Stanford Geothermal WorkshopFebruary 9-11, 2026

This study presents a data-driven framework for modeling, interpreting, and forecasting heat recovery behavior in hydrothermal and fractured geothermal reservoirs. By combining scientific computing with unsupervised and supervised machine learning, this approach offers a robust methodology for extracting dominant patterns to understand various geothermal systems. Several machine learning techniques were used to provide insights on key geophysical and thermal transport dynamics in low-enthalpy hydrothermal reservoirs. Self-Organizing Maps (SOM) cluster normalized production temperature profiles into four temporal regimes (early, early-intermediate, late-intermediate, and late). Non-negative Matrix Factorization with K-means (NMFK) extracts five latent temperature signatures tied to these regimes. Supervised models (XGBoost, Random Forest, and Deep Neural Networks) are trained on each regime, and interpreted using SHAP analysis to identify the influence of key features such as temperature ratio, thermal Peclet number, fluid expansion, and reservoir geometry on the thermal regimes. Next, the ML workflow is applied to varying fracture network configurations representing enhanced geothermal systems (EGS). Each model incorporates deterministic and stochastic fractures, dynamic well placements, and consistent thermal-fluid-rock properties. Simulation outputs including temperature and pressure data from fracture and matrix zones are processed with NMFK to reveal dominant temporal signatures. These patterns differentiate between fracture-dominated regimes, with early thermal breakthrough, and matrix-buffered regimes characterized by gradual heat conduction. The proposed framework integrates the understandings from dimensionless subsurface transport phenomena, machine learning techniques, and predictive modeling for geothermal reservoir characterization and sustainable energy development.

Topic: Modeling