Hybrid finite element and machine learning approach for estimating lateral capacity of partially rock-socketed monopiles

Abstract

This study combines finite element (FE) simulations and soft computing techniques to predict lateral capacity (H) of partially rock-socketed monopiles. A three-dimensional (3D) nonlinear FE model of monopile-soil interaction was developed and validated against centrifuge experimental data. The FE simulations were used to generate 240 monopile configurations, where the monopile length (L), diameter (D), lateral loading eccentricity (e) and sand relative density (D_r) were varied while keeping the rock-socketed length equal to the monopile diameter. We applied three machine learning algorithms to this database to establish predictive models: Adaptive Boosting (AdaBoost), Random Forest (RF), and Extreme Gradient Boosting (XGBoost), for which the hyper-parameters were optimized and evaluated by 10-fold cross-validation. The XGBoost demonstrated highest prediction accuracy (R²:0.986, RMSE: 5.43), outperforming RF and AdaBoost. Shapley Additive Explanations (SHAP) showed the relative importance of input features, ranking them as D > e > L > D_r. A parametric study further verified that the model outputs capture physical behavior and align with theoretical understanding by varying each input while keeping others constant at their mean values. A flexible prediction framework was developed using MAPE and conservatism level as guiding metrics. To unable real-world use, a user-friendly GUI was implemented based on XGBoost model, facilitating efficient prediction without requiring extensive FE computations.