Machine Learning–Based Rate of Penetration Prediction Using Multi-Well Field Data in Deviated Wells

Abstract

Reliable forecasting of drilling rate of penetration (ROP) remains a technical challenge due to its complex dependence on several operational, geological, and directional conditions that are highly nonlinear and well-specific. These difficulties are amplified in deviated wells, where traditional empirical relations often fail to capture the combined effects of lithological variability and directional changes. This study develops a machine learning (ML) model for ROP prediction using multi-well field data and evaluates candidate ML algorithms/models not only on numerical accuracy but also on their ability to reproduce reasonable trend responses to changes in key drilling parameters. The objective is to establish a validation approach that integrates statistical performance, cross-well generalization, and sensitivity behavior consistent with drilling mechanics. Field data from 18 wells were used to develop the model, while two additional wells were reserved for independent validation. A total of 91,140 drilling datasets, comprising 24 parameters/features, were collected. After preprocessing and selective outlier screening, 85,695 records were retained for model training and testing. Feature selection resulted in nine highly influential parameters: True Vertical Depth (TVD), Weight on Bit (WOB), rotational speed (Rotation), mud flow rates (FLOW), mud density, shale content (Shale), Sandstone content (Sandstone), inclination (Inc), and dogleg severity (DLS). Eleven supervised ML algorithms were evaluated, representing instance-based, tree-based, boosting-based, neural-network, and stacking-ensemble model families. Model performance was assessed using feature-importance ratings and statistical metrics, including train/test R2, MSE, RMSE, and MAE, which were used to select the best-performing model. Finally, sensitivity analysis was conducted to evaluate the selected model's ability to reproduce physically meaningful ROP responses. Model comparison across the training and testing splits demonstrated strong overall predictive performance, with training R2 values ranging from 0.91 to 0.99 and testing R2 values ranging from 0.88 to 0.94. Among the evaluated models, the Gradient Boosting (GB) model provided the best overall performance, with training R2 = 0.9940, testing R2 = 0.9399, RMSE = 36.88, and MAE = 21.52. The GB model also showed strong agreement with the averaged feature-importance ranking and reproduced physically meaningful ROP sensitivity trends for the dominantly influential drilling parameters. SHAP and feature-importance analyses confirmed that TVD, FLOW, WOB, shale content, and rotation were the most influential variables controlling ROP. External validation on two unseen wells further demonstrated strong generalization, with R2 = 0.9694 for Well-19 and R2 = 0.96 for Well-20. This study presents a physically informed framework for evaluating and selecting data-driven ML models for deviated wells. The approach combines conventional accuracy metrics with feature-importance consistency, reasonable trend prediction in sensitivity analysis, and strong agreement between model predictions and unseen-well data. Through this integrated evaluation, a model is identified that not only reproduces historical data accurately but also yields correct responses under a diverse range of varying input parameters. The proposed methodology establishes a reproducible basis for developing more reliable ROP forecasting tools for complex well trajectories.