Undrained uplift capacity prediction of open-caisson anchors in anisotropic clays using XGBoost integrated with mutation-based genetic algorithms

This study evaluates the undrained uplift capacity of open-caisson anchors embedded in anisotropic clay using Finite Element Limit Analysis (FELA) and a hybrid machine learning framework. The FELA simulations investigate the influence of the radius ratio ( R/B ), anisotropy ratio ( r e ), interface roughness factor ( α ), and inclination angle ( β ). Specifically, the results reveal that increasing β significantly enhances N c , especially as soil behavior approaches isotropy. Higher α improves resistance at steeper inclinations by mobilizing greater interface shear. N c increases with r e , reflecting enhanced strength under isotropic conditions. To enhance predictive accuracy and generalization, a hybrid machine learning model was developed by integrating Extreme Gradient Boosting (XGBoost) with Genetic Algorithm (GA) and Mutation-Based Genetic Algorithm (MGA) for hyperparameter tuning. Among the models, MGA-XGBoost outperformed GA-XGBoost, achieving higher predictive accuracy (R 2 = 0.996 training, 0.993 testing). Furthermore, SHAP analysis consistently identified soil anisotropy ( r e ) as the most influential factor in predicting uplift capacity, followed by interface roughness ( α ), inclination angle ( β ), and radius ratio ( R/B ). The proposed framework serves as a scalable decision-support tool adaptable to various soil types and foundation geometries, offering a more efficient and data-driven approach to uplift-resistant design in anisotropic cohesive soils. • New limit state solutions of the undrained uplift capacity factor of open-caisson anchors embedded in anisotropic cohesive soils using finite element limit analysis (FELA) • Developing a hybrid Extreme Gradient Boosting (XGBoost) model by coupling with Genetic Algorithm (GA) and Mutation-Based Genetic Algorithm (MGA) • Employing SHAP analysis, performance metrics, and residual distribution to evaluate the performance of proposed machine learning models