Prediction of Geothermal Reservoir Pressure and Temperature Using DataDriven, Physics-Informed, and Hybrid Neural Surrogates

Abstract

Comparisons between data-driven, physics-informed, and hybrid surrogate models are frequently complicated by variations in data, model size, or test splits for uncertainty analysis and optimization of geothermal systems. This study compares in a controlled way the simulation of a reduced dimensionless 3D thermo-hydro-mechanical problem. The pressure and temperature change in a coupled storagediffusion system and a scalar displacement potential provides a reduced mechanical closure. The parameter scenarios are calculated 48 times per run on 9 x 9 x 6 grid at 10 times. Complete scenarios are split into 32 training, 8 validation and 8 testing cases. Three random seeds are used when training the supervised coordinate network, the physics-informed neural network, and the hybrid residual-correction network, which are all based on a similar multilayer perceptron. In the hybrid model, mean relative L2 error of pressure is 0.0134, temperature is 0.0121, and mechanical potential is 0.0134. It will also produce the minimum normalized discrete residues. The pure network gets relative errors ranging from 0.2200 to 0.3022 for the field dependent. With the reported CPU cost, it turns out that the inclusion of automaticdifferentiation residual penalties does not help in generalization: the physics-informed model results in errors ranging from 0.2472 to 0.3246 and a longer training time. These results demonstrate that such a low fidelity physics prior can be more useful than just a residual penalty for this smaller benchmark. Geothermal reservoir, thermo-hydro-mechanical coupling, neural surrogate, physics-informed neural network, hybrid modelling, reproducibility.