Machine-learning models for classification of mudcake thickness in oil-based drilling fluids

Abstract

The performance of oil-based drilling fluids (OBDFs) is critical for maintaining drilling efficiency under conditions in which water-based fluids fail to perform effectively. Mudcake or filtercake thickness (MCT) is a key property of OBDFs because it directly influences wellbore stability and the extent of formation damage. Conventional methods for determining MCT rely on manual laboratory measurements that are time-consuming and cannot represent real-time downhole conditions, thereby limiting their value for operational decision-making and motivating the development of reliable predictive methods. This study proposes a novel framework for the near-real-time classification of MCT in OBDF systems. A comprehensive field dataset was compiled from five routinely measured, rapidly acquired laboratory parameters: mud weight, temperature, alkalinity, emulsion stability, and Marsh funnel viscosity. Four machine-learning classifiers—extreme gradient boosting (XGB), support vector machine, extreme learning machine, and multi-layer perceptron—were developed to classify MCT into three operationally meaningful categories: Excellent (≤1.0 mm), Good (1.0–2.2 mm), and Poor (>2.2 mm). The developed models achieved high classification accuracy, with the XGB algorithm demonstrating the most stable and reproducible performance (test accuracy: 0.98, F1-score: 0.98). Comprehensive evaluation—including receiver operating characteristic analysis, model confidence assessment, learning curves, group K-fold cross-validation, and external validation—confirmed the strong generalization capability of the XGB model, which achieved 94.81% accuracy when applied to completely unseen wells. Interpretability analysis using SHapley Additive exPlanations (SHAP) identified emulsion stability as the dominant predictive feature and revealed that parameter interactions (e.g., alkalinity–emulsion stability) vary dynamically across different MCT classes. The primary innovation of this study lies in reframing MCT prediction as a multi-class classification problem rather than a traditional regression task, enabling rapid and actionable interpretation of drilling fluid performance. The proposed framework provides a semi-automated, data-driven tool for near-real-time monitoring of MCT in OBDF systems, reducing reliance on slow laboratory measurements and supporting timely operational decisions that improve drilling efficiency and preserve wellbore integrity in complex drilling environments.