AI-Driven Optimization of Sustainable Soil Stabilization: A Kernel Learning Approach for Makuba-Treated Lateritic Pavements

Abstract

The stabilization of lateritic soils using agricultural waste materials, such as locust bean pod ash (LBPA) locally known as "Makuba" presents a sustainable alternative for improving pavement subgrades. However, predicting optimal stabilizer dosage and geotechnical performance remains challenging due to complex nonlinear interactions between soil composition, stabilizer chemistry, and compaction conditions. This preliminary proof-of-concept study develops a physics-guided, stabilization-anchored kernel machine learning framework to predict unconfined compressive strength (UCS), California bearing ratio (CBR), and plasticity index (PI) of Makuba-stabilized lateritic soil. The approach integrates experimental data (six mix proportions: 0–12% Makuba) with material science principles through a stabilization-anchored feature set incorporating clay fraction, linear shrinkage, CaO, and SiO₂ content. Three kernel regression models Gaussian Process Regression (GPR), Kernel Ridge Regression (KRR), and Support Vector Regression (SVR) were implemented with hyperparameters optimized via grouped 5-fold cross-validation to prevent data leakage. Results demonstrate that GPR with stabilization-anchored features achieves strong interpolation accuracy (R² = 0.982, RMSE = 6.78 kN/m² for UCS), significantly outperforming baseline models. Feature importance analysis confirms alignment with geotechnical principles, identifying Makuba content, clay fraction, and CaO as the most influential predictors. However, stabilization-band hold-out testing reveals a critical limitation: models trained on low Makuba content (0–6%) fail to generalize to high-content mixtures (9–12%), underscoring extrapolation risks. Uncertainty quantification via GPR provides predictive confidence intervals for risk-informed decision-making. The framework is proposed as a preliminary screening-level tool for dosage optimization and laboratory trial reduction. While promising, results warrant cautious interpretation and emphasize the need for expanded experimental validation prior to field implementation.