A Robust General Physics-Informed Machine Learning Framework for Energy Recovery Optimization in Geothermal Reservoirs

Abstract

Energy extraction from the Enhanced Geothermal System (EGS) relies on hydraulic fractures or natural fractures to migrate fluid and thus extract heat from surrounding rocks. However, due to the heterogeneity and complex multi-physics nature inside of fracture plane, high-fidelity physics-based forward simulation can be computationally intensive, creating a barrier for efficient reservoir management. A robust and fast optimization framework for maximizing the thermal recovery from EGS is needed. We developed a general reservoir management framework which is combining a low-fidelity forward surrogate model (fl) with gradient-based optimizers to speed up reservoir management process. Thermo-hydro-mechanical (THM) EGS simulation model is developed based on the finite element-based reservoir simulation. We parameterized the fracture aperture and well controls and performed the THM simulation to generate 2500 datasets. Further, we trained two different architectures of deep neural network (DNN) with the datasets to predict the dynamics (pressure and temperature), and this ultimately becomes the forward model to calculate the total net energy. Instead of performing optimization workflow with large amount of simulations from fh, we directly optimize the well control parameters based on geological parameters to the fl. As fl is efficient, accurate and fully differentiable, it is coupled with different gradient-based or gradient-free optimization algorithms to maximize the total net energy by finding the optimum decision parameters. Based on the simulation datasets, we evaluated the impact of fracture aperture on temperature and pressure evolution, and demonstrated that the spatial fracture aperture distribution dominates the thermal front movement. The fracture aperture variation is highly correlated with temperature change in the fracture, which mainly results from thermal stress changes. Compared to the full-fledged physics simulator, our DNN-based forward surrogate model not only provides a computational speedup of around 1500 times, but also brings high predictive accuracy with R2 value 99%. With the aids of the forward model fl, gradient-based optimizers run optimization 10 to 68 times faster than the derivative-free global optimizers. The proposed reservoir management framework shows both efficiency and scalability, which enables each optimization process to be executed in a real-time fashion.