Well Trajectory Effect on Robust Prediction of Rate of Penetration During Unconventional Oil Well Drilling
Accurate prediction of the rate of Penetration (ROP) is critical for optimizing drilling efficiency and reducing well construction costs in unconventional resource development. A significant gap exists in the overwhelming majority of machine learning (ML) ROP models, which have been trained exclusively on vertical well data. This systematically neglects the physical influence of wellbore geometry in deviated and horizontal wells that dominate modern unconventional drilling. However, this study addresses this gap by developing and rigorously comparing three ML algorithms: Artificial Neural Network (ANN), Gradient Boosting Machine (GBM), and Random Forest (RF), for real-time ROP prediction in a non-vertical unconventional shale well. A preprocessed dataset of 22,486 data points from the daily drilling records from a United States (U.S.) shale well was used. Eleven input features spanning four categories were employed: wellbore trajectory, depth-related, mechanical/operational parameters, and hydraulic parameters. Models were trained on a 70:15:15 train-validation-test split and evaluated using the coefficient of determination (R2), root mean square error (RMSE), and mean absolute error (MAE). All three models achieved strong predictive accuracy: ANN (R2 = 0.8780, RMSE = 0.06924), GBM (R2 = 0.9134, RMSE = 0.05833), and RF (R2 = 0.9961, RMSE = 0.01237). The RF model was superior and selected for feature importance analysis. An explicit ANN Equation was derived from extracted weights and biases, enabling model transparency and software-independent computation of ROP. Feature importance analysis on a categorical basis showed that mechanical/operational parameters were the most influential predictors, followed by depth-related, then trajectory parameters. A systematic literature review showed that in vertical conventional wells, WOB and bit speed consistently dominate as the primary ROP drivers. However, in the directional unconventional context of this study, azimuth ranked fifth and inclination seventh, together contributing 14.5%, which exceeds all hydraulic parameters. This result underscores the previously neglected role of wellbore geometry in ROP prediction for deviated wells. However, the findings in this study are limited to a single-well dataset from one U.S. shale formation, an inclination range of 0.18°–55.92° without a full horizontal lateral, and the absence of formation evaluation logs.