The 8th Forum of Materials Genome Engineering

GPUMD software and its applications to materials calculations

EXTENDED ABSTRACT: Efficient materials calculations, advanced experiments, and big data techniques are the foundations for materials genome engineering, and molecular dynamics (MD) simulation is one of the indispensable techniques for efficient materials calculations. Graphics processing units (GPUs) have played an important role in accelerating MD simulations. During the past several years, we have developed the graphics processing units molecular dynamics (GPUMD) software fully utilizing the computational power of GPUs [1]. A notable feature of GPUMD is that it contains a complete framework for generating and using the neuroevolution potential (NEP), which is a highly efficient machine-learning potential (MLP) [2]. First, we discuss one of the core techniques in GPUMD, i.e., a formalism for many-body potentials that is suitable for efficient GPU implementation. This formalism establishes the correct heat current for many-body potentials, laying down the foundation for accurate modelling of heat transport properties of materials. Then, we discuss the theoretical formalism and efficient implementation of NEP in GPUMD. NEP is a neural network potential with a single hidden layer, as implied by the GPUMD logo in Fig. 1. The training of NEP is based on an evolutionary algorithm, which does not depend on third-party packages such as TensorFlow and PyTorch, and we have achieved extremely high computational efficiency by employing native CUDA programming. Next, we discuss the applications of GPUMD and NEP in modelling the various properties of materials, such as mechanical, thermal, electrical, phase-transition, growth, irradiation, catalysis, and spectra. Last, we point out promising future works that could be done, based on GPUMD and NEP, in the field of materials calculations.

Keywords: molecular dynamics simulation; machine-learning potential; materials calculation
REFERENCES:
[1]Zheyong Fan et al., Comput. Phys. Commun., 218(10), (2017) 10-16
[2]Zheyong Fan et al., Phys. Rev. B, 104(10) (2021) 104309

Brief Introduction of Speaker

Zheyong Fan

Zheyong Fan has completed his PhD from Nanjing University in 2010 and did postdoctoral studies at Xiamen University and Aalto University. He is now a full professor at Bohai University, working on developing molecular dynamics methods and software, as well as a comprehensive machine-learning potential model for the periodic table. He has developed the GPUMD software and the NEP machine-learning potential framework. He has published over 100 peer-viewed papers with a total citation over 4000 and an H-index of 36 (Google Scholar), and has been in the Top 2% Scientist list for 2023 and 2024.