TensorKMC: Kinetic Monte Carlo Simulation of 50 Trillion Atoms Driven by Deep Learning on a New Generation of Sunway Supercomputer

Authors: Honghui Shang (Institute of Computing Technology, Chinese Academy of Sciences); Xin Chen and Xingyu Gao (Institute of Applied Physics and Computational Mathematics); Rongfen Lin (Tsinghua University, China); Lifang Wang (Institute of Applied Physics and Computational Mathematics); Fang Li, Qian Xiao, and Qiang Sun (National Supercomputing Center in Wuxi); Lei Xu (Institute of Computing Technology, Chinese Academy of Sciences); Leilei Zhu (University of Science and Technology of China); Fei Wang (Tsinghua University, China); Yunquan Zhang (Institute of Computing Technology, Chinese Academy of Sciences); and Haifeng Song (Institute of Applied Physics and Computational Mathematics)

Abstract: The atomic dynamics Monte Carlo (AKMC) method has played an important role in the multi-scale physical simulation; it is the bridge connecting the micro and macro worlds. Its accuracy is limited, however, by the empirical potentials. In this work, we propose a vacancy-centered tabulation algorithm to efficiently integrate ab initio fitted neural network potentials (NNPs) with AKMC. We port this to SW26010-pro and optimize the neural network work potentials with big fusion strategy for the new Sunway heterogenous computing units. We further optimize the memory usage to make TensorKMC capable of simulating up to 50 trillion atoms. TensorKMC can achieve excellent strong scaling performance using up to 24,960,000 cores, and linear weak scaling using up to 27,456,400 cores.


Presentation: file


Back to Technical Papers Archive Listing