Fourier-MIONet: Fourier-enhanced multiple-input neural operators for multiphase modeling of geological carbon sequestration

Overview

• Geologic carbon sequestration (GCS) is a technology that aims to reduce carbon dioxide in the atmosphere.
• Simulating multiphase flow in porous media, essential for understanding CO₂ migration and pressure fields in GCS, is computationally challenging.
• The authors develop a Fourier-enhanced multiple-input neural operator (Fourier-MIONet) to learn the solution operator for multiphase flow in porous media.
• Fourier-MIONet is more computationally efficient than existing models while maintaining similar prediction accuracy.

Plain English Explanation

The paper describes a new machine learning model called Fourier-MIONet that can simulate the movement of different fluids, like oil and gas, through a porous rock underground. This is an important problem for geologic carbon sequestration, which is a way to store carbon dioxide underground to reduce the amount in the atmosphere and slow climate change.

Simulating this fluid flow is very difficult using traditional computer models because the physics involved is highly complex and nonlinear. The Fourier-MIONet model uses a novel neural network architecture that can learn to predict the fluid flow much faster than existing methods, while still maintaining accurate results. This allows researchers to explore many different underground scenarios and optimize carbon sequestration processes more efficiently.

The key benefit of Fourier-MIONet is that it requires much less computational power and memory than other state-of-the-art models, making it practical for real-world applications. It can also make accurate predictions using just a few snapshots of the fluid flow over time, rather than needing many detailed simulations.

Technical Explanation

The Fourier-MIONet model builds on the recently developed multiple-input deep neural operators (MIONet) and incorporates the Fourier neural operator (FNO) into the network architecture. This allows the model to efficiently learn the solution operator for the partial differential equations governing multiphase flow in porous media.

Compared to the enhanced FNO (U-FNO) model, Fourier-MIONet has 90% fewer parameters, can be trained about 3.5 times faster, and requires much lower CPU and GPU memory (less than 15% and 35% respectively). Despite this significant reduction in computational cost, Fourier-MIONet can achieve similar prediction accuracy for the evolution of fluid saturation and pressure over time under various reservoir conditions.

The authors demonstrate that Fourier-MIONet can be trained with just 6 snapshots of the flow over time to accurately predict the solutions for 30 years. This excellent generalizability is enabled by the model's adherence to the physical principle that PDE solutions are continuous in time.

Critical Analysis

The paper provides a thorough evaluation of the Fourier-MIONet model's performance and computational efficiency compared to existing approaches. However, the authors do not discuss any potential limitations or caveats of the model.

It would be helpful to understand how Fourier-MIONet might perform on more complex or realistic multiphase flow scenarios, such as those involving heterogeneous rock properties, complex well configurations, or coupled processes like geochemistry. The authors also do not address potential issues with the model's ability to generalize to new types of reservoir conditions beyond those seen during training.

Additionally, while the computational efficiency of Fourier-MIONet is a significant advantage, the paper does not explore how this might impact the model's ability to capture fine-scale details or handle rapid changes in the flow field. Further investigation into the trade-offs between speed, accuracy, and resolution would strengthen the critical analysis.

Conclusion

The Fourier-MIONet model represents an important advancement in the field of neural operators for multiphase flow in porous media. By combining the strengths of MIONet and FNO, the authors have developed a computationally efficient model that can accurately predict the evolution of fluid flow and pressure in the subsurface over long timescales.

This work has significant implications for improving the feasibility of geologic carbon sequestration by enabling researchers to explore a wider range of reservoir scenarios and optimize the process in real-time. The model's ability to generalize from limited training data also suggests it may have broader applicability to other types of partial differential equation problems in science and engineering.

Overall, the Fourier-MIONet represents an important step forward in the development of computationally efficient and physically-consistent neural network models for complex multiphysics problems.