Enhancing Geothermal Drilling Performance: Predicting Rate of Penetration with Machine Learning Utilizing Geomechanical and Petrophysical Data

Abstract

Abstract Optimizing drilling efficiency is crucial for enhancing geothermal energy production, with the rate of penetration (ROP) being a critical performance metric. Since drilling operations represent the largest financial expense in the geothermal energy cycle, precise ROP prediction is vital for cost reduction. However, current machine learning (ML) models depend exclusively on drilling parameters, frequently overlooking essential geomechanical and petrophysical data. Incorporating these datasets can substantially enhance prediction accuracy, operational efficiency, and economic sustainability. This study focuses on improving ROP prediction by incorporating geomechanical, petrophysical, and drilling data into advanced machine learning (ML) models using the FORGE dataset. Following outlier removal, feature importance algorithms were used to identify the top 10 influential parameters. Six ML algorithms—Random Forest, Extra Trees, XGBoost, Gradient Boosting, KNearest Neighbors, and Support Vector Machine—were trained and optimized through grid search and cross-validation. The models were then applied to predict ROP using the selected features, and their performance was evaluated by comparing results from drilling-only data versus the combined dataset. Furthermore, a Python-based web application was developed to enable real-time ROP prediction, enhancing decision-making and operational efficiency in drilling operations. The results revealed that integrating geomechanical and petrophysical data significantly improves the accuracy of ROP predictions. Feature importance analysis highlighted geomechanical parameters—particularly maximum horizontal stress, minimum horizontal stress, and vertical stress—as critical factors influencing model performance. Among the evaluated algorithms, Random Forest (RF) delivered the highest accuracy with an R2 of 0.87, whereas K-Nearest Neighbors performed the least effectively with an R2 of 0.81. When using only drilling parameters, the model achieved an R2 of 0.82, indicating lower accuracy compared to the average performance achieved with the combined dataset. These findings emphasize the importance of incorporating diverse data types into ML models to enhance ROP prediction. Additionally, the developed web application proved highly reliable, accurately replicating the study's results and demonstrating its practical utility in geothermal drilling operations. In this study, we introduce a systematic integration of geomechanical and petrophysical data with traditional drilling parameters, marking the first time such a comprehensive approach has been applied to ROP prediction. This integration significantly enhances prediction accuracy. Furthermore, we developed a real-time web application that serves as a bridge between research and industry, offering drilling engineers a robust tool for making data-driven decisions.