Application of multiple machine learning algorithms for intelligent prediction of the strength of fine-grained natural soils

This study presents a novel intelligent approach for predicting the unconfined compressive strength (UCS) of fine-grained natural soils by utilizing machine learning (ML) techniques such as Gradient Boost (GB), random forest (RF), and Extreme Gradient Boost (XGB) on a large dataset obtained from multiple sources. A comprehensive testing initiative was conducted to assess the UCS, sieve analysis, Atterberg limits, and specific gravity of natural soils. To overcome the limitations of existing UCS predictive models in covering output variability for the fine-grained natural soil deposit, a diversity of input parameters defining natural soil attributes, such as the percentage of fines, sand, plasticity index (PI), specific gravity (Gs), and liquid limit (LL), were employed. Multiple ML models were developed through Python code with varying algorithm inputs, and the models with the best predicting abilities were analyzed. The ability of the ML models to predict based on the number of statistical performance indices (SPIs) such as correlation indices, i.e., coefficient of determination (R²), Nash–Sutcliffe efficiency (NSE), and Pearson correlation coefficient (PCC); and error indices, i.e., root mean square error (RMSE), Willmott index (WI), and mean absolute error (MAE), were analyzed and found to be reasonable based on SPIs. Based on the rank analysis of SPIs, the XGB model was proposed to predict the UCS value of natural soils. Sensitivity and parametric analyses revealed that LL has the most significant effect on prediction in the proposed model, pursued by PI, fines, sand, and Gs. The proposed XGB approach is a potentially effective asset to geologists and engineers to predict the UCS for new datasets of natural soils and liquid limits ranging between 20 and 40.