Accelerating Geothermal Modeling with Low- and High-Fidelity Fourier Neural Operators

Geothermal reservoir models are costly to build and calibrate, and generating a single forecast can take hours. Operators tasked with field planning and optimization are constrained by the speed of these forecast simulations, limiting the number of scenarios they can explore. Machine learning can be a powerful tool to speed up computationally expensive tasks, but standard approaches using Neural Networks (NNs) struggle to produce accurate results and require thousands of high-fidelity simulation outputs for training, an impractical expense in geothermal modeling workflows. Fourier Neural Operators (FNOs) have recently emerged as a compelling alternative, outperforming NNs in many physical modeling tasks. FNOs have been shown to produce fast and accurate first-order estimates of oil and gas reservoirs with high spatial fidelity. Furthermore, FNOs can generalize between multiple resolutions, allowing surrogate models trained on low-resolution simulation data to make high-resolution predictions. We present a FNO surrogate model of a geothermal reservoir at multiple resolutions, including natural state production phase simulations. Once trained, these surrogate models generate results in ~400 ms, whereas actual simulations require ~5 minutes, enabling rapid scenario testing and accelerated model calibration. These tools are intended to compliment, rather than replace, traditional geothermal simulation workflows and enable faster, more flexible reservoir forecasting.