Hybrid Stacking Ensemble Learning for Direct Prediction and Geotechnical Interpretation of Liquefaction Factor of Safety in the Port Sudan Coastal Plain

Soil liquefaction is a major seismic hazard that threatens coastal infrastructure. Yet, accurate prediction of the liquefaction factor of safety (FS) remains challenging because of the complex nonlinear interactions among geotechnical and stress-state variables. This study proposes a novel leakage-aware Hybrid Stacking Ensemble (HSE) framework for the direct prediction and geotechnical interpretation of FS in the Port Sudan coastal plain. The proposed model combines Random Forest, Extra Trees, and Gradient Boosting Regressors through a RidgeCV meta-learner while restricting the predictor set to independent spatial, geotechnical, and stress-state variables to prevent data leakage. A database comprising 534 Standard Penetration Test (SPT)-based observations from 61 boreholes was used to evaluate the proposed model against K-Nearest Neighbors (KNN), Artificial Neural Network (ANN), Random Forest, and XGBoost. The Hybrid Stacking Ensemble achieved the highest predictive performance, with a test RMSE of 0.058, an MAE of 0.032, and an R2 of 0.985, demonstrating improved accuracy, robustness, and generalization over the benchmark models. SHAP-based interpretation identified N1(60)cs as the dominant predictor of FS, followed by relative density, depth, and stress-related variables, confirming that the proposed framework learned physically meaningful relationships consistent with established liquefaction mechanics. The contribution of this work lies in the development of a Hybrid Stacking Ensemble framework for the direct prediction of the liquefaction factor of safety (FS), integrated with a leakage-free modeling workflow and explainable machine learning techniques, rather than in the development of a new machine learning algorithm.