Data-driven prediction of drilling fluid rheology: a hybrid machine learning approach

Accurate estimation of rheological properties are important for cuttings transport, pressure control, wellbore stability and operational safety. Early detection of sudden changes in rheological behavior can help prevent costly operational problems. But laboratory measurement of rheological property is time-consuming and typically performed only once or twice daily. This study presents machine learning framework for predicting key drilling fluid rheological properties such as Plastic Viscosity (PV), Yield Point (YP), and Gel Strengths (10s and 10 m), from three routinely measured inputs: Marsh Funnel Viscosity (MFV), solids percentage and fluid density. Seven single and ensemble architectures were compared using 388 observations from nine gas and geothermal wells within water-based mud system. Model performance was assessed in terms of predictive accuracy and computational efficiency. Hybrid Stacking achieved a R² of 0.963, compared with 0.902 for the single-model SVM–PSO comparator. The differences among Hybrid Stacking (0.963), Full Stacking (0.962), and Stacked GBM + RF (0.958) were too small to support a meaningful ranking. SHapley Additive exPlanations (SHAP) identified MFV as the dominant predictor of PV, YP, and Gel Strength (10 m); for Gel Strength (10s), Density and MFV are comparable only in the Full Stacking model. The study goes beyond earlier task-specific models by providing a controlled comparison of seven architectures for simultaneous direct prediction of four rheological properties under one evaluation pipeline. Further validation using unseen wells and prospective field data is required before operational deployment.