GNN Surrogates Forecast CO2 Migration in Complex Geological Formations

Researchers have developed a novel graph neural network (GNN) surrogate model designed to predict the movement of CO2 plumes within intricate geological formations. This data-driven machine learning approach aims to accurately reproduce the physical behavior of multiphase flows, specifically focusing on CO2 storage scenarios. The GNN model reformulates geological benchmarks as graphs, where computational cells are nodes and interactions are edges. It incorporates an anisotropic message-passing mechanism, which accounts for directional transport influenced by grid geometry, permeability, and geological heterogeneity. The model also uses an autoregressive residual formulation for temporal evolution, trained with multi-step supervision. Evaluated on the SPE11A benchmark, an industry standard for CO2 storage assessment, the proposed GNN surrogate demonstrates competitive forecasting capabilities for key indicators like gas saturation and liquid-phase density. It maintains moderate cumulative errors over extended prediction horizons, offering a promising tool for monitoring CO2 storage.