What question did this study set out to answer?

The aim is to predict torque and drag forces in directional drilling using machine learning methods.

March 13, 2026

Real-Time Prediction of Drilling Torque and Drag Using XGBoost on MWD Data from Directional Wells

Key Points

The aim is to predict torque and drag forces in directional drilling using machine learning methods.
Utilized Measurement While Drilling (MWD) data from directional wells.
Implemented XGBoost, Random Forest, and Support Vector Regression algorithms.
Applied data preprocessing techniques including Z-score outlier removal and MinMax scaling.
Used an 80/20 train-test split and 5-fold cross-validation to validate model performance.
XGBoost achieved an R2 of 0.9235 for torque and 0.9762 for drag.
Random Forest attained R2 values of 0.9253 for torque and 0.9749 for drag.
Support Vector Regression obtained R2 values of 0.9283 for torque and 0.9781 for drag.
Feature importance analysis indicated WOB, inclination, and RPM were crucial predictors.

Abstract

The exact prediction of the torque and drag forces is necessary to improve directional drilling operations and reduce non-productive time in the complicated wellbore environments. Nonlinear relationships between downhole forces and parameters of drilling are however complex and not properly modeled using conventional empirical frameworks. This reduces the effectiveness in drilling process. The study discusses the design of a machine learning model, which uses Measurement While Drilling (MWD) data of directional wells in determining the real-time torque and drag. It deployed and tested three superior algorithms, namely, XGBoost, random forest (RF), and support vector regression (SVR). Our well data points were used in this assessment and the input features used were inclination, azimuth, hookload, Weight on Bit (WOB) and Revolutions per Minute (RPM). Our outlier removal was implemented based on Z-score cut-offs, we normalised the features with MinMax scaling, and we down-sampled with 30s temporal intervals. An 80 by 20 train-test split in conjunction with 5-fold cross-validation was used to validate the model. XGBoost demonstrated good performance in predicting torque where it attained a R2 of 0.9235 and an RMSE of 2.059 Nm. It had a R2 of 0.9762 and RMSE of 5.294 kN in terms of prediction of drag. The performance of the Random Forest model was also comparable and the model attained R2 of 0.9253 in torque and 0.9749 in drag. Conversely, Support Vector Regression (SVR) technique had R2 values of 0.9283 and 0.9781 in predicting torque and drag respectively. Further, the feature importance analysis indicated WOB, inclination, and RPM to be important predictors of both targets. The SHAP analysis conducted (SHAPley Additive Explanations) provided the contribution of each feature and thus increased the level of transparency of the model. This technology makes real-time drilling optimisation possible through the use of field-deployable computational efficiency that is used to minimise the effect of non-productive time and enhance operational safety.

Bookmark

Real-Time Prediction of Drilling Torque and Drag Using XGBoost on MWD Data from Directional Wells

Key Points

Abstract

Cite This Study