Physics-Guided Feature Engineering and Machine Learning for Small-Sample Prediction of Offshore Monopile Self-Weight Penetration

Abstract

Reliable prediction of monopile self-weight penetration depth is essential for drivability assessment and installation planning but remains challenging because of nonlinear pile–soil interaction and limited field data. This study compiles records from 143 large-diameter monopiles at five Chinese offshore wind farms and develops a framework combining physics-guided feature engineering, conventional machine learning, and a physics-regularized neural network (PRNN). Variable-length profiles were encoded using up to 10 actual stratigraphic layers, and raw geotechnical features were compared with API-derived unit shaft and tip resistances. Eight conventional models and the PRNN were evaluated on a representative 80:20 pile-level split; robustness was further assessed using 50 repeated splits, leave-one-site-out (LOSO) validation, and paired bootstrap analysis, with the API method as a mechanics-based baseline. On the representative split, XGBoost achieved the highest R2 (0.790), lowest RMSE (2.762 m), and lowest MAPE (12.49%), while Random Forest yielded the lowest MAE (2.268 m). A controlled matched-network ablation showed that equilibrium-based regularization moderately improved mean predictive accuracy and physical consistency. The framework is therefore promising for within-domain prediction, while cross-site application requires further validation.