Benchmarking Mixture of Experts against Stacking ensembles for predicting permeability of leachate-contaminated soils

• Mixture of Experts (MoE) applied as a novel approach in geotechnics. • MoE effectively captures the clustered behavior of contaminated soil materials • MoE showed superior performance over competing Stacking models • Sensitivity analysis identified pH, sulfate, time, pressure, and PI as main drivers • GP provided independent validation of key permeability-controlling features Accurate prediction of permeability in leachate-contaminated soils is crucial for landfill design, yet laboratory testing is often costly and impractical under heterogeneous conditions. This study proposes a selection-based Mixture of Experts (MoE) framework to overcome the limitations of conventional fusion-based ensemble models in representing clustered and nonlinear soil behavior. A comprehensive database of 494 experimental records was compiled from published studies, and the effects of missing-value treatment and outlier handling were systematically examined. The optimized MoE model achieved superior predictive metrics (R² = 0.89, RMSE = 7.11 × 10 ⁻10 , MAE = 3.19 × 10 ⁻10 ), consistently outperforming four stacking ensemble models. Sensitivity analyses identified pH, sulfate concentration, time, confining pressure, and plasticity index (PI) as the dominant permeability controls. To enhance practical applicability, an interpretable Genetic Programming (GP) model was developed using these key features, yielding reliable accuracy (R² ≈ 0.87) and robust generalization across multiple test sets.