Enhancing Molecular Dynamics Simulations of Aqueous Systems with Deep Neural Network Force Fields

Molecular dynamics simulations have been widely utilized in various scientific fields to study a variety of physical systems. However, the accuracy of these type of simulations strongly depends on the model used to describe the atomic interactions. Although ab initio molecular dynamics (AIMD) based on density functional theory (DFT) offers high accuracy, it is limited to small systems and relatively short simulation times. In this context, Neural Network Force Fields (NNFFs) play an important role by providing a way to overcome these limitations. In this study, we investigate NNFFs designed at the DFT level to describe liquid water, with a focus on the size and quality of the training data set. We show that training data set should be sampled from uncorrelated snapshots with various system conditions and configurations, which provides a good distribution over the phase space allowing one to significantly reduce the amount of data and the size of the NN required to have accurate NNFF. Furthermore, our results demonstrate that structural properties of water are less dependent on the size of the training data set compared to dynamical properties, such as the diffusion coefficient.