一种耦合CFD与深度学习的气固快速模拟方法

doi:10.11949/0438-1157.20230711

摘要/Abstract

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:

TQ 021.1

温凯杰, 郭力, 夏诏杰, 陈建华. 一种耦合CFD与深度学习的气固快速模拟方法[J]. 化工学报, 2023, 74(9): 3775-3785.

Kaijie WEN, Li GUO, Zhaojie XIA, Jianhua CHEN. A rapid simulation method of gas-solid flow by coupling CFD and deep learning[J]. CIESC Journal, 2023, 74(9): 3775-3785.

图/表 16

图1 ConvLSTM网络结构（修改自文献[19-20]）

Fig.1 The network structure of the ConvLSTM (Revised from Refs.[19-20])

图2 深度学习与OpenFOAM耦合流程图T—模拟物理时长;t—当前物理时长;tOF—OpenFOAM单次运行物理时长;tDNN—深度学习模型单次预测物理时长

Fig.2 Flowchart of coupling deep learning and OpenFOAM

图3 二维床示意图

Fig.3 Schematic drawing of 2D bed

表1 模拟体系主要参数

Table 1 Main parameters of the simulation

变量	数值
颗粒直径（d_p）/µm	275
固相密度（ρ_s）/(kg/m³)	2500
气相密度（ρ_g）/(kg/m³)	1.225
气相黏度（μ_g）/(Pa·s)	1.84×10^-5
反应器尺寸/m	0.28×1.5
床层高度/m	0.4
网格数量	56×300
模拟时间步长/s	2×10^-4
写入文件时间步长/s	0.01
总模拟时长/s	10

图4 气体速度为0.46 m/s时床层的时均空隙率径向分布对比

Fig.4 Comparison of radial distribution of time mean voidage at gas velocity of 0.46 m/s

表2 神经网络参数

Table 2 Parameters of the neural network

Name	Type	Kernel	Stride	Padding	Ch I/O
Encoder	CNN	3×3	1×1	1×1	5/16
	ConvLSTM	3×3	1×1	1×1	16/64
	CNN	3×3	2×2	1×1	64/64
	ConvLSTM	3×3	1×1	1×1	64/96
	CNN	3×3	2×2	1×1	96/96
	ConvLSTM	3×3	1×1	1×1	96/96
Forecast	ConvLSTM	3×3	1×1	1×1	96/96
	CNN	4×4	2×2	1×1	96/96
	ConvLSTM	3×3	1×1	1×1	96/96
	CNN	4×4	2×2	1×1	96/96
	ConvLSTM	3×3	1×1	1×1	96/64
	CNN	3×3	1×1	1×1	64/16
	CNN	1×1	1×1	0×0	16/5

图5 损失函数曲线

Fig.5 Variation of loss function

表3 测试集上的模型性能

Table 3 Model performance on the test set

Physical quantity	MSE	PSNR
ε_s	6.67×10^-4	31.76
U_g_x	1.04×10^-5	49.83
U_g_y	1.69×10^-5	47.72
U_s_x	1.53×10^-5	48.15
U_s_y	2.37×10^-5	46.25

图6 神经网络中εs物理场的输入(a)、预测的最后一帧(b)和目标(c)图像

Fig.6 Images of the input (a), the last predicted frame (b) and the corresponding target (c) of εs in the neural network

图7 神经网络预测值与目标值数据分布

Fig.7 The data distribution of predicted and target values of the neural network

图8 耦合计算流程

Fig.8 Calculating process of coupling DNN and OpenFOAM

图9 网格平均颗粒体积分数变化

Fig.9 Change of mesh average particle volume fraction

图10 深度学习模型预测εs、 Ug 、 Us 时固含率校正后的典型物理量轴向分布对比[(a)~(c)]及继续处理小量后的轴向分布对比[(d)~(f)]

Fig.10 Comparison of typical axial distribution after correction of solid volume fraction [(a)—(c)] and typical axial distribution after further treatment of very small values [(d)—(f)] when deep learning model predicts εs, Ug and Us

图11 深度学习模型预测εs、 Ug 、p时典型物理量的轴向分布对比

Fig.11 Comparison of typical axial distribution when deep learning model predicts εs, Ug and p

图12 不同深度学习模型预测跨度下εs轴向分布对比

Fig.12 Comparison of axial distribution of εs under different deep learning model prediction spans

图13 不同深度学习模型预测跨度下耦合流程的计算耗时与加速比

Fig.13 Calculation time and acceleration ratio of coupling process under different deep learning model prediction spans

参考文献 40

1	Steyaert S, Pizurica M, Nagaraj D, et al. Multimodal data fusion for cancer biomarker discovery with deep learning[J]. Nature Machine Intelligence, 2023, 5(4): 351-362.
2	Chen X Y, Lv H L. Intelligent control of nanoparticle synthesis on microfluidic chips with machine learning[J]. NPG Asia Materials, 2022, 14: 69.
3	Yu T H, Boob A G, Volk M J, et al. Machine learning-enabled retrobiosynthesis of molecules[J]. Nature Catalysis, 2023, 6(2): 137-151.
4	Karagiorgi G, Kasieczka G, Kravitz S, et al. Machine learning in the search for new fundamental physics[J]. Nature Reviews Physics, 2022, 4(6): 399-412.
5	Vinuesa R, Brunton S L. Enhancing computational fluid dynamics with machine learning[J]. Nature Computational Science, 2022, 2(6): 358-366.
6	Zhu L T, Chen X Z, Ouyang B, et al. Review of machine learning for hydrodynamics, transport, and reactions in multiphase flows and reactors[J]. Industrial & Engineering Chemistry Research, 2022, 61(28): 9901-9949.
7	Brunton S L, Noack B R, Koumoutsakos P. Machine learning for fluid mechanics[J]. Annual Review of Fluid Mechanics, 2020, 52: 477-508.
8	Kim B, Azevedo V C, Thuerey N, et al. Deep fluids: a generative network for parameterized fluid simulations[J]. Computer Graphics Forum, 2019, 38(2): 59-70.
9	Fukami K, Fukagata K, Taira K. Super-resolution reconstruction of turbulent flows with machine learning[J]. Journal of Fluid Mechanics, 2019, 870: 106-120.
10	Yang Z, Lu B N, Wang W. Coupling artificial neural network with EMMS drag for simulation of dense fluidized beds[J]. Chemical Engineering Science, 2021, 246: 117003.
11	Muralidhar N, Bu J, Cao Z, et al. Physics-guided deep learning for drag force prediction in dense fluid-particulate systems[J]. Big Data, 2020, 8(5): 431-449.
12	Guo L, Wu J, Li J H. Complexity at mesoscales: a common challenge in developing artificial intelligence[J]. Engineering, 2019, 5(5): 924-929.
13	Maulik R, Sharma H, Patel S, et al. A turbulent eddy-viscosity surrogate modeling framework for Reynolds-averaged Navier-Stokes simulations[J]. Computers & Fluids, 2021, 227: 104777.
14	Tompson J, Schlachter K, Sprechmann P, et al. Accelerating eulerian fluid simulation with convolutional networks[C]//Proceedings of the 34th International Conference on Machine Learning. New York, NY, USA: ACM, 2017: 3424-3433.
15	Ajuria I E, Alguacil A, Bauerheim M, et al. Towards an hybrid computational strategy based on deep learning for incompressible flows[C]//Proceedings of the AIAA Aviation 2020 Forum. Reston, Virginia: AIAA, 2020: AIAA2020-3058.
16	Obiols-Sales O, Vishnu A, Malaya N, et al. CFDNet: a deep learning-based accelerator for fluid simulations[C]//Proceedings of the 34th ACM International Conference on Supercomputing. New York, NY, USA: ACM, 2020: 1-12.
17	Jeon J, Lee J, Kim S J. Finite volume method network for the acceleration of unsteady computational fluid dynamics: non-reacting and reacting flows[J]. International Journal of Energy Research, 2022, 46(8): 10770-10795.
18	Bazai H, Kargar E, Mehrabi M. Using an encoder-decoder convolutional neural network to predict the solid holdup patterns in a pseudo-2D fluidized bed[J]. Chemical Engineering Science, 2021, 246: 116886.
19	Shi X J, Gao Z H, Lausen L, et al. Deep learning for precipitation nowcasting: a benchmark and a new model[C]// Proceedings of the 31st International Conference on Neural Information Processing Systems. New York, NY, USA: Curran Associates Inc., 2017: 5622-5632.
20	Shi X J, Chen Z R, Wang H, et al. Convolutional LSTM network: a machine learning approach for precipitation nowcasting[C]//Proceedings of the 28th International Conference on Neural Information Processing Systems. Cambridge, MA, USA: MIT Press, 2015: 802-810.
21	Wang Y B, Long M S, Wang J M, et al. PredRNN: recurrent neural networks for predictive learning using spatiotemporal LSTMs[C]//Proceedings of the 31st International Conference on Neural Information Processing Systems. New York, NY, USA: Curran Associates Inc., 2017: 879–888.
22	Wang Y B, Gao Z F, Long M S, et al. PredRNN++: towards a resolution of the deep-in-time dilemma in spatiotemporal predictive learning[C]// Proceedings of the 35th International Conference on Machine Learning. New York, NY, USA: ACM, 2018: 5123-5132.
23	Yu B, Yin H T, Zhu Z X. Spatio-temporal graph convolutional networks: a deep learning framework for traffic forecasting[C]//Proceedings of the 27th International Joint Conference on Artificial Intelligence. Palo Alto, CA, USA: AAAI Press, 2018: 3634–3640.
24	Zhou S Y, Xie W H, Lu Y X, et al. ConvLSTM-based wave forecasts in the South and East China Seas[J]. Frontiers in Marine Science, 2021, 8: 680079.
25	Gonzalez F J, Balajewicz M. Deep convolutional recurrent autoencoders for learning low-dimensional feature dynamics of fluid systems[EB/OL]. 2018, arXiv: 1808.01346. .
26	Weiner A, Semaan R. flowTorch—a Python library for analysis and reduced-order modeling of fluid flows[J]. Journal of Open Source Software, 2021, 6(68): 3860.
27	Maulik R, Sharma H, Patel S, et al. Deploying deep learning in OpenFOAM with TensorFlow[C]//Proceedings of the AIAA Scitech 2021 Forum. Reston, Virginia: AIAA, 2021: AIAA2021-1485.
28	Jeon J, Lee J, Eivazi H, et al. Physics-informed transfer learning strategy to accelerate unsteady fluid flow simulations[EB/OL]. 2022, arXiv: 2206.06817. .
29	Lu L Q, Gao X, Dietiker J F, et al. Development of a filtered CFD-DEM drag model with multiscale markers using an artificial neural network and nonlinear regression[J]. Industrial & Engineering Chemistry Research, 2022, 61(1): 882-893.
30	Hochreiter S, Schmidhuber J. Long short-term memory[J]. Neural Computation, 1997, 9(8): 1735-1780.
31	王雅琳, 潘雨晴, 刘晨亮. 基于GSA-LSTM动态结构特征提取的间歇过程监测方法[J]. 化工学报, 2022, 73(9): 3994-4002.
	Wang Y L, Pan Y Q, Liu C L. Intermittent process monitoring based on GSA-LSTM dynamic structure feature extraction[J]. CIESC Journal, 2022, 73(9): 3994-4002.
32	刘立邦, 杨颂, 王志坚, 等. 基于改进WOA-LSTM的焦炭质量预测[J]. 化工学报, 2022, 73(3): 1291-1299.
	Liu L B, Yang S, Wang Z J, et al. Prediction of coke quality based on improved WOA-LSTM[J]. CIESC Journal, 2022, 73(3): 1291-1299.
33	Cho K, van Merrienboer B, Gulcehre C, et al. Learning phrase representations using RNN encoder–decoder for statistical machine translation[C]//Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP). Stroudsburg, PA, USA: Association for Computational Linguistics, 2014: 1724-1734.
34	van der Hoef M A, van Sint Annaland M, Deen N G, et al. Numerical simulation of dense gas-solid fluidized beds: a multiscale modeling strategy[J]. Annual Review of Fluid Mechanics, 2008, 40: 47-70.
35	Ishii M. Thermo-fluid Dynamic Theory of Two-phase Flow[M]. Paris: Evrolles, 1975.
36	Derivation Wachem B., implementation, and validation of computer simulation models for gas-solid fluidized beds[D]. Amsterdam: Delft University of Technology, 2000.
37	Taghipour F, Ellis N, Wong C. Experimental and computational study of gas-solid fluidized bed hydrodynamics[J]. Chemical Engineering Science, 2005, 60(24): 6857-6867.
38	Shin D, Gupta S K. A new circuit simplification method for error tolerant applications[C]//2011 Design, Automation & Test in Europe. Washington, DC, USA: IEEE, 2011: 1-6.
39	Moosazadeh M, Andalib A. A new robust color digital image watermarking algorithm in DCT domain using genetic algorithm and coefficients exchange approach[C]//2016 Second International Conference on Web Research (ICWR). Washington, DC, USA: IEEE, 2016: 19-24.
40	Shih T K, Chang R C, Lu L C, et al. Adaptive digital image inpainting[C]//18th International Conference on Advanced Information Networking and Applications. Washington, DC, USA: IEEE, 2004: 71-76.

[1]	张思雨, 殷勇高, 贾鹏琦, 叶威. 双U型地埋管群跨季节蓄热特性研究[J]. 化工学报, 2023, 74(S1): 295-301.
[2]	肖明堃, 杨光, 黄永华, 吴静怡. 浸没孔液氧气泡动力学数值研究[J]. 化工学报, 2023, 74(S1): 87-95.
[3]	周绍华, 詹飞龙, 丁国良, 张浩, 邵艳坡, 刘艳涛, 郜哲明. 短管节流阀内流动噪声的实验研究及降噪措施[J]. 化工学报, 2023, 74(S1): 113-121.
[4]	江河, 袁俊飞, 王林, 邢谷雨. 均流腔结构对微细通道内相变流动特性影响的实验研究[J]. 化工学报, 2023, 74(S1): 235-244.
[5]	何松, 刘乔迈, 谢广烁, 王斯民, 肖娟. 高浓度水煤浆管道气膜减阻两相流模拟及代理辅助优化[J]. 化工学报, 2023, 74(9): 3766-3774.
[6]	王玉兵, 李杰, 詹宏波, 朱光亚, 张大林. R134a在菱形离散肋微小通道内的流动沸腾换热实验研究[J]. 化工学报, 2023, 74(9): 3797-3806.
[7]	袁佳琦, 刘政, 黄锐, 张乐福, 贺登辉. 泡状入流条件下旋流泵能量转换特性研究[J]. 化工学报, 2023, 74(9): 3807-3820.
[8]	尹刚, 李伊惠, 何飞, 曹文琦, 王民, 颜非亚, 向禹, 卢剑, 罗斌, 卢润廷. 基于KPCA和SVM的铝电解槽漏槽事故预警方法[J]. 化工学报, 2023, 74(8): 3419-3428.
[9]	岳林静, 廖艺涵, 薛源, 李雪洁, 李玉星, 刘翠伟. 凹坑缺陷对厚孔板喉部空化流动特性影响研究[J]. 化工学报, 2023, 74(8): 3292-3308.
[10]	邢雷, 苗春雨, 蒋明虎, 赵立新, 李新亚. 井下微型气液旋流分离器优化设计与性能分析[J]. 化工学报, 2023, 74(8): 3394-3406.
[11]	高燕, 伍鹏, 尚超, 胡泽君, 陈晓东. 基于双流体喷嘴的磁性琼脂糖微球的制备及其蛋白吸附性能探究[J]. 化工学报, 2023, 74(8): 3457-3471.
[12]	诸程瑛, 王振雷. 基于改进深度强化学习的乙烯裂解炉操作优化[J]. 化工学报, 2023, 74(8): 3429-3437.
[13]	闫琳琦, 王振雷. 基于STA-BiLSTM-LightGBM组合模型的多步预测软测量建模[J]. 化工学报, 2023, 74(8): 3407-3418.
[14]	汪林正, 陆俞冰, 张睿智, 罗永浩. 基于分子动力学模拟的VOCs热氧化特性分析[J]. 化工学报, 2023, 74(8): 3242-3255.
[15]	徐野, 黄文君, 米俊芃, 申川川, 金建祥. 多源信息融合的离心式压缩机喘振诊断方法[J]. 化工学报, 2023, 74(7): 2979-2987.