An Advanced Graph Neural Pipeline for CO2 Plume Migration in Complex Geological Formations
Please login to view abstract download link
Deployment of carbon capture and storage (CCS) technologies is expensive and risky, accurate and effective prediction of CO2 plume migration, especially when facing intricate subsurface formations is critical to reduce uncertainties and ensure viability of projects. Traditional numerical simulators, while reliable, incur prohibitively high computational costs when simulating long-term storage scenarios or performing uncertainty quantification studies. Recent surrogate modeling approaches, such as Fourier Neural Operators (FNOs), DeepOnets, and Graph Neural Networks (GNNs) [1], have demonstrated good accuracy and substantial speedups, particularly in many-query computations, such as parameter exploration, optimization, and uncertainty quantification, and are increasingly used in digital twins and digital shadows. GNNs naturally accommodate unstructured meshes, typical of geological reservoirs. In this work, we deploy an advanced end-to-end GNN pipeline, based on MeshGraphNet [2], to generate surrogates for CO2 plume migration. The model has three modules: an encoder that maps initial node and edge feature matrices into latent embeddings; a processor composed of sequential message-passing (M-P) steps that iteratively update the embeddings and capture spatial dependencies; and a temporal module via a Chebyshev graph-convolutional LSTM that integrates spectral graph convolutions, enabling the model to capture spatial and temporal correlations in a unified latent space. Finally, a decoder maps the latent node embeddings back to the physical space. An autoregressive training strategy over rolling windows is adopted to extend the prediction horizon while mitigating error accumulation. We apply the GNN model to build a surrogate for the well-known 11th SPE CSP Version 11A test case, which is a 2D case that has been calibrated against laboratory results and conditions. The governing equations form a fully coupled system that combines the Darcy momentum balance with N component-wise mass-conservation equations and an EOS flash constraint for local phase equilibrium. The benchmark presents several challenges: a sharp gas-water interface and a rapid onset of convective mixing with massive finger development make training a surrogate for gas saturation and phase composition very difficult.
