An Optimized Kolmogorov–Arnold Network for Predicting the Uplift Capacity of Open-Caisson Anchors in Dense Sand Using Barnacles Mating Optimizer

This study investigates the drained uplift capacity (N_g) of open-caisson anchors embedded in dense sand using an integrated numerical and data-driven framework. Finite Element Limit Analysis (FELA), incorporating the Bolton failure criterion, is employed to capture the stress-dependent strength and dilatancy behavior of dense sand. Parametric analyses reveal that the embedment ratio (H/B) is the dominant parameter governing uplift resistance. In contrast, geometric inclination (β) at large embedment depths exhibits practically negligible influence, while increasing the radius ratio (R/B) reduces N_g under deep embedment conditions. Soil strength parameters also show systematic influence; increasing the critical-state friction angle ( \(\phi_{cv}\) ) increases N_g, particularly at large embedment depths where confinement effects dominate. To enable rapid prediction of uplift capacity, a Kolmogorov–Arnold Network (KAN) optimized using the Barnacles Mating Optimizer (BMO) is developed as a surrogate model. The KAN–BMO model achieves excellent predictive performance (R² = 0.9998) for both training and testing datasets while maintaining consistency with the physical trends identified by FELA. SHAP-based interpretability analysis confirms that H/B remains the dominant governing parameter, demonstrating that the surrogate model preserves the underlying mechanistic hierarchy. The proposed framework integrates physically based numerical analysis with data-driven prediction to support practical uplift capacity assessment.