Efficient Phase Equilibrium Calculations of Shale Reservoir Fluids—Part III: Deep Learning–Based Acceleration and Integration into Compositional Simulation

Summary Compositional simulation is pivotal for optimizing carbon dioxide (CO2)-injection strategies in shale reservoirs, where nanopore confinement effect and capillary pressure significantly alter fluid phase behavior. However, computational bottlenecks in phase equilibrium calculations (PECs) hinder efficient simulation. In this study, we develop an artificial neural network (ANN)–based framework integrating two independent models: one for phase stability analysis (PSA) with a probability threshold and another for flash calculations providing initial values to a standard algorithm. A tailored training data generation scheme addresses compositional variations during CO2 injection. By incorporating confinement effect and capillarity—critical for shale reservoirs—our framework achieves higher accuracy in phase behavior prediction. We have completed seamless integration of this novel PEC framework into Stanford University’s next-generation compositional simulator Automatic Differentiation-Based General Purpose Research Simulator (ADGPRS), achieving high-performance simulation. Field-scale validation on China’s Gulong shale reservoir (137,580 active grids) demonstrates the framework’s industrial applicability: Confinement effect and capillary pressure increase cumulative oil/gas production by 6–21% and 14–47%, respectively, depending on pore size distribution. Standalone PEC tests demonstrate more than 80% computation time reduction, while coupled with ADGPRS, the approach accelerates compositional simulation of Gulong shale reservoir by more than 40%. This work provides a computationally efficient solution for CO2-enhanced oil recovery (EOR) and carbon sequestration in shale reservoirs, bridging the gap between theoretical algorithms and industrial-scale implementation.