Rapid CO2-water multi-phase flow prediction with machine learning
Abstract/Contents
- Abstract
- To reach net-zero emissions globally, an estimated five to ten gigatons of CO2 must be captured and permanently stored annually. Scaling up the carbon capture and storage industry requires extensive computational support that will be very challenging without accurate and fast algorithms. However, current numerical simulation approaches are very expensive due to the multi-phase, multi-physics, and multi-scale nature of CO2 storage. This thesis demonstrated that machine learning provides an unprecedented tool for tackling this complex problem. We present a series of works for modeling CO2 storage with progressively more complex problem settings, from a demo scale to a 3D basin scale. To demonstrate the viability of using machine learning approaches for modeling CO2 storage, we first proposed a convolutional neural network (CNN) model to predict CO2 plumes in a 2D demo-scale heterogeneous formation. The CO2 plume is subject to the complex interplay of gravity, viscous, and capillary forces. Also, we applied engineering controlling parameters of injection time and injection perforation location to control the plume. The proposed CNN model was able to predict the CO2 plume with very high accuracy and computational efficiency, which are the two important foundations of using machine learning approaches to support the computational needs required by CO2 storage project development. Building on these foundations, we introduced CCSNet, a machine-learning modeling suite for CO2 storage in flat saline formations. CCSNet fully explores the potential of a CNN-based algorithm to provide a practical alternative to numerical simulation. To train the modeling suite, we generated a dataset with 20,000 simulations to cover most of the realistic reservoir conditions, rock properties, geological models, and injection designs for CO2 storage. CCSNet consists of six CNN models to predict saturation distributions, pressure buildup, fluid densities, and molar fractions of each fluid. CCSNet demonstrates that all the necessary outputs to CO2 storage, including mass balance, can be accurately and efficiently predicted by machine learning. As we increase the problem-setting complexity, we also investigate machine learning architectures that are more effective for the CO2 storage problem. Although CNN models are fast and accurate, they often require training over huge numerical simulation data sets to achieve desirable performance. As a result, the computational costs of the data collection process can lead to significant overburdens when using machine learning-based approaches. To overcome this challenge, we proposed U-FNO, an enhanced Fourier neural operator (FNO) architecture for multiphase flow. FNO is a novel architecture that learns the solution operator efficiently in the Fourier domain. U-FNO outperforms both the original FNO and a state-of-the-art CNN benchmark in gas saturation and pressure buildup predictions. It unlocks the potential for tackling 3D problems and makes the vision for a general-purpose machine learning tool significantly more practical. Taking FNO's advantage of data efficiency, my latest work, Nested FNO, tackles one of the most challenging CO2 storage settings: basin-scale 3D simulation in dipping reservoirs with multiple injection wells. Using a semi-adaptive local grid refinement approach and a hierarchy of FNO models, this work enables high-fidelity full-physics to flow predictions for real-world CO2 storage projects with diverse reservoir conditions, geology, and injection schemes. It provides a fully functional simulator alternative with five orders of magnitude speed-up at unprecedented high resolutions. Once trained, machine learning models provide significant computational benefits because the inferences of machine learning models are often very cheap. The four pieces of work discussed above generally speeds up prediction by 4 to 6 orders of magnitudes. We demonstrate the speed-up with a web application, https://ccsnet.ai, that provides instant prediction to academics and the general public. This also promotes equity in CO2 storage project development and knowledge adoption.
Description
| Type of resource | text |
|---|---|
| Form | electronic resource; remote; computer; online resource |
| Extent | 1 online resource. |
| Place | California |
| Place | [Stanford, California] |
| Publisher | [Stanford University] |
| Copyright date | 2023; ©2023 |
| Publication date | 2023; 2023 |
| Issuance | monographic |
| Language | English |
Creators/Contributors
| Author | Wen, Gege |
|---|---|
| Degree supervisor | Benson, Sally |
| Thesis advisor | Benson, Sally |
| Thesis advisor | Durlofsky, Louis |
| Thesis advisor | Tchelepi, Hamdi |
| Degree committee member | Durlofsky, Louis |
| Degree committee member | Tchelepi, Hamdi |
| Associated with | Stanford Doerr School of Sustainability |
| Associated with | Stanford University, Department of Energy Resources Engineering |
Subjects
| Genre | Theses |
|---|---|
| Genre | Text |
Bibliographic information
| Statement of responsibility | Gege Wen. |
|---|---|
| Note | Submitted to the Department of Energy Resources Engineering. |
| Thesis | Thesis Ph.D. Stanford University 2023. |
| Location | https://purl.stanford.edu/kw803yf5830 |
Access conditions
- Copyright
- © 2023 by Gege Wen
- License
- This work is licensed under a Creative Commons Attribution Non Commercial No Derivatives 3.0 Unported license (CC BY-NC-ND).
Also listed in
Loading usage metrics...