The Prediction of CO2 Plume Using Neural Network Based On the Swin Transformer

Abstract

Investigating the migration of carbon dioxide (CO₂) fluids is essential for the effective monitoring in the geological sequestration of CO₂. Traditional numerical simulation methods are often time-consuming and computationally expensive. Recently, deep learning methods, particularly convolutional neural networks (CNNs), have gained traction for predicting CO₂ plume migration. However, these approaches typically require extensive training datasets and tend to emphasize local information. To overcome these limitations, we introduce a visual attention model along with a novel neural network based on the Swin Transformer architecture to forecast CO₂ plume migration in heterogeneous geological formations. A significant challenge in conventional machine vision is the translational invariance of input images, which can hinder performance. To address this issue, we integrate relevant physical prior knowledge into our model. Compared with U-net and Transformer, the model exhibits highest predictive performance, with an R² score of 0.9741 and the test set root mean squared error (RMSE) reaching 0.0245. These results indicate that this approach enables the network to effectively extract both local and global features, maximizing the use of limited datasets and enhancing the understanding of CO₂ migration patterns. Additionally, the model demonstrates strong capabilities for global information learning and generalization. These advantages, therefore, facilitate the extensive application of the visual attention model in predicting CO₂ migration.