Construction And Application Of Precise Atomic Potential Based On "Deepmd-kit

The preparation of materials is the foundation for the development of new materials.Although various advanced experimental methods have been used for the preparation of new materials,the determination of experimental parameters is still based on the experience of researchers.With the improvement of computer computing power,molecular dynamics simulation provides a virtual means for the preparation of new materials,which can significantly reduce experimental costs and accelerate the development of new materials.However,traditional molecular dynamics simulations cannot effectively guide experiments due to inaccurate results,while molecular dynamics simulations based on first principles calculations cannot effectively simulate the preparation of actual materials due to excessive computational resource consumption.Therefore,how to balance computational accuracy and efficiency has become one of the difficulties in the field of molecular dynamics simulation.With the emergence of machine learning,the use of deep learning to train neural networks to represent atomic interaction potentials has provided the possibility to solve the aforementioned difficulties.In this article,we take the dynamic process of germanium（Ge）atoms on graphene/copper（111）substrates as the research object,and use deep potential molecular dynamics（DPMD）method to train deep neural networks using first-principles data from small-scale systems.We obtain a deep neural network potential that can accurately simulate large-scale systems.The main results are as follows:（1）We found that the adsorption energy of IIIA-VA main group metal atoms on a monolayer graphene/Cu（111）substrate was enhanced by more than 0.15 e V compared to monolayer graphene,especially the adsorption energy of Ge atoms on graphene/Cu（111）was increased by more than 0.27 e V compared to monolayer graphene,indicating that the substrate played a significant enhancing role in the process of monolayer graphene adsorbing metal atoms.When graphene on the substrate undergoes ripple due to stress regulation,the Ge atoms at the bottom of the fluctuations aggregate and undergo quasi one-dimensional restricted growth due to enhanced adsorption by the copper substrate.（2）The neural network interatomic interaction potential of Ge epitaxial growth model on graphene/Cu（111）was obtained using the DPMD method,with an average energy error of3.505×10^-4e V and force error of 4.627×10^-2e V/?for each atom,achieving the accuracy of first-principles calculations.We use the deep neural network potential to simulate the system with the lattice constant of 13.302?for a long time,and the obtained radial distribution function of the atom is very consistent with the results of the first principle calculation,which proves that the deep neural network potential can accelerate the molecular dynamics simulation while maintaining the computational accuracy.When deep neural network potentials are used to simulate the system of 70.944?×61.439?×31.038?at high temperatures,there will be some atomic bond breaking phenomena,indicating that the migration of deep neural network potentials in large-scale systems is a problem that should be paid attention to in new molecular dynamics simulation methods based on machine learning.We will use reinforcement learning in future research to address this issue.