Physical-informed deep learning framework for CO2-injected EOR compositional simulation

CO2 injection for enhanced oil recovery (CO2-EOR) is one of the industry's most widely applied techniques for CO2 utilization. People use the compositional simulation technique to clearly describe CO2 injection processes for EOR, which is always computational-expensive. The main objective of this study is to utilize deep learning (DL) techniques to describe phase behaviors and accelerate compositional simulations. First, we constructed a new DL-based compositional simulation framework. Instead of solving equations of state for every grid cell, this framework can aggregate millions of phase behavior solvers and determine all mixtures' phase status in one single round of calculation. This trait can significantly improve the efficiency of solving numerical equations. Next, we illustrated how to develop DL-based phase behavior models to determine phase status accurately. Our newly proposed single-phase identification model can resolve the difficulty in inaccurate singlephase labeling within compositional simulations, greatly benefiting the numerical convergence of simulation. Finally, we tested abundant complicated cases covering near-miscible and miscible gas injection processes and demonstrated the new DL-based method's strength in accelerating the compositional simulation. Besides, we also presented the DL-based method's more substantial power in numerical convergences that it could succeed in cases where the standard simulation method fails. This work can contribute to more efficiently and effectively describing complex fluid displacement processes, benefiting CO2 injection EOR and other utilizations like CO2 sequestrations.