GNN surrogate forecasts CO2 plume migration for carbon storage

A team led by Rodrigo S. Luna and colleagues at multiple institutions introduced a novel Graph Neural Network (GNN) surrogate designed to accelerate CO2 migration forecasting during geological carbon storage. The model reframes the SPE11A benchmark—which simulates sharp gas-water interfaces, advection, and convective fingering—as a graph where each computational cell becomes a node and edges encode transmissibility-based interactions enriched with geometric attributes. To handle directional transport caused by grid geometry, permeability contrasts, and heterogeneity, the network uses an anisotropic message-passing mechanism: edge embeddings conditioned on local geometry bias message aggregation toward physically relevant transport directions. Temporal evolution is modeled in latent space via an autoregressive residual formulation trained with multi-step supervision, enabling stable rollouts over extended time horizons.

On the SPE11A benchmark, the proposed GNN surrogate produces competitive forecasts for two key indicators—gas saturation and liquid-phase density—with cumulative errors that remain moderate even over long forecasting horizons. This approach offers a significantly faster alternative to full-physics simulations, potentially enabling real-time monitoring and scenario screening for carbon storage projects. By replacing costly numerical solvers with a learned emulator that respects key physical constraints, the method could help operators optimize injection strategies, detect leaks early, and reduce computational barriers to large-scale CCS deployment. The paper appears as a preprint on arXiv (2606.17180) and is currently under journal review.