On the interpretability of machine and deep learning techniques for predicting CBR of stabilized soil containing agro-industrial wastes

Problematic soils pose challenges to light infrastructure, such as pavements, due to their swelling and collapse characteristics. Traditional soil stabilizers like cement and lime, while successful, have limits in cost, environmental impact, and energy use. This study explores agricultural and industrial byproducts as alternative stabilizers. It aims to determine the California Bearing Ratio (CBR) of stabilized soils using modern Machine and Deep Learning (MDL) techniques. MDL models used are Multivariate Adaptive Regression Splines (MARS), Artificial Neural Networks (ANN), M5P Model Trees (M5P-MT), Extreme Gradient Boosting (XGBoost), Locally Weighted Polynomials (LWP), and Long Short-Term Memory (LSTM) networks. Two modeling approaches were created: approach I with 12 input variables and approach II with 7. Key input variables were Atterberg limits, Ordinary Portland Cement (OPC), Optimum Moisture Content (OMC), Maximum Dry Density (MDD), dust, and ashes. Feature importance was assessed using Sklearn permutation importance and SHapley Additive Explanation (SHAP). Six statistical measures were used to assess model effectiveness: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), RMSE-to-Standard Deviation Ratio (RSD), Variance Accounted For (VAF), 95% Uncertainty (U95), and Correlation Coefficient (R). All models had high prediction accuracy (R > 0.95), with method II being more straightforward. The LSTM model achieved the highest performance (R = 0.98; RMSE = 3.12), with XGBoost producing comparable results. The most influential variables, according to SHAP analysis, were OPC, additive type, and plasticity index. This study demonstrates the potential of Agricultural and Industrial Wastes (AIWs) in soil stabilization and the usefulness of MDL models, notably LSTM, in accurately predicting CBR levels.