Surrogate Modeling for Geothermal Systems: Accelerating Optimization, History Matching, and Uncertainty Quantification

Abstract

High-fidelity geothermal reservoir simulators are essential for forecasting subsurface pressure and temperature distributions but are computationally prohibitive for rapid optimization, history matching, and uncertainty quantification. Surrogate modeling offers a practical alternative by learning simulator input–output mappings and delivering fast predictions that can be embedded in engineering decision loops. In this study, we present a systematic comparison of surrogate models for steady-state geothermal applications, evaluating convolutional encoder–decoder architectures (U-Net), neural operator approaches (FNO, U-FNO, and Fourier-MIONet). Models are assessed using decision-relevant criteria, including predictive accuracy, physics consistency, generalization to unseen geologic heterogeneity, data efficiency, and computational cost. Our results demonstrate that neural operator–based surrogates achieve strong predictive performance from limited training data while reducing inference time by orders of magnitude relative to full-physics simulation. We highlight key trade-offs between surrogate architectures and provide practical guidelines for selecting models that balance fidelity, robustness, and speed in geothermal optimization, history matching, and uncertainty quantification workflows.