The full text of this item is not available at this time because the student has placed this item under an embargo for a period of time. The Libraries are not authorized to provide a copy of this work during the embargo period, even for Texas A&M users with NetID.
Application of Fourier Neural Operator and Transfer Learning for Rapid Forecasting of Geological Carbon Storage
Abstract
Geological carbon storage (GCS) is a crucial technique for the mitigation of anthropogenic carbon dioxide (CO2) emissions into the atmosphere. Despite its potential, GCS presents numerous risks, including leakage risks and potential adverse impacts on surrounding environments. Numerical simulators are critical in addressing these challenges by accurately predicting pressure and saturation distribution within the reservoir to ensure long-term storage integrity. However, numerical simulators often require long computational times because of the complexity of the multi-physics nature of these problems. There is a sense of urgency associated with the deployment of GCS projects, and this requires faster approaches to model these processes in large areas of review.
In recent years, machine learning has been recognized as a potentially powerful tool in the evaluation and analysis of numerical simulation results because of its ability to extract patterns from complex data. It can enable reliable predictions of reservoir behavior, improved production forecasting, optimal well placement, and real-time monitoring. The Fourier Neural Operator (FNO) method is a promising machine learning technique for the prediction of the distributions of pressure and phase saturations in CO2 sequestration simulations. Unlike commercial reservoir simulators, FNO-based models can provide accurate predictions with much lower computational costs, enabling more simulations in less time. However, most of these models are limited to specific scenarios and conditions.
In this proposed study, I implemented an updated architecture, padding, feature engineering, and data sparsity management to improve a FNO workflow used to train a model on a set of scenarios with fixed injection locations and constant injection rates for GCS. I then use transfer learning techniques to leverage the knowledge gained from this initial training in order to enhance the performance of the model when applied to a wider array of scenarios. The main goal of this study is to demonstrate how Transfer Learning (TL) can accurately predict CO2 the distributions of pressure and phase saturations under varying geological and operational conditions, such as different injection rates and well locations.
The results of this study strongly suggest that the improved FNO workflow surpasses the outcomes obtained by means of traditional numerical reservoir simulations, reducing the computational time by approximately 97% and the relative mean error to less than 1% for the predictions of the distributions of both the CO2 pressure and saturation. Moreover, this work offers insights into the application of TL to reservoir simulation, demonstrating that today's computational resources can effectively transfer knowledge from a pre-existing model to other related tasks, while reducing the need for samples by 78% and maintaining a relative mean error below 5%. This is achieved while effectively delineating the areas of review for such projects, despite occasional over-predictions. This work provides well-documented evidence of the potential of TL to reduce the computational time and the data requirements for reservoir simulation in GCS projects, offering a more efficient and faster means of conducting these simulations. Thus, this study provides compelling evidence supporting the continued investigation into future applications of this promising approach.
Subject
Geological Carbon SequestrationReservoir Simulation
Machine Learning
Transfer Learning
Fourier Neural Operator
Citation
Calvo Nunez, Andres Felipe (2023). Application of Fourier Neural Operator and Transfer Learning for Rapid Forecasting of Geological Carbon Storage. Master's thesis, Texas A&M University. Available electronically from https : / /hdl .handle .net /1969 .1 /200129.