(502d) Construction of Coarse-Grained Neural Network Potentials for Predicting Water Properties

Accurate yet computationally trackable interatomic potential will be invaluable in many fields of science and engineering. However, many existing interatomic potentials are frequently limited by their fixed functional form, high computational cost, transferability for a wide range of environmental conditions. In particular, the transferability of force field is very poor for polar and hydrogen bonded liquids such as water. Recent advances in machine learning methods have shown promising pathways to improve the accuracy and transferability of inter-atomic potentials. In this direction, here, we attempt to develop a coarse-grained neural network (CG-NN) potential that predict the potential energy surface (PES) of water very accurately. Although the atomic forces are not trained to the network, the analytical derived forces from the predicted energies shows appreciable agreement with the actual atomic forces on the coarse-grained atoms. Further Monte Carlo (MC) and molecular dynamics (MD) simulations based on CG-NN potential predict reasonably accurate thermophysical properties of water. We also analyze the correlations between the predictability of the CG-NN model with its topology and the choice of symmetry functions that represent a PES.