(258c) Machine Learning Potentials for Thermodynamic and Transport Properties of Bulk Liquids

Classical molecular dynamics (MD) and Monte Carlo (MC) methods rely on force fields to describe the potential energy landscape that governs molecular interactions. A recurrent strategy to parametrize force fields for molecular systems relies on the availability of thermodynamic and transport experimental measurements, such as density or viscosity, which might not be available for the system under study at the conditions of interest. An alternative route to obtain such properties is to conduct ab-initio molecular dynamics (AIMD) simulations, which do not require pre-existing experimental measurements. Despite this clear advantage, AIMD is prohibitively expensive, severely limiting the accessible length and time scales required to compute the physical properties of interest.

Recently, neural networks have raised a significant amount of interest due to their ability to reproduce multi-dimensional potential energy surfaces with high accuracy. They have been proposed as surrogates to costly AIMD simulations, as they faithfully reproduce the underlying quantum mechanical energies and forces with less computational demands. We utilize neural network force fields (NNFFs) that use local descriptors to compute transport and thermodynamic properties of several organic liquids. The NNFFs were trained from energies and forces obtained from AIMD simulations. Taking advantage of the computational efficiency of NNFFs, we are able to reach the nanosecond and nanometer time and length scales required to compute transport and thermodynamic properties such as density, viscosity or diffusion coefficients consistent with the underlying quantum mechanical potentials.