Accelerating High-Fidelity Three-Phase Equilibrium Calculation for CO ₂ -EOR and Storage: A Tie-Simplex Augmented Neural Network Approach

CO2 injection is widely recognized as an efficient, economically attractive, and environmentally relevant technique for enhanced oil recovery (EOR) and concurrent geological sequestration. However, the accurate modeling of complex multiphase hydrocarbon systems remains a significant computational bottleneck in compositional simulations. This study introduces, to our knowledge, the first AI-driven surrogate framework tailored for three-phase hydrocarbon equilibrium calculations, coupling artificial neural network (ANN) models with a rigorous thermodynamic solver to achieve rapid calculations while preserving thermodynamic consistency. To improve model generalization under dynamic compositional variations during CO2 injection, raw data sets are constructed by integrating Latin Hypercube Sampling (LHS) for base oils, P-x phase diagram generation, and adaptive sampling. Subsequently, tie-simplex theory enables physically informed data augmentation at a negligible computational cost. The integrated framework includes a classifier for phase stability analysis, along with dual initializers for two- and three-phase flash calculations. By embedding conservative probability thresholds and a thermodynamic fallback mechanism, the system reduces phase misclassification risk and routes uncertain cases to the rigorous solver, thereby maintaining numerical stability and solution accuracy. Extensive validation shows that the models achieve over 99% classification accuracy, restrict misidentification rates to 0.00%–0.0388%, and reduce flash iterations to 35.56%–43.99% of conventional baselines, thereby improving overall computational efficiency by 65.73% to 86.80%. The framework is evaluated across fluid mixtures of varying complexity, and computational acceleration becomes more pronounced as fluid complexity increases. Ultimately, this hybrid methodology establishes a thermodynamically consistent acceleration layer for efficient compositional simulation of complex CO2-hydrocarbon systems.