Acta Chimica Sinica ›› 2022, Vol. 80 ›› Issue (5): 598-606.DOI: 10.6023/A22010003 Previous Articles     Next Articles

Article

基于神经网络势与增强采样的气相水团簇成核过程研究

徐森a, 吴丽铃b, 李震宇a,b,*()   

  1. a 中国科学技术大学 合肥微尺度物质科学国家研究中心 合肥 230026
    b 中国科学技术大学 化学物理系 合肥 230026
  • 投稿日期:2022-01-03 发布日期:2022-05-31
  • 通讯作者: 李震宇
  • 作者简介:
    庆祝中国科学院青年创新促进会十年华诞.
  • 基金资助:
    国家自然科学基金(21825302); 国家自然科学基金(21573201)

Nucleation of Water Clusters in Gas Phase: A Computational Study Based on Neural Network Potential and Enhanced Sampling

Sen Xua, Liling Wub, Zhenyu Lia,b()   

  1. a Hefei National Laboratory for Physical Sciences at the Microscale, University of Science and Technology of China, Hefei 230026, China
    b Department of Chemical Physics, University of Science and Technology of China, Hefei 230026, China
  • Received:2022-01-03 Published:2022-05-31
  • Contact: Zhenyu Li
  • About author:
    Dedicated to the 10th anniversary of the Youth Innovation Promotion Association, CAS.
  • Supported by:
    National Natural Science Foundation of China(21825302); National Natural Science Foundation of China(21573201)

Due to their low density in atmosphere, theoretical simulations of the nucleation of gas-phase molecules are computationally very expensive. In this study, neural network potential (NNP) is combined with enhanced sampling techniques to effectively investigate the nucleation of water clusters in gas phase. The neural network potential is trained based on water-cluster energies and forces from density functional theory (DFT). The problem that the binding between water molecules is too weak in the previous empirical force field model has been solved in the NNP. This NNP potential is then applied to Monte Carlo simulations in grand canonical ensemble with enhanced sampling methods such as aggregation-volume-bias Monte Carlo (AVBMC) and transition-matrix Monte Carlo (TMMC) to realize a random walk among different cluster sizes. Probability distribution of water cluster sizes and the corresponding Gibbs free energies can then be obtained. Subsequently, the evaporation rates of water clusters can be calculated via umbrella sampling Monte Carlo simulations in canonical ensemble combined with variational transition state theory (VTST). We observe a big change of free energy and evaporation rate from tetramer to pentamer. A statistical analysis of the number of hydrogen bonds suggests that more hydrogen bonds are required to be broken in the evaporation reaction of tetramer compared to that of trimer and pentamer. Structure analysis indicates that, although the ground state of the pentamer has a two-dimensional ring structure, three-dimensional hydrogen bond network begins to form in pentamer at finite temperature. Therefore, it is a two-dimensional to three-dimensional transition from tetramer to pentamer. The fact that the most probable configuration of pentamer is different from the lowest energy configuration demonstrates the importance of molecular simulations. Simply finding the lowest energy configuration via global geometry optimization and then calculating the free energy within a harmonic approximation of vibrations are not a universal protocol for cluster systems. Methods used in this study are expected to be applicable for more complicated multicomponent systems, which opens an avenue for the research of particulate matter formation in atmosphere.

Key words: neural network, Monte Carlo, water, nucleation, enhanced sampling