Thermal–Hydraulic–Mechanical Embed-to-Control-and-Observe (THM-E2CO): Toward Real-Time Surrogate Modeling of Geothermal Systems

Abstract

This study presents a deep learning framework, Thermal-Hydraulic-Mechanical Embed-to-Control-and-Observe (THM-E2CO), designed for high-fidelity reduced-order modeling of geothermal reservoirs governed by tightly coupled poro-thermo-elastic processes. The architecture extends the E2CO paradigm to address the complex nonlinear interactions between heat transfer, fluid flow, and rock deformation. Unlike static surrogate models, this framework is specifically engineered to predict and capture the temporal evolution of porosity and permeability as they respond to thermal drawdown and mechanical loading within the reservoir. The THM-E2CO architecture integrates a 3D convolutional encoder-decoder to compress and reconstruct full field pressure and temperature distributions while preserving critical thermal-hydraulic-mechanical couplings. A nonlinear latent-state transition network models the temporal evolution of the reservoir under varying operational controls. Crucially, the model is trained to observe the feedback loops where temperature gradients and pore pressure fluctuations drive mechanical deformation, subsequently altering the flow properties of the rock matrix. The learning process is regularized through a composite loss function that enforces data reconstruction fidelity, latent consistency, and conservation of mass and energy, ensuring physical plausibility across the coupled domains. Using CMG STARS as the high-fidelity simulator, a diverse ensemble of training data was generated via CMOST-driven variations in injection rates and producer bottom hole pressures. The dataset comprises over 1,000 simulations spanning a 30-year production period, featuring a six-well configuration. This wide operational space allows the model to learn robust coupled behavior and the nonlinear degradation or enhancement of reservoir transmissibility over time. The novelty of this work lies in the framework’s ability to explicitly capture the poro-thermo-elastic response of the reservoir, providing a rapid alternative to computationally expensive geomechanical solvers. By embedding the physics of stress-dependent permeability into the latent space, the model achieves high-fidelity reproduction of full physics simulations with computational accelerations exceeding 30,000 times. This development provides a scalable foundation for the prediction of rapid geothermal performance, uncertainty quantification, and real time optimization under complex geomechanical conditions.