Multisource Heterogeneous Monitoring Data Fusion and Dynamic Migration Characterization for CO 2 Geological Sequestration

Carbon capture and geological sequestration have emerged as important strategies for reducing atmospheric carbon dioxide (CO 2 ) emissions and mitigating global climate change. Recent advancements in subsurface monitoring, reservoir simulation, and deep learning techniques have significantly improved the capability to analyze CO 2 plume migration and long‐term storage behavior in geological formations. However, existing monitoring frameworks still face major challenges in integrating heterogeneous monitoring data, accurately characterizing dynamic plume migration, and handling uncertainty under varying geological conditions. To address these limitations, this study proposes a multisource heterogeneous monitoring data fusion framework for dynamic CO 2 migration characterization using the Sleipner CO 2 plume dataset and National Risk Assessment Partnership (NRAP) simulation dataset. The framework integrates plume boundary maps, pressure field maps, and CO 2 saturation maps through input‐level (feature‐level) fusion to jointly learn spatial representations of CO 2 migration. Geological reservoir properties, including permeability, porosity, formation depth, and oil saturation, are utilized for reservoir characterization and interpretation of migration behavior rather than as direct convolutional neural network (CNN) input features. A hybrid CNN–long short‐term memory (LSTM) architecture is employed in which the CNN extracts spatial reservoir features while the LSTM models temporal migration dynamics and sequential plume evolution behavior. Experimental evaluation demonstrated that the proposed framework achieved superior prediction performance with a mean squared error (MSE) of 0.0020, root mean square error (RMSE) of 0.0456, mean absolute error (MAE) of 0.0274, coefficient of determination ( R 2 ) score of 0.9738, mean absolute percentage error (MAPE) of 0.0841, and Intersection over Union (IoU) value of 0.9448. The obtained results confirmed improved spatial overlap accuracy, reduced prediction error, and enhanced uncertainty‐aware migration.