Predictive performance of thermophysical properties of supercritical carbon dioxide using Gaussian process regression and deep learning

Abstract

Carbon dioxide has important usage in many engineering applications, such as regenerative cooling in scramjets and supercritical power cycles, and thermophysical properties are essential in numerical simulation. For the typical cubic real-fluid equation of states (EOSs), such as Soave-Redlich-Kwong (SRK) and Peng-Robinson (PR) models, their reliability and accuracy are relatively poor in the high-pressure compressed-liquid, pseudo-boiling, and near-critical regions. In the region around the critical point, the renormalization group theory was developed to improve the predictive capability, but the associated complexity increases the computational cost substantially. To alleviate the situation, the present study leverages the recent advance in machine learning and proposes different data-driven models using Gaussian process regression (GPR) and deep learning for property evaluations of carbon dioxide. Gaussian process regression with various kernel functions and deep feedforward neural network (DFNN) model are explored to estimate the properties of carbon dioxide in a wide range of thermodynamic states. Both GPR and DFNN models show excellent agreement with the standard database, with the corresponding absolute average relative deviation (AARD) below 0.5% and 2.5%, respectively. The developed data-driven models can be potentially incorporated into large-scale computation of propulsion systems in a more accurate and efficient manner than the cubic EOSs.