Accurate prediction of soil stress–strain behavior remains a major challenge in geotechnical engineering due to the inherent heterogeneity, nonlinearity, and sparsity of soil datasets. Conventional laboratory and in-situ testing methods are often expensive, time-consuming, and sensitive to sampling disturbances, which limits their efficiency in large-scale engineering applications. To address these challenges, this study proposes an optimized stacking ensemble framework that integrates advanced tree-based learning algorithms with metaheuristic optimization. The selected base learners Light Gradient Boosting (LGB), Extreme Gradient Boosting (XGB), Random Forest (RFR), and Histogram-based Gradient Boosting (HGB) were chosen for their complementary strengths in capturing nonlinear interactions, handling high-dimensional inputs, and maintaining robustness under sparse and heterogeneous data conditions. These models are optimized using the Puma Optimization (PO) algorithm and combined through a stacking strategy to enhance predictive stability and generalization performance. A dataset comprising 1, 410 samples was compiled literature data, witch the K-fold cross-validation was employed to evaluate model robustness. The proposed stacking model, particularly the optimized XGBPO ensemble, achieved superior predictive accuracy with a coefficient of determination (R2) of 0. 9914 in the testing phase, outperforming individual and hybrid models. Interpretability and sensitivity analyses further identified dry density (₃), void ratio (e0), and degree of saturation (Sₑ) as the most influential factors governing soil compressibility behavior. The proposed framework provides a scalable, reliable, and computationally efficient alternative to traditional geotechnical testing methods, offering improved predictive accuracy and practical applicability for infrastructure design and decision-making under complex soil conditions.
Benemaran et al. (Fri,) studied this question.