Generating a Machine-learned Equation of State for Fluid Properties

TL;DR

This paper demonstrates that machine learning models like neural networks and Gaussian processes can effectively replicate traditional equations of state for fluids, enabling improved prediction of thermodynamic properties.

Contribution

It introduces a machine-learned approach to model fluid properties, replacing traditional analytical equations of state with neural networks and Gaussian processes for better accuracy and flexibility.

Findings

Machine learning models accurately predict critical properties.

ML models outperform traditional EoS in extrapolation.

Large dataset confirms viability of ML for thermophysical predictions.

Abstract

Equations of State (EoS) for fluids have been a staple of engineering design and practice for over a century. Available EoS are based on the fitting of a closed-form analytical expression to suitable experimental data. The underlying mathematical structure and the underlying physical model significantly restrain the applicability and accuracy of the resulting EoS. This contribution explores the issues surrounding the substitution of analytical EoS for machine-learned models, in particular, we describe, as a proof of concept, the effectiveness of a machine-learned model to replicate statistical associating fluid theory (SAFT-VR-Mie) EoS for pure fluids. By utilizing Artificial Neural Network and Gaussian Process Regression, predictions of thermodynamic properties such as critical pressure and temperature, vapor pressures and densities of pure model fluids are performed based on molecular descriptors. To quantify the effectiveness of the Machine Learning techniques, a large data set is constructed using the comparisons between the Machine-Learned EoS and the surrogate data set suggest that the proposed approach shows promise as a viable technique for the correlation, extrapolation and prediction of thermophysical properties of fluids.