Leveraging Artificial Intelligence for Field-Scale Thermal Conductivity Prediction in the UTAH FORGE Geothermal System

Abstract Predicting thermal conductivity in Enhanced Geothermal Systems (EGS) is a complex challenge due to the dynamic interplay of thermal, hydraulic, mechanical, and geological factors that vary significantly with depth. Traditional approaches such as core-based measurements and numerical simulations, while effective, are often constrained by high costs, limited spatial coverage, and computational intensity. This challenge is especially apparent at the Utah FORGE site, where no field-wide thermal conductivity profile currently exists—data is only available from the 58-32 pilot well. However, an accurate understanding of thermal properties and heat extraction across the entire field is essential for effective reservoir characterization, stimulation design, and long-term energy extraction planning. To fill this gap, this study presents a validated machine learning (ML) framework that integrates drilling and petrophysical log data to predict thermal conductivity at scale and can be useful for the entire structure. Five ML models were developed and evaluated: Random Forest (RF), Extreme Gradient Boosting (XGBoost), Decision Tree (DT), Artificial Neural Network (ANN), and a Hybrid ensemble combining DT, RF, and XGBoost. Input features included petrophysical logs—gamma ray, porosity, sonic velocity—and drilling data such as weight on bit, rate of penetration, rotary speed, and fluid flow rate. The dataset comprised information from three wells: 58-32 pilot and water, 16A (78)-32, and 16B (78)-32. Models were trained and validated using an 80:20 train-test split and optimized via cross-validation to ensure robust generalization. Among all models, Random Forest delivered the best performance with an R² of 0.994 and RMSE of 0.0355, exhibiting smooth prediction profiles and strong alignment with experimental values. The Hybrid ensemble model further enhanced prediction stability, particularly at greater depths. In contrast, XGBoost performed well in data-dense regions but introduced noise in deeper zones. Decision Tree showed acceptable results near the surface but suffered from overfitting, and ANN, while able to model complex relationships, showed higher residual variance. Feature importance and SHAP analysis revealed that engineered parameters—such as gamma ray–porosity interactions (GR-SPHI-Interaction) and porosity-to-sonic ratios (SPHI/DTCO)— contributed significantly to model performance. These results highlight the value of AI-driven tools in enhancing geothermal reservoir characterization and supporting future field-scale planning. Overall, this study highlights how scalable and interpretable ML models—particularly Random Forest and Hybrid ensembles—can bridge critical data gaps and support high-resolution thermal characterization for geothermal field development and stimulation planning at Utah FORGE.