中山大学学报(自然科学版) ›› 2023, Vol. 62 ›› Issue (5): 67-77.doi: 10.13471/j.cnki.acta.snus.2022A064

• 研究论文 • 上一篇    下一篇

面向边缘计算的神经网络MPS加速算法

王卓霖1(), 江俊扬1, 杨耿超1(), 姚清河1(), 蒋子超1, 张仪2   

  1. 1.中山大学航空航天学院,广东 深圳 518107
    2.特文特大学机器人和机电一体化系,荷兰 恩斯赫德 7522NB
  • 收稿日期:2022-07-27 接受日期:2023-01-03 出版日期:2023-09-25 发布日期:2023-09-28
  • 通讯作者: 杨耿超,姚清河 E-mail:wangzhlin3@mail2.sysu.edu.cn;yanggch8@mail.sysu.edu.cn;yaoqhe@mail.sysu.edu.cn
  • 作者简介:王卓霖(1998年生),男;研究方向:计算流体力学;E-mail:wangzhlin3@mail2.sysu.edu.cn
  • 基金资助:
    国家自然科学基金(11972384);广东省基础与应用基础研究基金(2021B1515310001);国家重点研发计划(2020YFA0712502)

An MPS algorithm accelerated by neural network on edge computing

Zhuolin WANG1(), Junyang JIANG1, Gengchao YANG1(), Qinghe YAO1(), Zichao JIANG1, Yi ZHANG2   

  1. 1.School of Aeronautics and Astronautics,Sun Yat-sen University,Shenzhen 518107,China
    2.Department of Robotics and Mechatronics,University of Twente,Enschede,7522 NB,the Netherlands
  • Received:2022-07-27 Accepted:2023-01-03 Online:2023-09-25 Published:2023-09-28
  • Contact: Gengchao YANG,Qinghe YAO E-mail:wangzhlin3@mail2.sysu.edu.cn;yanggch8@mail.sysu.edu.cn;yaoqhe@mail.sysu.edu.cn

摘要:

在无网格法中,移动粒子半隐式方法(MPS)利用流体的不可压性构造泊松方程,在获得了压力求解准确性的同时也带来了巨大的计算量,导致其不适合实现大规模流体模拟. 针对上述问题提出新型算法NN-MPS,将泊松方程的求解转化为利用神经网络求解回归问题. NN-MPS算法通过构建每一步的流场特征与压力的预测模型,实现泊松方程的快速求解. 本文进一步将基于NN-MPS算法的泊松方程求解过程移植到Atlas 200 DK上,实现边缘侧加速求解泊松方程. 本文采用多种溃坝模型进行数值实验,结果表明,本文的MPS加速方法具有低成本、高速度且较少精度损失的特点,求解速度实现了一定的提升. 本文同时也验证了边缘计算设备在计算流体力学领域应用的可行性.

关键词: 边缘计算, 神经网络, NN-MPS, 计算流体力学

Abstract:

As a meshless method, the moving particle semi-implicit(MPS) method forms the pressure Poisson equation by using the incompressibility of fluid, which not only obtain the accuracy of pressure but also bring a high cost of calculation. Therefore, it is not appropriate for MPS to solve large-scale fluid simulations. IA new algorithm NN-MPS is proposed to solve the above problems, which transforms the solution of the Poisson equation into a regression problem by using neural network. The NN-MPS algorithm realizes the quick solution of the Poisson equation by constructing the prediction model of flow field features and pressure at each step. In this work, the procedure of solving the pressure Poisson equation is further transported to Atlas 200 DK device for a faster speed of solving procedure. Results show that the acceleration method of MPS mentioned in this work has the characteristics of low cost, high speed, and low accuracy loss, and the solution speed has been improved. We also verified the feasibility of applying the edge computing device in the field of CFD.

Key words: edge computing, neural network, NN-MPS, CFD

中图分类号: