A Novel Machine-Learning Assisted Phase-Equilibrium Calculation Model for Liquid-Rich Shale Reservoirs

Abstract

In composition reservoir simulation, fluid phase behavior is determined by vapor-liquid equilibrium (VLE) calculations. VLE calculations can consume more than half of the CPU time of compositional reservoir simulations. To accelerate the VLE calculations, machine learning (ML) technique is introduced. In this work, we developed a novel ML-assisted VLE calculation model for shale reservoirs. Our model has two main innovations compared with previous ML-assisted VLE calculation models. Firstly, the extended Peng-Robinson equation of states (PR-C EOS) is incorporated for VLE calculation. Previous models used the conventional Peng-Robinson equation of states (PR EOS), which becomes inaccurate when the pore diameter reduces to the scale of nanometers. With PR-C EOS, fluid characteristics can be accurately modeled under nano-scale conditions, making our model applicable to shale reservoirs. Secondly, in our model, a general set of pseudo components is selected to cover different fluid types. Previous models are designed for a specific type of hydrocarbon mixture. There are two parts to our model: stability analysis and flash calculation. In the stability analysis, the multi-layer perceptron (MLP) is trained to predict whether the fluid is in single-phase or two-phase condition. The equilibrium ratios are estimated using a physics-informed neural network (PINN) in the flash calculation. The application of ML techniques accelerates the CPU time by two orders of magnitude without losing too much accuracy. This work provides the framework of incorporating ML into VLE calculation and develops a ML-assisted VLE calculation model that is suitable for various hydrocarbon mixtures in shale reservoirs.