Representing liquid-vapor equilibria of Ternary systems using neural networks

Abstract

We develop a method based on neural networks for efficiently interpolating equations of state (EOS) for liquid-vapor equilibria of ternary mixtures. We investigate the performance of neural networks both when experimental data are available and when only simulation data are available. Simulation data are obtained from Gibbs Ensemble Monte Carlo simulations, using the TraPPE-EH molecular model. Our investigation uses the mixture of carbon dioxide, methane, and ethane as a validation example, for which experimental data exist. Analysis of the error in a neural-network-generated liquid-vapor coexistence curve shows that the resulting interpolation is robust and accurate, even in the case where the network is trained on a few data points. We use this observation to construct a methodology for accurately locating liquid-vapor equilibria of ternary mixures wihout using any experimetnal data.