Explainable machine learning surrogates for lateral pile response prediction in soft clay

Abstract

Accurate prediction of lateral pile response requires computationally intensive nonlinear finite element analysis, limiting its use in routine design and large-scale parametric studies. This study presents a machine learning surrogate framework trained on 15,351 Beam on Nonlinear Winkler Foundation simulations, generated by Latin Hypercube Sampling across an eleven-variable parametric space covering pile geometry, soil properties, loading conditions, and head fixity. Six structural response quantities were predicted simultaneously: maximum bending moment (M_max), lateral deflection (δ_max), shear force (V_max), pile head rotation (θ_head), peak soil reaction (p_max), and the soil capacity utilization ratio (p/p_ult). Extreme Gradient Boosting achieved R² of 0.989 (M_max), 0.819 (δ_max), 0.999 (V_max), 0.824 (θ_head), 0.988 (p_max), and 0.981 (p/p_ult) on a 20% hold-out set, with normalized root mean squared error not exceeding 5.86% across all targets. SHAP TreeExplainer analysis confirmed that lateral load and pile diameter are the dominant predictors, with opposing roles: lateral load governs moment demand while diameter controls deflection and rotation. Undrained shear strength ranked third across most targets. Head fixity reduces deflection by approximately 30 to 40% at high loads while increasing peak bending moment by 25 to 30%, indicating a clear trade-off between deflection control and structural moment demand. Three practical design tools were derived from the surrogate: a joint fixity and eccentricity design chart, closed-form power-law expressions for hand calculation, and a soil utilization index quantifying proximity to plastic failure. The framework provides a computationally efficient and physically interpretable basis for sensitivity analysis and preliminary design of laterally loaded piles in soft clay.