Modeling of thermophysical properties and vapor-liquid equilibrium using Gaussian process regression

Thermophysical properties are essential in the numerical simulation of fluid flows and combustion in propulsion and power-generation systems. Existing property evaluation tools, such as NIST REFPROP, analytical methods based on cubic equations of state (EOSs), and deep neural network (DNN) models, tend to be inaccurate against experimental data in certain thermodynamic regions. The objective of this paper is to develop data-driven models based on Gaussian process regression (GPR), which correlates directly with experimental data in a statistical manner and can accurately predict thermodynamic and transport properties and vapor-liquid equilibrium (VLE) behavior across the entire range of thermodynamic regions. Different kernel functions in GPR training are examined for various data types. As specific examples, experimental property data of carbon dioxide and VLE behavior of four binary mixtures are collected extensively from literature for training and validation of GPR models. Besides, physical constraint is incorporated during the training process of VLE behavior via logit transformation. The results show that GPR with exponential kernel presents excellent performance for evaluation of density, constant-pressure specific heat, speed of sound, and viscosity, with prediction errors substantially lower than those of Peng-Robinson (PR) EOS, NIST REFPROP, and DNN models. GPR with Gaussian kernel shows minimal errors (less than 0.6%) for determination of saturation pressure and gas-phase composition against experimental counterparts for VLE of hydrocarbon/nitrogen binary mixtures. The developed GPR models can be programmed and incorporated into large-scale simulations to achieve more reliable calculations of practical fluid flows where real-gas effect is important.