Cross-Well ROP Prediction Using Physics-Informed Machine Learning in Horizontal Wells

Abstract

Cross-well prediction of rate of penetration (ROP) is critical for enabling data-driven drilling optimization, yet transfer-ability across wells with varying geology remains a key challenge. We investigate cross-well ROP prediction using five real-world horizontal wells, treating each well as an independent unit. A unified set of mechanical and directional features is augmented with physics-informed quantities, including mechanical specific energy, logarithmic MSE, torque × RPM, and depth-rolling MSE, computed per well to prevent data leakage. Gradient-boosted decision trees are trained on one or multiple wells and evaluated on held-out target wells, with performance assessed via RMSE, MAE, and R2. Results indicate that cross-well transfer is viable when geological and strati-graphic conditions are compatible: section-aligned transfers achieve RMSE <12 ft/hr with R2 >0.96, whereas section-mismatched transfers show degraded performance. Physics-informed features substantially improve accuracy, particularly in Top→Top and Bottom→Bottom transfers, while ensemble models trained on multiple wells increase robustness when training wells are geologically similar to the target well. Depth-stratified analysis shows predictive performance improves with depth, reflecting increased geological continuity. Sensitivity analysis highlights that stable drilling regimes (medium RPM, moderate WOB) yield the most reliable predictions, while extreme operating conditions limit generalization. These findings demonstrate that cross-well ROP prediction is conditionally reliable, with success dependent on geological compatibility, depth, and operational context. Ensemble and physics-informed approaches enhance model robustness, but careful selection of training wells is essential to avoid negative transfer. This work provides practical guidance for field deployment of cross-well machine learning models and informs strategies for geology and operation-aware predictive drilling.