Massively parallelization strategy for material simulation using high‐dimensional neural network potential. Issue 10 (10th November 2018)

• Record Type: Journal Article
• Authors: Shang, Cheng; Huang, Si‐Da; Liu, Zhi‐Pan
• Other Names: Mo Yirong guestEditor.; Wu Wei guestEditor.; Cao Zexing guestEditor.
• Is Part Of: Journal of computational chemistry. Volume 40:Issue 10(2019)
• Page Start: 1091
• Page End: 1096
• Publication Date: 2018-11-10
• Subjects: neural network -- parallelization -- structure descriptor; Chemistry -- Data processing -- Periodicals; 542.85
• DOI: 10.1002/jcc.25636
• Languages: English
• ISSNs: 0192-8651

Abstract

Abstract : The potential energy surface (PES) calculation is the bottleneck for modern material simulation. The high‐dimensional neural network (HDNN) technique emerged recently appears to be a problem solver for fast and accurate PES computation. The major cost of the HDNN lies at the computation of the structural descriptors that capture the geometrical environment of atoms. Here, we introduce a massive parallelization strategy optimized for our recently developed power‐type structural descriptor. The method involves three‐levels: from the top to the bottom the parallelization is over atoms first, then, over structural descriptors and finally over the n ‐body functions. We illustrate the parallelization method in a boron crystal system and show that the parallelization efficiency is maximally 100%, 58%, and 34% at each level. © 2018 Wiley Periodicals, Inc. Abstract : A massive parallelization strategy is designed to speed up the computation of high‐dimensional neural network based on the power‐type structural descriptors. The parallelization framework is designed to be hierarchical with three‐levels to make the best use of computational resources. Using this framework, the authors achieved a speed‐up of 411 on 560 cores for the neural network computation of a 560‐atom Boron solid.