A hybrid molecular dynamics/machine learning framework to calculate the viscosity and thermal conductivity of Ar, Kr, Xe, O, and Ν

Abstract

In this paper, molecular dynamics (MD) simulations and machine learning (ML) methods are combined to obtain the transport properties, such as viscosity and thermal conductivity, of five basic elements, which are computationally hard to obtain at the nanoscale and extremely demanding to estimate accurately through an experimental procedure. Starting from an experimental database from literature sources, we extend the (P-T) space on which the transport properties are calculated by employing MD simulations and ML predictions, in a synergistic mode. Results refer to all fluid states (gas, liquid, supercritical), under ambient and supercritical conditions, suggesting an alternative path that can be accurately followed to bypass expensive experiments and costly numerical simulations. Nine different ML algorithms are exploited and assessed on their prediction ability, with tree-based architectures achieving increased accuracy on the implied data set. The proposed computational platform runs fast in a common python Jupyter environment, both for MD and ML, and can be adjusted and extended for the calculation of material properties both in interpolation and extrapolation applications.