitting model and parameter combination with the lowest residuals, comparing favorably to empirical regression models. Machine-learning-based regression models suggest an optimal combination of properties while considering model performance and handling missing values for uncovering the relationships between hydraulic conductivity and multiple, influential, soil properties.