An Artificial Intelligence–Assisted Framework for Accelerating Multiphase Equilibrium Calculations in Carbon Dioxide-Enhanced Oil Recovery and Storage Simulations

Summary

Carbon dioxide (CO2)-enhanced oil recovery (EOR) (CO2-EOR) coupled with geological storage supports improved hydrocarbon recovery while providing a pathway for long-term subsurface CO2 management. Reliable assessment of these processes requires compositional simulation, but repeated oil/gas/water phase-equilibrium calculations can dominate runtime. To address this bottleneck, this study develops a simulator-coupled artificial intelligence (AI)–assisted framework that accelerates these thermodynamic calculations. It combines a phase classifier and three phase-specific flash initializers with conservative confidence thresholds and fallback checks. A physics-based flash solver determines the final equilibrium state. A simulation-oriented data-generation strategy samples thermodynamic states along CO2 injection trajectories and near phase boundaries. Performance was assessed through static pressure-composition (P-x) benchmarks and coupled 2D water-alternating-gas (2D-WAG) and 3D tests. Threshold-accepted phase predictions achieved an accuracy greater than 99%, while standalone P-x calculation time was reduced by 70.59–85.61%. In the heterogeneous 2D-WAG and 3D tests, the AI-assisted results closely matched the reference saturation, production, and spatial responses. The thermodynamic-module runtime was reduced by 88.66% in the 2D-WAG simulation and 82.66% in the 3D simulation, lowering overall runtime by 59.86% and 44.21%, respectively. Physics-based flash correction preserved agreement with the reference model while reducing computational expense. To the best of our knowledge, this is the first AI-assisted oil/gas/water equilibrium workflow coupled with a compositional simulator, establishing a basis for broader thermodynamic acceleration in reservoir modeling.