Explainable Machine-Learning-Based Prediction of Equivalent Circulating Density Using Surface-based Drilling Data

Abstract

When drilling wells for energy explorations, it is important to regulate the formation pressures appropriately to prevent kicks, which can lead to unimaginable loss of lives and properties. This is usually done by controlling the equivalent circulating density (ECD), which responds to the dynamic conditions that occur during drilling. The conventional approach to determine ECD is via mathematical modeling or downhole measurements. However, the downhole measurement tools can be very expensive, and the mathematical models do not provide a high degree of accuracy. Some previous authors have proposed using machine learning (ML) techniques to improve the degree of accuracy of the ECD predictions. In this work, we employed an extreme gradient-boosting (XGBoost) methodology to predict ECD values. The model's accuracy was determined using correlation coefficients (R2) and root mean square errors (RMSE) as their performance metrics. The results showed a strong prediction capability with an R2 and RMSE of 1.00 and 0.0005 for the training data and an R2 and RMSE of 0.989 and 0.023 for the testing/blind data set, respectively. Lastly, we used SHAP (SHapley Additive exPlanations) approach, a concept borrowed from cooperative game theory, to enhance the interpretability and the explainability of the model prediction values.