Accurate forecasting of CO₂ plume migration, geochemical alteration, and geomechanical deformation is essential for secure geological carbon storage, yet high-fidelity thermo-hydro-mechanical-chemical (THMC) simulators remain too expensive for rapid uncertainty screening and digital-twin-style updates. This study develops a graph-based multi-output surrogate trained on THMC simulation outputs to predict three engineering observables relevant to storage performance and risk assessment: gas saturation (Sg), formation-water pH, and vertical surface displacement. The framework is evaluated over a century-long period from 2031 to 2130, spanning both injection and post-injection phases. Three baseline architectures are compared under a strict realization-level split, namely a multilayer perceptron (MLP), a convolutional neural network (CNN), and a graph neural network (GNN). Among the tested baselines, the GNN delivers the strongest overall accuracy and the most faithful spatial reconstruction, particularly for plume front geometry and deformation localization under unseen geological realizations. At the same time, the proposed model should be interpreted as an emulator of simulator-derived observables rather than as a conservation-enforcing THMC solver. We therefore position the framework as a fast forecasting layer complemented by uncertainty quantification and physics-aware plausibility checks, not as a replacement for full-physics simulation in regulatory or site-specific design studies. Millisecond-scale inference and calibrated ensemble prediction intervals make the approach useful for rapid scenario screening, monitoring support, and digital-twin workflows, while future work should incorporate explicit conservation constraints, richer nonlinear geomechanics, and larger training ensembles.
Thanh et al. (Wed,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: