Machine Learning-Driven Prediction of Filtration Properties for Drilling Fluids: Enabling Real-Time Monitoring and Automation

Abstract Drilling fluids play a crucial role in ensuring wellbore stability, cooling drill bits, and transporting cuttings to the surface. Among the various properties of drilling fluids, filtration behavior is particularly significant, as it directly affects wellbore integrity, formation damage, and overall drilling efficiency. These effects are further amplified under high-pressure, high-temperature (HPHT) conditions, where fluid loss control becomes more complex due to extreme environmental factors. Traditionally, filtration properties are assessed using laboratory tests such as the API filter press, which, despite their accuracy, are time-consuming and conducted intermittently. The infrequent nature of these tests delays real-time monitoring and decision-making, increasing the risk of drilling complications. To address this limitation, machine learning (ML) presents a powerful alternative, enabling continuous, data-driven predictions of filtration properties based on readily available drilling parameters. This study aims to develop a machine learning-driven predictive model capable of accurately estimating the filtration properties of oil-based drilling fluids in real-time. The objective is to integrate this predictive framework into drilling operations, reducing reliance on labor-intensive laboratory testing while enhancing real-time monitoring and decision-making. A comprehensive dataset comprising operational drilling records from multiple wells in a Middle Eastern field was utilized to develop the predictive model. Key input parameters, including Solid Content, Oil-to-Water Ratio (O/W), Mud Weight (MW), Funnel Viscosity (FV), and the output feature: Filtration Volume (Filt. Vol.) that were recorded under HPHT conditions (500 psi, 250°F) to capture realistic drilling scenarios. The data underwent rigorous preprocessing, including outlier removal using the Z-score method and skewness correction via logarithmic transformation, ensuring its quality and reliability for model training. A Random Forest Regressor was selected as the predictive model due to its superior performance in handling complex, nonlinear relationships. The model was optimized through hyperparameter tuning using a grid search approach combined with k-fold cross-validation to enhance its predictive accuracy and generalizability. The optimized machine learning model demonstrated strong predictive capability, achieving a correlation coefficient (CC) of 0.97 and a mean absolute percentage error (MAPE) of 3.94% in the training phase. When tested on unseen data, the model maintained its reliability, yielding a CC of 0.98 and a MAPE of 3.23%, highlighting its robustness in predicting filtration properties across diverse conditions. Feature importance analysis revealed that Solid Content, O/W ratio, and MW were the most influential parameters in determining filtration volume, providing valuable insights into fluid behavior under HPHT conditions. The successful development of an ML-driven predictive model for filtration properties presents a transformative step toward real-time drilling fluid monitoring. By integrating this model into drilling workflows, operators can optimize fluid performance, mitigate formation damage, and proactively manage wellbore stability without the delays associated with traditional laboratory testing.