机器学习分子动力学辅助材料凝固形核研究进展

陈名毅, 胡俊伟, 余耀辰, 牛海洋

Advances in Machine Learning Molecular Dynamics to Assist Materials Nucleation and Solidification Research

CHEN Mingyi, HU Junwei, YU Yaochen, NIU Haiyang

图3 CdSe结晶机制的研究[69]

Fig.3 Research on the nucleation mechanism of CdSe[69] (a) flowchart of the active learning procedure to train the deep neural network (DNN) potential (AIMD—ab initio molecular dynamics, DPMD/DP-WTMetaD—DNN-based molecular dynamics (DPMD) simulations and their variants accelerated by WTMetaD (DP-WTMetaD), DFT—density functional theory) (b) distribution of training set configurations in the two-dimensional space mapped with the principal component analysis (PCA) (Typical structures are marked in the plot. Green and brown spheres are Se and Cd, respectively. ZB—zinc blende, WZ—wurtzite) (c, d) error of the DNN potential relative to the DFT calculation. The values of mean absolute error (MAE) correspond to the potential energy (E) (c) and the force (F) (d) are stamped in the plot[69] (NN—neural network)