Field Calibrated Dual Porosity THM Modeling and Physics-Guided Machine Learning for Probabilistic Enhanced Geothermal System Forecasting at Utah FORGE

Abstract

Enhanced Geothermal Systems (EGS) depend on connected fractures for fluid circulation and heat extraction, but those same pathways control pressure buildup, thermal contraction, transmissivity evolution, stress redistribution, and the potential for induced slip. This study develops a field anchored, two-dimensional reduced order thermo-hydro-mechanical response model of the Utah Frontier Observatory for Research in Geothermal Energy (Utah FORGE) injector-producer corridor and couples it with audited machine-learning surrogates for probabilistic geothermal-performance and fracture-stability screening. Public Well 58-32 geophysical logs, X-ray-diffraction mineralogy, thermal-conductivity measurements, and processed 2024 Wells 16A/16B circulation records are integrated through an explicit evidence hierarchy, while published fracture, thermal, and stress information constrains the model domain and parameter ranges. The parent model represents thermal decline, quasi-steady pressure support, thermo-poroelastic stress redistribution, fracture-aperture evolution, cubic-law permeability feedback, and orientation-dependent frictional criticality. Two 20-year operating cases are evaluated: a prospective 60 kg/s injection case and a 26.5 kg/s field-rate reference case. A 2,500-realization physics-generated ensemble is used to train surrogate models for five independent outputs, followed by 50,000 Monte Carlo evaluations. The apparent wellhead pressure-difference/rate index decreases from 0.688 to 0.650 MPa/(L/s), equivalent to a 1.89% cubic-law aperture increase if the full trend is attributed to fracture transmissivity. At Year 20, the 60 kg/s case predicts a production temperature of 180.3°C and thermal power of 32.7 MW, compared with 183.8°C and 14.8 MW for the lower-rate case. The higher-rate case produces a maximum aperture change of 8.10%, a maximum permeability ratio of 1.263, and a friction ratio of approximately 1.07 for the south-striking, moderately west-dipping fracture set, whereas the lower-rate case remains below the simplified frictional threshold. Independent 750-case validation yields (R^2) values of 0.997 for production temperature, 0.9996 for proxy injection pressure, 0.9992 for aperture change, and 0.9958 for thermal-breakthrough time. Maximum friction ratio is less transferable ((R^2=0.605)); therefore, geomechanical Monte Carlo outputs are interpreted as conditional screening results rather than seismic-hazard probabilities. The study establishes a traceable workflow in which real observations constrain a reduced-order THM replica, physics-based simulations generate the training manifold, ML enables rapid uncertainty propagation, and probabilistic analysis defines a heat-recovery-versus-mechanical-criticality operating envelope for risk-aware EGS design.