Stratified Construction of Neural Network Potentials and Co-Evolutionary Global Structure Optimization for Acceleration of Ab Initio Materials Prediction

摘要

Density functional theory has gained prominence in the computational community as a successful materials modeling and prediction tool, demonstrated by numerous successful applications of the theory. However, computationally expensive scaling of the approach with the system size is restricting its widespread use, particularly, in the unconstrained search for novel materials. This limitation motivated the development of accurate and efficient interatomic interaction models for acceleration of the first- principles materials prediction. Neural network interatomic potentials, proposed as an alternative means of total energy calculation, have shown promising performance in mapping the ab initio energy landscape. Present challenges, such as reliable sampling of configuration space for generating training dataset and efficient construction of neural network models are topics of active research in the community.The present work aims to address the methodological challenges in constructing neural network potentials, improve the global structure optimization efficiency, ex- amine the developed methodologies' performance in predicting new materials, and explore the applicability of first-principles approaches in materials modeling.To improve the efficiency and reliability of the neural network potentials, a stratified training scheme for the construction of the interaction model for multicomponent systems and an evolutionary scheme for dataset generation are introduced. Further- more, a multitribe co-evolutionary approach is developed to accelerate the global optimization via promoting the suitability of the search results with exchanging the best candidate members among the nearby populations. After extensive benchmark- ing tests, the proposed method developments have been employed for a systematic comparison of neural network potentials and a widely used empirical potential performance in identifying the ab initio ground states of Cu-Pd-Ag nanoparticles. The search results revealed that neural network potentials significantly outperform the empirical Gupta model in guiding the first-principle search for novel phases and that the co-evolutionary global optimization scheme results in substantial improvement in convergence rate of the search process. As an example of using first-principle methods for guiding materials discovery and interpreting experimental results, density functional theory is employed to study a synthesized high-pressure phase of near-stoichiometric LiB compound. The simulations of the complex evolution of the compound success- fully interpreted the experimental observations and verified a previously predicted phase.