Digital Pore Pressure Prediction for Well Drilling Using Machine Learning in a Deep Shale Gas Field

Abstract Abnormal pressure and wellbore instability are the main challenges during drilling in shale gas reservoirs. Traditionally, the pre-drill predictions of pore pressure and wellbore stability are executed manually by geomechanics engineers. The procedures are usually complicated and take time, the results also highly depend on the executor's expertise. All these make pore pressure and wellbore stability prediction far from ideal and automatic. In this study, we utilized machine learning methods to perform prediction in a simpler manner. The digital models were trained with existing well data, geology data and drilling data, and were correlated with spacing coordinates, that contains geological structure information. The models on formation materials properties are trained and learnt with patterns recognition; the pore pressure, earth stresses and wellbore stability are trained with physics-based hybrid algorithm. The trained models are then used to predict pore pressure and mud weight window at any point in subsurface or along any planned well trajectory, to identify drilling risks and recommend solutions. This approach was applied and validated in a deep shale gas field in Sichuan basin, China. In this field, the main shale gas reservoirs are overpressured and severe drilling complexities were encountered in drilling. Horizonal development wells are planned to drill to enhance production. This requires pre-drill pore pressure and wellbore stability prediction. Due to multiple abnormal pressure mechanisms and subsurface complexity, manual methodology is usually time-consuming, and the results are not consistent with different executors. With the developed new machine learning method, the digital models were trained with eleven geology surfaces and well data from eight existing wells. The trained model was used to predict pore pressure and mud weight window, including formation collapse pressure, mud loss pressure and breakdown pressure. The machine learning prediction of planned horizontal well Y14H and well Y15H were then compared against manual results calculated by geomechanics experts. The digital results matched well with manual results. The actual drilling results of well Y15H also confirmed the accuracy of the machine learning method. In well Y15H drilling, there were no drilling complexities and hole enlargements as using mud weight optimized with machine learning prediction. After well completed, the results showed that the pore pressure difference was only 0.5% between downhole measurement, 53.1MPa, and machine learning prediction, 52.8MPa. The minimum horizontal stress difference was about 5% between machine learning prediction, 72.88MPa, and downhole measurement, 76.76MPa. This field study confirmed the accuracy, effectiveness, and efficiency of machine learning method.