Data-Driven Prediction of Oil-Based Drilling Fluid Rheology: An Integrated Automated Machine Learning and Shapley Additive Explanations Interpretation Framework

Résumé fourni par la source

Summary Deep oil and gas resources have become a key strategic frontier for future reserve replacement and production growth. However, the complex downhole conditions in deep formations tend to reduce hole-cleaning efficiency and increase the risk of wellbore instability, making precise regulation of drilling fluid rheology essential for safe and efficient drilling operations. At present, field rheology control still relies largely on empirical judgment and therefore suffers from a clear time lag. Although data-driven methods can improve the efficiency of rheology prediction, existing models are generally limited by insufficient automation in model development, cumbersome hyperparameter optimization, and inadequate engineering interpretability, which restricts their use in accurate field prediction and sound operational decision-making. To address these limitations, this study proposes an intelligent prediction method for oil-based drilling fluid rheology by integrating an automated machine learning (AutoML) framework with Shapley additive explanations (SHAP). Built on the AutoGluon framework, the proposed method can automatically perform data processing, hyperparameter optimization, and multimodel-stacked ensembling, enabling the development of high-accuracy predictive models for apparent viscosity (AV), plastic viscosity (PV), and yield point (YP) with limited manual intervention. The results show that the optimal ensemble model, WeightedEnsemble_L3, achieved coefficients of determination (R2) of 0.875, 0.884, and 0.862 for AV, PV, and YP, respectively, on the independent test set, with corresponding root mean square error (RMSE) of 2.722, 2.359, and 0.612. Its predictive accuracy and generalization performance were superior to those of mainstream machine learning models, such as Bayesian-optimized categorical boosting (CatBoost), and AutoML frameworks, such as fast and lightweight AutoML (FLAML). Global SHAP interpretation showed that funnel viscosity, drilling fluid density (DY), and solids content were the dominant factors controlling rheological variation, with overall positive contributions. Local SHAP interpretation further quantified the contribution direction and relative weight of each feature under representative operating conditions, thereby improving the transparency and engineering credibility of the model predictions. This study provides a methodological basis for building accurate and trustworthy intelligent rheology-prediction systems for drilling fluids, with practical value for rheology risk warning and fine control under deep and complex drilling conditions.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.