Machine learning approach for prediction of safe mud window based on geochemical drilling log data

Abstract

Background: Accurate prediction of the safe mud window (SMW) is critical for drilling operations to prevent costly risks such as blowouts, mud loss, and wellbore instability. Traditional geomechanical methods for SMW determination face challenges in handling complex, nonlinear relationships within drilling datasets. Purpose: This study aims to develop robust machine learning (ML) models to predict two key SMW parameters—Mud Pressure below shear failure (MWsf) and tensile failure (MWtf)—using geochemical drilling log data from Middle Eastern carbonate reservoirs. Methods: Hybrid ML models combining Least Squares Support Vector Machine (LSSVM) and Multilayer Perceptron (MLP) with optimization algorithms (Gray Wolf Optimization, GWO; Grasshopper Optimization Algorithm, GOA) were trained on 2,820 data points from three wells. Input variables included drilling time, caliper, weight on bit, flow rate, and rheological properties. Model performance was evaluated using RMSE, R2, and cross-validation. Results: The LSSVM-GWO model outperformed others, achieving RMSE values of 58.01 (MWsf) and 95.42 (MWtf) with R2 > 0.99. Flow speed, rotor solids, and fan readings strongly influenced MWsf, while WOB, gel strengths, and flow rate impacted MWtf. Generalization testing on a third well confirmed robustness (RMSE: 50.26 for MWsf, 70.89 for MWtf). Conclusion: The LSSVM-GWO framework provides a reliable, data-driven solution for SMW prediction, enabling safer and more efficient drilling operations. This approach reduces operational risks and highlights the potential of hybrid ML models in reservoir management.