AI-enhanced soil classification with incomplete CPT data for offshore wind farm

Abstract

Accurate soil classification is fundamental to offshore wind farm foundation design, yet conventional cone penetration test (CPT) based methods often require complete datasets that are costly and challenging to obtain in offshore environments. This study presents an artificial intelligence (AI) enhanced framework for soil classification based on the Robertson Classification, with a particular emphasis on robustness under incomplete CPT data. A comprehensive synthetic CPT database comprising 229,808 samples was generated using both uniform and statistically distributed sampling strategies to represent a wide range of realistic soil conditions. Among the four evaluated machine learning models, the random forest model achieved the best performance, with an R & sup2; of 0.99 and a classification accuracy of 92.53%. Simulations of missing CPT input parameters reveal that reliable predictions can be maintained even under incomplete data scenarios. Feature importance indicates that cone tip resistance (qc), sleeve friction (fs) and effective stress (sigma'v), are the dominant factors governing soil classification. Prediction uncertainty using Monte Carlo simulations shows model performance within a 95% confidence interval. Overall, the proposed AI-enhanced framework provides a robust and practical solution for CPT-based soil classification using incomplete datasets for offshore wind farm geotechnical design.