Molecular dynamics simulations for the molecular polarization of salt-free and salt-containing liquids with Stockmayer fluids and ensemble neural networks

We develop our simulation method for the molecular dielectric response via a Stockmayer fluid (dipolar molecule), combined with neural networks using machine-learning techniques. We first show that despite the drastic simplification of polar molecules with a minimal set of molecular parameters, our coarse-grained molecular dynamics simulations are consistent with the experimental data of the dielectric constants of various salt-free and salt-containing organic solvents. To substantially reduce the computational cost due to the statistical analysis of the simulation data and the parametrization of the model parameters, we also construct surrogate models for the dielectric constant of the solvents using ensemble neural networks. We show that our models can predict the qualitative trends of the simulation results by taking up a small number of samples from a population of statistically-noisy simulation data.