Machine learning-based intelligent modeling of hydraulic conductivity of sandy soils considering a wide range of grain sizes

This study presents novel intelligent modeling of the hydraulic conductivity (k) of sandy soil by employing machine learning (ML) algorithms i.e., artificial neural network (ANN), multi-expression programming (MEP) and genetic expression programming (GEP) on a large dataset. For this purpose, an extensive testing program was carried out to evaluate the k-value, gradation and compaction characteristics of a wide spectrum of sandy soils. A broad range of input parameters defining geological characteristics i.e., large, medium and small grain sizes (D), gradation parameters and dry density (γd) were engaged to resolve the limitation of existing k-value predictive models to cover output variability for the different combinations of D-value within a sandy soil deposit. Several possible models were generated by algorithm-guided iterations and varying algorithm inputs; thereof best models were scrutinized. The performance of all the scrutinized ML-based models was found to be reasonable based on various key performance indices (KPIs), i.e., error indices, correlation indices and variance analysis. Finally, the GEP-based model was proposed to predict the k-value based on the best performance among all the models. The proposed model also showed the most reasonable performance on an independent dataset to tackle the aforementioned k-value variability issue based on the Taylor diagram analysis and validation indices in comparison to the existing models in the literature and a singular D-value-based model using the current dataset. The parametric and sensitivity analyses showed that D10 is the most sensitive parameter in the proposed model followed by D50, D60, D5, D30, γd, and coefficient of uniformity (Cu), subsequently.