Hybrid Machine Learning and Metaheuristic Optimization Framework for Predicting Lateral Earth Pressures on Rigid Retaining Walls With Geofoam Inclusions in Sand Backfill

ABSTRACT Expanded polystyrene (EPS) geofoam is increasingly used in retaining structures as a compressible inclusion to reduce lateral earth pressures. However, the complex interactions among soil, EPS, and wall components—together with the limitations of analytical and numerical approaches—pose significant challenges in accurately predicting lateral pressures on rigid retaining walls with geofoam inclusions. This study introduces an intelligent hybrid Sparrow Search Algorithm–Machine Learning (SSA–ML) framework to overcome these deficiencies through autonomous hyperparameter optimization, interpretability enhancement, and uncertainty quantification. Six ML models—Generalized Neural Regression (GNRR), Support Vector Machine (SVM), Extreme Learning Machine (ELM), Decision Tree (DT), Random Forest (RF), and Multilayer Perceptron (MLP)—were optimized using SSA based on a comprehensive database compiled from experimental and numerical studies. Model performance was systematically assessed through K ‐fold cross‐validation and comparative analysis. The kernel density estimation (KDE) method was applied to generate probabilistic interval predictions, thereby quantifying the inherent uncertainty in the modeling process. Furthermore, the SHapley Additive exPlanations (SHAP) approach was employed to interpret the influence and directionality of key parameters on predicted lateral pressures. Results demonstrate that the proposed SSA–MLP model achieves superior predictive accuracy and robustness compared to other benchmark models, while maintaining clear physical interpretability. This hybrid, interpretable, and uncertainty‐aware framework provides a reliable data‐driven tool for analyzing lateral earth pressures on soil–EPS–wall systems and offers new insights for the design and optimization of compressible inclusion retaining walls.