Accelerating Porous Media Flow Simulations With Fourier Neural Operators: An Application to Geologic Storage of CO2

Abstract

This study aims to develop surrogate models to accelerate decision-making processes related to porous media flows, using geologic storage of carbon dioxide ($C{O}_2$) as an example. Several engineering problems, including selection of subsurface $C{O}_2$ storage sites, often requires costly and complex simulations of flow fields. In this work, a Fourier Neural Operator (FNO) based model is developed for real-time, high-resolution simulation of $C{O}_2$ plume migration. The model is trained on a comprehensive dataset derived from realistic subsurface parameters and achieves a computational speed-up of $\mathcal{O}({10}^3\mathrm{to}{10}^5)$ when compared to numerical simulators used in this work, with only a minimal reduction in predictive accuracy. Super-resolution experiments are also investigated to reduce the computational cost of training the FNO-based models. Additionally, various strategies are proposed to enhance the reliability of model predictions, which is crucial for evaluating actual geological storage sites. This framework, based on NVIDIA's PhysicsNeMo library, enables rapid screening of sites for CCS. This work scales data-driven models to realistic 3D systems that better reflect real-life subsurface aquifers and reservoirs, paving the way for building next-generation digital twins for subsurface CCS applications. The workflows and strategies discussed can be easily adapted to other material systems and energy solutions, such as geothermal reservoir modeling, flow batteries, fuel cells, and hydrogen storage.