With the energy demand arising globally, geothermal recovery by Enhanced Geothermal Systems (EGS) becomes a promising option to bring sustainable energy supply along with mitigating 𝐶𝑂2 emission. However, reservoir management of EGS primarily relies on reservoir simulation, which is quite expensive due to the reservoir heterogeneity, the interaction of matrix and fractures, and the intrinsic multi-physics coupled nature. Therefore, a robust optimization framework is critical for the management of EGS. We develop a general Physics-Informed Machine Learning (PIML) framework for reservoir management with multiple optimization options. A robust forward surrogate model 𝑓𝑙 is developed based on a convolutional neural network, and it successfully learns the nonlinear relationship between input reservoir model parameters (e.g., fracture permeability field) and interested state variables (e.g., temperature field and produced fluid temperature). 𝑓𝑙 is trained using simulation data from the EGS coupled thermal-hydro simulation model by sampling reservoir model parameters. As 𝑓𝑙 is accurate, efficient and fully differentiable, EGS thermal efficiency can be optimized following two schemes: (1) training a control network 𝑓𝑐 to map reservoir geological parameters to reservoir decision parameters by coupling it with 𝑓𝑙 ; (2) directly optimizing the reservoir decision parameters based on coupling the existing optimizers with 𝑓𝑙 . We evaluate the impact of reservoir model parameters on the thermal recovery based on simulation datasets through sensitivity analyses, and demonstrated that injection mass rate dominates thermal recovery. Further, the forward model 𝑓𝑙 performs accurate and stable predictions of evolving temperature fields (relative error 1.27 ± 0.89%) in EGS and the time series of produced fluid temperature (relative error 0.26 ± 0.46%), and its speedup to the counterpart high-fidelity simulator is 4564 times. When optimizing with 𝑓𝑐 , we achieve thermal recovery with a reasonable accuracy but significantly low CPU time during inference, 0.11 𝑠𝑒𝑐𝑜𝑛𝑑𝑠/task. When optimizing with Adam optimizer, we achieve the objective perfectly with relatively high CPU time, 4.58 𝑠𝑒𝑐𝑜𝑛𝑑𝑠/task. This is because the former optimization scheme requires a training stage for 𝑓𝑐 but its inference is non-iterative, while the latter scheme requires an iterative inference without training. We also investigate the option to use 𝑓𝑐 inference as an initial guess for Adam optimizer, which decreases Adam’s CPU time but achieves excellent convergence in the objective function. This is the highest recommended option among the three evaluated. The efficiency, scalability and accuracy observed in our reservoir management framework makes it highly applicable to near real-time reservoir management in EGS as well as other similar system management processes.