Predictive modelling and experimental evaluation of red mud and paddy stubble stabilised soil using Bayesian optimised machine learning

Abstract This study investigates the combined use of red mud (RM) and paddy stubble (PS) as sustainable additives for soil stabilization, integrating experimental characterization with advanced machine learning (ML) predictive modelling. Experimental tests were conducted to evaluate index properties, compaction, unconfined compressive strength (UCS), CBR, triaxial, and consolidation behaviour of stabilized soil composites. The optimal blend (soil + 25% RM + 1% PS) achieved a 250% increase in UCS after 28 days of curing. Complementary tests confirmed improved bearing capacity, shear strength, and reduced compressibility. To enable robust UCS prediction, six regression models including Linear Regression (LR), Polynomial Regression (PR), Decision Tree (DT), Random Forest (RF), XGBoost, and Gaussian Process Regression (GPR) were developed and evaluated. Ensemble-based models (XGBoost, RF, DT) significantly outperformed traditional regression, with XGBoost achieving the highest testing accuracy (R 2 = 0.991, RMSE = 12.1 kPa), followed closely by RF (R 2 = 0.99) and DT (R 2 = 0.99) demonstrating their utility as a reliable predictive tool for geotechnical design. GPR provided reliable predictions (R 2 = 0.985) with uncertainty quantification. Residual analysis, Taylor diagrams, and REC curves confirmed the superior stability and generalization of ensemble methods. Sobol sensitivity analysis identified curing period as the dominant factor controlling UCS, while PS and specific gravity influenced strength through higher-order interactions. LassoCV reduced polynomial feature count by 40% with negligible accuracy loss, offering an interpretable alternative. Collectively, this study demonstrates that RM-PS combined stabilization is a viable, sustainable approach for geotechnical applications, and that ensemble machine learning provides a high-fidelity predictive framework for UCS estimation, facilitating reliable engineering practice.