化工学报 ›› 2023, Vol. 74 ›› Issue (9): 3775-3785.DOI: 10.11949/0438-1157.20230711

• 流体力学与传递现象 • 上一篇    下一篇

一种耦合CFD与深度学习的气固快速模拟方法

温凯杰1,2(), 郭力1,2, 夏诏杰1,2, 陈建华1()   

  1. 1.中国科学院过程工程研究所多相复杂系统国家重点实验室,北京 100190
    2.中国科学院大学化学工程学院,北京 100049
  • 收稿日期:2023-07-10 修回日期:2023-09-02 出版日期:2023-09-25 发布日期:2023-11-20
  • 通讯作者: 陈建华
  • 作者简介:温凯杰(1998—),女,硕士研究生,wenkaijie21@mails.ucas.ac.cn
  • 基金资助:
    国家自然科学基金面上项目(22078327)

A rapid simulation method of gas-solid flow by coupling CFD and deep learning

Kaijie WEN1,2(), Li GUO1,2, Zhaojie XIA1,2, Jianhua CHEN1()   

  1. 1.State Key Laboratory of Multiphase Complex System, Institute of Process Engineering, Chinese Academy of Sciences, Beijing 100190, China
    2.School of Chemical Engineering, University of Chinese Academy of Sciences, Beijing 100049, China
  • Received:2023-07-10 Revised:2023-09-02 Online:2023-09-25 Published:2023-11-20
  • Contact: Jianhua CHEN

摘要:

在计算流体力学领域,深度学习被用于重建流场、预测曳力、求解泊松方程、加速流体模拟等方面的研究。为了加速气固两相流的模拟计算,使用卷积长短时记忆网络对物理量进行预测,并基于LibTorch实现深度学习模型预测与OpenFOAM的耦合。通过与单纯OpenFOAM模拟结果对比,发现深度学习模型预测存在颗粒体积分数不守恒、极小数值预测不准确的问题,先后通过体积分数校正和网格数据过滤消除了前述影响。选取不同的三个物理量组合进行深度学习模型预测以加速CFD计算,在同样选取颗粒体积分数和气体速度的条件下,对比了增加预测颗粒速度和压力对耦合流程计算结果的影响,其中,增加预测颗粒速度的计算结果较为准确,经分析发现这可能与求解器对不同变量的调用顺序有关。此外,研究了不同深度学习模型预测跨度比(1∶1、3∶1、5∶1、10∶1)下的耦合计算结果以及加速效果,发现在一定误差范围内,当前耦合计算流程最高可实现9倍左右的加速,且加速比与跨度比呈近似线性关系。

关键词: 两相流, 计算流体力学, 神经网络, 加速计算, 深度学习

Abstract:

In the field of computational fluid dynamics (CFD), deep learning has been used for reconstructing flow field, predicting drag force, solving the Poisson equation, accelerating simulation and so on. In order to accelerate the simulation of gas-solid two-phase flow, the neural network of convolutional long short-term memory was used to predict physics quantities and it was coupled with OpenFOAM through LibTorch. Comparing the results of coupled calculations with those of pure OpenFOAM calculations, it was found that the present deep learning model prediction had problems such as non-conservative particle volume fraction and inaccuracy of very small values. These problems were eliminated through volume fraction correction and grid data filtering. Then different combinations of three physical quantities were selected for accelerating CFD by the deep learning model prediction. Given model prediction of the particle volume fraction and the gas velocity, the influence of adding prediction of the particle velocity and the pressure on the results of the coupling process was compared. It was found that the former was more accurate, and the difference may be related to the invoking order of different variables in the solver. In addition, the coupled calculation results and acceleration effects under different deep learning model prediction span ratios (1∶1, 3∶1, 5∶1, 10∶1) were studied, and it was found that within a certain error range, the current calculation process can accelerate by about 9 times, and the acceleration ratio has an approximately linear relationship with the span ratio.

Key words: two-phase flow, CFD, neural network, computing acceleration, deep learning

中图分类号: