A Comprehensive Review of Machine Learning Application to Flash Calculations in Compositional Reservoir Simulators

Abstract Reservoir engineering often involves dealing with formations that contain several chemical species and show complex phase behaviors. One of the most critical aspects of this field is calculating phase equilibrium, which is usually achieved through numerical simulations of multi-component, multi-phase flow in porous media. These simulations are known as flash calculations and describe the phase behavior of specific fluid mixtures. Flash calculations are typically performed using reservoir simulators that are based on equations of state (EOS), such as the Peng–Robinson (PR) and the Soave–Redlich–Kwong (SRK). While EOS-based flash calculations are known for their accuracy in describing phase behavior within reservoirs, they can be computationally intensive and time-consuming. Machine learning (ML), a branch of artificial intelligence, offers a promising alternative by predicting desired outputs through learning complex patterns among fluid properties of the reservoir. ML models have the potential to outperform traditional reservoir simulators in predicting phase equilibrium by significantly reducing the computational time required for flash calculations. This paper reviews various machine learning models developed over the years as alternatives to traditional flash calculations. It also explores the application of ML in both stability and phase split tests, discussing their limitations and providing recommendations for further improvements.