Development of a Rapid Deep-Learning-Assisted Multiphase Multicomponent Numerical Simulation Protocol

Compositional reservoir simulations can represent the phase behavior of multicomponent fluids with high accuracy, but their practical use is often limited by the heavy computational cost of iterative phase-stability and flash calculations. In contrast, black-oil models are much faster, yet their simplified treatment of phase behavior restricts their reliability in problems where phase equilibrium strongly affects the flow performance, such as CO₂ injection. This study develops a deep-learning-based workflow that replaces conventional iterative flash calculations. The proposed model performs both classification and regression tasks, identifying phase stability and predicting component equilibrium ratios. After training, the expert system is fully coupled to a compositional simulator and evaluated using three benchmark cases. It is first tested for phase-envelope reconstruction, where blind tests show mean relative errors below 1% for the liquid fraction and phase compositions, while accurate behavior near the critical region confirms thermodynamic consistency. It is then integrated into a compositional solver for a five-spot gas-injection problem, where predicted oil and gas production rates show mean relative errors of below 0.2%. Tests on the Norne field model and a million-cell multiwell case demonstrate major efficiency gains, including a 517% faster flash calculation, a 2464% faster phase-stability testing, and over 2000% flash calculation improvement for grids exceeding a million cells.