Thermodynamics-Informed Neural Network (TINN) for Phase Equilibrium Calculations Considering Capillary Pressure

Tao Zhang,Shuyu Sun

doi:10.3390/en14227724

Abstract

The thermodynamic properties of fluid mixtures play a crucial role in designing physically meaningful models and robust algorithms for simulating multi-component multi-phase flow in subsurface, which is needed for many subsurface applications. In this context, the equation-of-state-based flash calculation used to predict the equilibrium properties of each phase for a given fluid mixture going through phase splitting is a crucial component, and often a bottleneck, of multi-phase flow simulations. In this paper, a capillarity-wise Thermodynamics-Informed Neural Network is developed for the first time to propose a fast, accurate and robust approach calculating phase equilibrium properties for unconventional reservoirs. The trained model performs well in both phase stability tests and phase splitting calculations in a large range of reservoir conditions, which enables further multi-component multi-phase flow simulations with a strong thermodynamic basis.

Highlights

With the large amount of air pollution caused by the traditional energy consumption represented by coal, oil and firewood, human society has been paying more and more attention to natural gas as a relatively clean energy [1]
In order to overcome the problem that a limited number of expensive experimental data were used as the training ground truth, an NVT flash calculation was first carried out in [35], and the results provided enough data for training and testing
In order to accurately describe the complex thermodynamic laws in the phase equilibrium investigations, it is difficult for the model trained by the deep neural network to fully avoid the overfitting problem caused by the high complexity, which affects the final prediction effect

Summary

Introduction

With the large amount of air pollution caused by the traditional energy consumption represented by coal, oil and firewood, human society has been paying more and more attention to natural gas as a relatively clean energy [1]. It is of critical importance to design a thermodynamically consistent flash calculation scheme considering capillary pressure to accurately predict phase behaviors for the reliable simulation of fluid flow and transport in shale gas reservoirs. Both the conventional NPT and NVT flash calculation schemes are implemented by iterative calculations, which is too time-consuming to be adapted in large-scale reservoir simulators, especially when a large number of components are involved.

Thermodynamic Foundations

Deep Learning Technique

Results

Method IDteerIeatpetrilMavetaeiervNntehinNVogTVdaTfllgflaoasrhsithhm DeeItperlaetairvneinNgVaTlgfolarsithhm