A shallow machine learning method based on geothermal drilling data: A case study of well 58–32 at the U.S. FORGE site

Abstract

Enhanced geothermal system (EGS) development relies on efficient drilling methods. Traditional physics-driven numerical simulation models, while effective, are computationally demanding due to the complexity of deep reservoir conditions. In contrast, data-driven models offer a simpler approach by deriving statistical features from training data to create predictive models. This study utilizes diagnostic drilling data from the U.S. FORGE site to train Back Propagation neural network (BPNN) and Random Forest regressor models. The Random Forest model, with a training time of 4.981 s, achieves a test R2 of 0.9995, surpassing the BPNN in computational efficiency by 91.82 % and accuracy by 1.04 %. By effectively explaining 99.95 % of drilling depth values using features such as maximum rate of penetration (ROP), pump pressure, torque, and flow in, the model demonstrates the potential of shallow machine learning on extensive datasets in geothermal energy development. This research not only validates the efficacy of data-driven models in optimizing drilling performance but also highlights their role in cost reduction for future drilling projects. The findings present valuable insights for the geothermal industry, paving the way for enhanced drilling strategies and improved resource management.