Predicting downhole rock friction angles in complex geological settings: Machine learning approaches and application to the Xihu sag

Abstract

Traditional laboratory methods for determining rock internal friction angles are costly, time-intensive, and limited by core sample quality, yet few studies address this in oil and gas drilling contexts. This research introduces a groundbreaking machine learning framework to predict offshore downhole internal friction angles, leveraging data from the East China Sea’s Xihu Sag. Using four algorithms and six well logging parameters selected via Spearman correlation analysis, we pioneered a systematic comparison of sequential versus random data partitioning. Sequential partitioning markedly reduced model accuracy and generalizability, most notably for KNN, while random partitioning enhanced performance. The XGBoost model excelled, achieving over 99 % accuracy in real downhole predictions with random splitting, showcasing unmatched accuracy, robustness, and generalization. We applied these predictions to assess wellbore stability in the Pinghu Formation using ABAQUS, demonstrating practical utility. This cost-effective, efficient alternative to traditional tests fills a critical gap in drilling research, offering novel insights into data partitioning and model selection. These findings not only advance predictive methodologies in complex geological settings but also provide a robust reference for future studies, underscoring the framework’s scientific rigor and high-value contributions.