Rapid Prediction of a Liquid Structure from a Single Molecular Configuration Using Deep Learning

CH Li and B Gilbert and S Farrell and P Zarzycki, JOURNAL OF CHEMICAL INFORMATION AND MODELING, 63, 3742-3750 (2023).

DOI: 10.1021/acs.jcim.3c00472

Molecular dynamics simulation is an indispensable toolfor understandingthe collective behavior of atoms and molecules and the phases theyform. Statistical mechanics provides accurate routes for predictingmacroscopic properties as time-averages over visited molecular configurations- microstates. However, to obtain convergence, we need a sufficientlylong record of visited microstates, which translates to the high-computationalcost of the molecular simulations. In this work, we show how to usea point cloud-based deep learning strategy to rapidly predict thestructural properties of liquids from a single molecular configuration.We tested our approach using three homogeneous liquids with progressivelymore complex entities and interactions: Ar, NO, and H(2)Ounder varying pressure and temperature conditions within the liquidstate domain. Our deep neural network architecture allows rapid insightinto the liquid structure, here probed by the radial distributionfunction, and can be used with molecular/atomistic configurationsgenerated by either simulation, first-principle, or experimental methods.

Return to Publications page