Stanford Geothermal WorkshopFebruary 9-11, 2026

This study applies supervised machine learning using Extreme Gradient Boosting (XGBoost) with physics-informed formulations to predict fracture intensity (P32) and fracture aperture in deep geothermal reservoirs using drilling and logging-while-drilling data from the Utah FORGE project. To reduce reliance on costly image logs and enable real-time, ahead-of-bit fracture characterization, we apply a depth-based machine learning (ML) prediction strategy that trains the model on shallow-depth data and predicts fracture properties across the remaining well interval. This approach relies on extrapolative, rather than interpolative, predictions and therefore involves a trade-off between prediction accuracy and logging requirements compared to traditional machine learning approaches that depend on full-length wellbore logs. Model performance is evaluated against conventional machine learning approaches, including baseline models without physics-informed formulations, and further compared under scenarios with and without physics-informed features, as well as with or without reduced feature sets. Results show that training the model on approximately 50-60% of the shallow well interval is sufficient to achieve reliable ahead-of-bit predictions of fracture intensity and aperture. Additionally, wavelet-based feature transformations enhance predictive accuracy, and the inclusion of physics-informed formulations further improves performance, particularly in predicting fracture intensity.

Topic: FORGE