Enhanced geothermal systems (EGS) can provide a sustainable and renewable solution to the new energy transition. Its potential relies on the ability to create a reservoir and to accurately evaluate its evolving hydraulic properties to predict fluid flow and estimate ultimate thermal recovery. Here we develop a hybrid machine learning (ML) model to predict permeability evolution of intermediate-scale (~10 m) hydraulic stimulation experiments at the Sanford Underground Research Facility (the EGS Collab project). We present a 3D map in situ permeability evolution for two of stimulation episodes in this project using microearthquakes (MEQs) data and injection histories of wellhead pressure and flow rate. This map includes both average reservoir permeability evolution over time and local fracture permeability distribution within the evolved reservoir. Compared with the ground truth of average permeability calculated from the well data, our predicted average permeability for these two episodes has a MSE value less than 2.9E-4 and R2 higher than 0.93, indicating that average permeability predicted by machine learning is consistent with good agreement with field observation. Additionally, distributed fracture permeability calculated by empirical equation over time shows the process of fracture propagation and identify the potential fluid path for geothermal reservoir.