Efficient Geothermal Reservoir Simulation Using Deep Learning Surrogates and Multiscale Interpolation Techniques

Accurate prediction of subsurface temperature distributions is essential for geothermal reservoir assessment, thermal performance evaluation, and decision support in reservoir management. However, repeated high-resolution numerical simulations are computationally expensive, particularly when multiple scenarios, heterogeneous petrophysical fields, and varying grid resolutions must be analyzed. This study presents a U-Net-based surrogate modeling framework for fast geothermal temperature field prediction on structured grids, coupled with interpolation strategies for handling unseen grid resolutions and intermediate time instances. Training and evaluation data are generated using the MATLAB Reservoir Simulation Toolbox (MRST) (24.1.0.2578822 (R2024a) Update 2) under multiple porosity–permeability realizations and at several grid resolutions (130 × 73, 67 × 37, 36 × 19, and 20 × 11) on a 2D grid. Data preprocessing and reshaping techniques are used to preserve spatial correspondence across resolutions. For fixed trained grids, the surrogate directly predicts temperature fields from porosity, permeability, and time inputs. For unseen grids, a grid interpolation strategy combines predictions from neighboring trained resolutions using weighted blending based on target grid cell count, followed by spatial resizing to the requested resolution. In addition, time interpolation is used to estimate temperature maps at intermediate time steps between predicted/simulated snapshots. The proposed framework enables rapid generation of temperature maps while maintaining spatial structure, making it suitable for efficient geothermal screening and multiscale scenario analysis.