Predicting self-diffusion coefficients of small molecular fluids using machine learning and the statistical associating fluid theory for Mie segments

In our previous work [Šlepavičius et al.,“Application of machine-learning algorithms to predict the transport properties of Mie fluids,” J. Chem. Phys. 159, 024127 (2023)], we applied three machine learning (ML) models to predict the self-diffusion coefficient of spherical particles interacting via the Mie potential. Here, we introduce an optimization approach using the so-called statistical associating fluid theory for Mie segments and available vapor–liquid equilibria data to obtain molecular parameters for both Mie and Lennard-Jones potentials to describe the diffusion coefficient of 16 molecules described as a single sphere. Our ML models utilize these molecular parameters to predict the self-diffusion of these molecules. We conduct a comparative analysis between the molecular parameters derived from our thermodynamic approach and those obtained through direct fitting of the experimental self-diffusion coefficients. Our findings indicate that the predictive accuracy remains largely unaffected by the specific repulsive and attractive exponents of the Mie potential employed, provided that the fitting of the molecular parameters is precise. The Mie parameters obtained within a thermodynamic framework exhibit a higher coefficient of determination (R2) and absolute average relative deviation values compared to those derived from molecular parameters derived from fitting the self-diffusion coefficient, indicating their superior precision at higher values of the self-diffusion coefficient. Despite this discrepancy, the overall precision of both methodologies remains comparable. Given the abundance of precise thermodynamic data in contrast to self-diffusion data, we advocate the thermodynamic fitting approach as the preferred method for acquiring accurate Mie coefficients, essential to predict self-diffusion coefficients with ML and semi-empirical models.