Data-Driven Prediction of Rheological Properties in Invert Emulsion Drilling Fluids: A Comparative Analysis of Machine Learning Models

Abstract Invert emulsion drilling fluids (IEDFs) are water-in-oil emulsions, where the internal phase typically consists of calcium chloride brine, and the external phase is a non-aqueous fluid. These fluids are extensively utilized in drilling operations for challenging shale formations due to their superior shale inhibition, excellent lubrication, and high-temperature stability. The rheological properties of drilling fluids play a crucial role in drilling hydraulics, influencing operational safety and hole cleaning efficiency. While rheological measurements are routinely performed by drilling fluid engineers on-site, real-time prediction of these properties can significantly enhance drilling precision and operational efficiency. In this study, a dataset including 1,701 observations and 11 critical properties such as flowline temperature (FLT), mud weight (MW), funnel viscosity (FV), plastic viscosity (PV), yield point (YP), high pressure high temperature (HPHT) fluid loss, whole mud alkalinity (Pom), electrical stability (ES), CaCl2 wt%, oil-water ratio and water, oil and corrected solid volume fractions collected from 21 wells in 3 different regions were used. Rheological properties in the dataset were measured at 150 °F with a rotational viscometer. The dataset's quality and reliability are enhanced by eliminating irrelevant, noisy, or inconsistent information. The outlier detection was performed using the Quartiles method with a threshold factor 1.5. Three feature selection algorithms, F-Test, MRMR, and RReliefF, were used for predicting key rheological properties—Plastic Viscosity (PV) and Yield Point (YP)—in IEDFs. The model's performance was assessed through various statistical metrics, including Coefficient of Determination (R2) and Mean Absolute Percentage Error (MAPE). The RReliefF algorithm outperformed other feature selection methods in improving model performance in predicting PV and YP. Among the models tested, Gaussian process regression showed superior accuracy in predicting PV, while machine learning models showed lower performance in predicting YP.