Machine learning for predicting thermodynamic properties of pure fluids and their mixtures

Abstract

Establishing a reliable equation of state for largely non-ideal or multi-component liquid systems is challenging because the complex effects of molecular configurations and/or interactions on the thermodynamic properties must generally be taken into account. In this regard, machine learning holds great potential for directly learning the thermodynamic mappings from existing data, thereby bypassing the use of equations of state. The present study outlines a general machine learning framework based on high-efficiency support vector regression for predicting the thermodynamic properties of pure fluids and their mixtures. The proposed framework is adopted in conjunction with training data obtained from a high-fidelity database to successfully predict the thermodynamic properties of three common pure fluids. The predictions demonstrate extremely low mean square errors. Moreover, little loss in the prediction accuracy is obtained for ternary mixtures of the pure fluids at the cost of a modest increase in the volume of training data provided by state-of-the-art molecular dynamics simulations. Our results demonstrate the promising potential of machine learning for building accurate thermodynamic mappings of pure fluids and their mixtures. The proposed methodology may pave the way in the future for the rapid exploration of novel or complex systems with potentially exceptional thermodynamic properties.