基于AutoML的湍流建模

doi:10.11871/jfdc.issn.2096-742X.2020.04.010

数据与计算发展前沿 ›› 2020, Vol. 2 ›› Issue (4): 121-131.

doi: 10.11871/jfdc.issn.2096-742X.2020.04.010

所属专题：下一代互联网络技术与应用

基于AutoML的湍流建模

任荟颖^1,²(),王婧¹(),王彦棡^1,^*()

1.中国科学院计算机网络信息中心, 北京 100190
2.中国科学院大学, 北京 100049

收稿日期:2020-03-30 出版日期:2020-08-20 发布日期:2020-09-10
通讯作者: 王彦棡
作者简介:任荟颖,中国科学院计算机网络信息中心,在读硕士研究生,主要研究方向为人工智能应用。
本文主要承担工作：实验设计,论文撰写。
Ren Huiying, is a graduate student of the at Computer Network Information Center of Chinese Academy of Sciences. Her main research direction is artificial intelligence application.
In this paper, she is responsible for experiment design and paper writing.
E-mail: renhuiying@cnic.cn|王婧,中国科学院计算机网络信息中心,高级工程师,主要研究方向为人工智能算法应用、多媒体信息检索及大数据分析。
本文主要承担工作：论文撰写。
Wang Jing, M.S., is a senior engineer of Computer Network information Center, Chinese Academy of Sciences. Her research interests include artificial intelligence algorithm application, multimedia information retrieval and big data analysis.
In this paper, she is responsible for paper writing.
E-mail: wangjing@cnic.cn|王彦棡,中国科学院计算机网络信息中心,博士,研究员,主要研究方向为人工智能应用,高性能计算。
本文主要承担工作：科学问题凝结,把握论文总体结构。
Wang Yangang, PHD, is a research fellow at Computer Network Information Center of Chinese Academy of Sciences. His main research directions are artificial intelligence application and high-performance computing.
In this paper, he is responsible for condensing scientific problems and grasping the overall structure of the paper.
E-mail: wangyg@sccas.cn
基金资助:
国家重点研发计划“大规模并行计算的工具库和领域相关基础软件包”(2017YFB0202202);“中国科技云”建设工程(二期)超算资源池建设(XXH13503)

Turbulence Modeling Based on AutoML

Ren Huiying^1,²(),Wang Jing¹(),Wang Yangang^1,^*()

1. Computer Network Information Center, Chinese Academy of Sciences, Beijing 100190, China
2. University of Chinese Academy of Sciences, Beijing 100049, China

Received:2020-03-30 Online:2020-08-20 Published:2020-09-10
Contact: Wang Yangang

摘要/Abstract

摘要：

【背景】 湍流问题涉及到工程中的诸多领域,其重要性不言而喻。雷诺平均Navier-Stokes(RANS)方程提供了一种计算时间平均湍流量的有效方法,由于其计算易处理性而被广泛使用。随着深度学习技术的发展,采用数据驱动的方法建模RANS模型受到了研究者广泛的关注。【方法】本文提出了一种数据驱动建模RANS模型的方法,该方法以数值软件模拟结果为基础,利用深度学习技术构造湍流模型。由于在湍流问题中,不同的系统初始条件不同,数据的质量千差万别,难以使用统一的神经网络结构进行训练。因此本文采用AutoML(自动机器学习)的方法自动搜索神经网络的结构并进行自动调参。此外,本文发现通过混合多种初始条件下的数据进行模型训练,可以提高深度学习模型的拟合精度,增强其鲁棒性。【结果】本文选取OpenFOAM中的经典算例内壁台阶流模拟作为数据来源进行实验。实验表明,该模型在预测雷诺应力时具有较好的精度和效率,表明数据驱动方法在湍流模拟中具有良好的应用前景。【局限】为了更好的在湍流领域应用深度学习技术,下一步的研究重点在于如何将深度学习模型与湍流数值模拟软件耦合。【结论】目前,针对湍流机器学习的系统研究相对较少。在现有工作经验的基础上,机器学习在未来的湍流模型化中必将扮演着更加重要的角色。

关键词: 湍流模型, 自动机器学习, 数据驱动, 雷诺平均方程, 深度学习

Abstract:

[Background] The turbulence problem involves many fields in engineering, and its importance is self-evident. The Reynolds Average Navier-Stokes (RANS) equation provides an effective method for calculating the time-averaged turbulence, and it is widely used because of its ease in calculation. With the development of deep learning technology, data-driven modeling of RANS model has been widely concerned by researchers. [Methods] In this paper, a data-driven method for modeling RANS is proposed. Based on the results of numerical simulation software, the method uses deep learning technology to construct turbulence models. Because of the different initial conditions of different turbulence systems and various qualities of data, it is difficult to use a unified neural network structures for training. Therefore, we used the method of AutoML (automatic machine learning) in deep learning to solve the problem, which can automatically build appropriate network structures and choose proper hyper-parameters for different datasets. In addition, this paper also improves the method by mixing the data under various initial conditions to train the deep learning model, which greatly increases accuracy and the robustness of our model. [Results] In this paper, a typical example in OpenFOAM about the step flow simulation of the inner wall is selected as the data source for the experiment. The experimental results show that the model has good accuracy and efficiency in prediction of Reynolds stress, which indicates that the data-driven method has a great application prospect in turbulence simulation. [Limitations] In order to better apply deep learning technology in the field of turbulence modeling, the most important issue needs to be solved in the next step is how to couple deep learning models with turbulence numerical simulation software. [Conclusions] At present, there are few systematic researches about turbulence machine learning. Based on the current work, machine learning will play a more important role in the future turbulence modeling researches.

Key words: turbulence modeling, auto machine learning, data-driven, RANS, deep learning

任荟颖,王婧,王彦棡. 基于AutoML的湍流建模[J]. 数据与计算发展前沿, 2020, 2(4): 121-131.

Ren Huiying,Wang Jing,Wang Yangang. Turbulence Modeling Based on AutoML[J]. Frontiers of Data and Computing, 2020, 2(4): 121-131.

图/表 14

图1

图2

图3

图4

图5

表1

表2

图6

表3

图7

图8

图9

图10

图11

参考文献 29

[1]	李树深 . 数据与计算是科技创新的巨大驱动力[J]. 数据与计算发展前沿, 2019,1(1):1.DOI: 10.11871/jfdc.issn.2096-742X.2019.01.001.PID:21.86101.2/jfdc.2096-742X.2019.01.001.
[2]	廖方宇, 洪学海, 汪洋, 褚大伟 . 数据与计算平台是驱动当代科学研究发展的重要基础设施[J]. 数据与计算发展前沿, 2019,1(1):2-10.DOI: 10.11871/jfdc.10-1649.2019.01.002.PID:21.86101.2/jfdc.10-1649.2019.01.002.
[3]	王彦棡, 王珏, 曹荣强 . 人工智能计算与数据服务平台的研究与应用[J]. 数据与计算发展前沿, 2019,1(2):86-97.DOI: 10.11871/jfdc.issn.2096-742X.2019.02.008.PID:21.86101.2/jfdc.2096-742X.2019.02.008.
[4]	POPE S B J T F . Turbulent Flows [M]. IOP Publishing, 2001.
[5]	TRACEY B, DURAISAMY K , ALONSO J of Conference. Application of Supervised Learning to Quantify Uncertainties in Turbulence and Combustion Modeling [C]. AIAA Aerospace Sciences Meeting Including the New Horizons Forum & Aerospace Exposition. 2013: 259.
[6]	TRACEY B D, DURAISAMY K , ALONSO J of Conference. A machine learning strategy to assist turbulence model development [C]. 53rd AIAA aerospace sciences meeting. 2015: 1287.
[7]	MING M, LU J, TRYGGVASON G . Using statistical learning to close two-fluid multiphase flow equations for a simple bubbly system[J]. Physics of Fluids, 2015,27(9):092101. doi: 10.1063/1.4930004
[8]	ZHANG Z J , DURAISAMY K of Conference. Machine Learning Methods for Data-Driven Turbulence Modeling [C]. AIAA Computational Fluid Dynamics Conference. 2015: 2460.
[9]	LING J, TEMPLETON J J P O F . Evaluation of machine learning algorithms for prediction of regions of high Reynolds averaged Navier Stokes uncertainty[J]. 2015,27(8):085103.
[10]	LING J, KURZAWSKI A, TEMPLETON J J J O F M . Reynolds averaged turbulence modelling using deep neural networks with embedded invariance[J]. 2016,807:155-166.
[11]	KUTZ J N . Deep learning in fluid dynamics[J]. Journal of Fluid Mechanics, 2017,814:1-4. doi: 10.1017/jfm.2016.803
[12]	CHANG C-W, DINH N T . Classification of machine learning frameworks for data-driven thermal fluid models[J]. International Journal of Thermal Sciences, 2019,135:559-579. doi: 10.1016/j.ijthermalsci.2018.09.002
[13]	CHANG C-W, DINH N T . Reynolds-averaged turbulence modeling using type I and type II machine learning frameworks with deep learning[J]. arXiv preprint arXiv:1804.01065, 2018.
[14]	BERGSTRA J, KOMER B, ELIASMITH C , et al. Hyperopt: a python library for model selection and hyperparameter optimization[J]. 2015,8(1):014008.
[15]	HUTTER F, HOOS H , LEYTON-BROWN K of Conference. An evaluation of sequential model-based optimization for expensive blackbox functions [C]. Proceedings of the 15th annual conference companion on Genetic and evolutionary computation. 2013: 1209-1216.
[16]	JIN H, SONG Q , HU X of Conference. Auto-keras: An efficient neural architecture search system[C]. Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 2019: 1946-1956.
[17]	PHAM H, GUAN M Y, ZOPH B , et al. Efficient neural architecture search via parameter sharing [J]. arXiv preprint arXiv:1802.03268, 2018.
[18]	MOCKUS I . On Bayesian methods of seeking an extremum[M], Optimization techniques IFIP technical conference. Springer, 2020.
[19]	RASMUSSEN C E, WILLIAMS C K I . Gaussian Processes for Machine Learning[M]. Berlin, Springer, 2003: 63-71.
[20]	KUTZ J N J J O F M . Deep learning in fluid dynamics[J]. 2017,814:1-4.
[21]	张伟伟, 朱林阳, 刘溢浪, 寇家庆 . 机器学习在湍流模型构建中的应用进展[J]. 空气动力学学报, 2019,37(03):444-454 .
[22]	VOS R, FAROKHI S . Introduction to Transonic Aerodynamics[M]. Springer, 2015.
[23]	PITZ R W, DAILY J W . Combustion in a turbulent mixing layer formed at a rearward-facing step[J]. AIAA Journal, 1983,21(11):1565-1570. doi: 10.2514/3.8290
[24]	AHMED U, PROSSER R . A Posteriori Assessment of Algebraic Scalar Dissipation Models for RANS Simulation of Premixed Turbulent Combustion[J]. Flow Turbulence & Combustion, 2018,100(1):39-73.
[25]	WELLER H G, TABOR G, JASAK H , et al. A Tensorial Approach to Computational Continuum Mechanics Using Object Orientated Techniques[J]. Computers in physics, 1998,12(6):620-631 . doi: 10.1063/1.168744
[26]	ABADI M, BARHAM P, CHEN J , et al. of Conference. Tensorflow: A system for large-scale machine learning [C]. 12th {USENIX} Symposium on Operating Systems Design and Implementation({OSDI} 16). 2016: 265-283.
[27]	NAIR V , HINTON G E of Conference. Rectified Linear Units Improve Restricted Boltzmann Machines [C]. Proceedings of the 27th International Conference on International Conference on Machine Learning. 2010: 807-8140.
[28]	IOFFE S, SZEGEDY C . Batch normalization: Accelerating deep network training by reducing internal covariate shift [J]. arXiv preprint arXiv:1502.03167, 2015 .
[29]	KINGMA D, BA J . Adam: A Method for Stochastic Optimization [J]. arXiv preprint arXiv:1412.6980, 2014 .

初始条件
速度	0m/s
压强	0bar
边界条件
入口速度	(10,0,0)m/s
入口压强	0
出口流速	0
出口压强	0bar
传递性质
层流粘性系数	10^-5 m²/s

数据集	网格大小	流速	层流粘性系数	数据集大小	备注
训练集A1	114700	10m/s	1.0e-5	1147000	时间窗口[0.01,0.1],均匀采样十次
训练集A2	28675	10m/s	1.0e-5	1147000	时间窗口[0.01,0.1],均匀采样四十次
训练集A3	458800	10m/s	1.0e-5	4588000	时间窗口[0.01,0.1],均匀采样十次
训练集A4	114700	5m/s	1.0e-5	1147000	时间窗口[0.01,0.1],均匀采样十次
训练集A5	114700	20m/s	1.0e-5	1147000	时间窗口[0.01,0.1],均匀采样十次
训练集A6	114700	10m/s	0.5e-5	1147000	时间窗口[0.01,0.1],均匀采样十次
训练集A7	114700	10m/s	2.0e-5	1147000	时间窗口[0.01,0.1],均匀采样十次
验证集B	114700	10m/s	1.0e-5	1147000	时间窗口[0.01,0.91],均匀采样十次

参数名称	详情
隐藏层数	17
隐藏单元数	512
学习率	0.02993
学习率衰减指数	0.41701
L2正则系数	0.06730

基于AutoML的湍流建模

Turbulence Modeling Based on AutoML

RichHTML

PDF

可视化

摘要/Abstract

引用本文

使用本文

图/表 14

参考文献 29

相关文章 15

编辑推荐

Metrics

本文评价

[1]	许淞源,刘峰. ESDRec：一种面向地球大数据平台的数据推荐模型[J]. 数据与计算发展前沿, 2023, 5(1): 55-64.
[2]	陈琼,杨咏,黄天林,冯媛. 小样本图像语义分割综述[J]. 数据与计算发展前沿, 2021, 3(6): 17-34.
[3]	蒲晓蓉,黄佳欣,刘军池,孙家瑜,罗纪翔,赵越,陈柯成,任亚洲. 面向临床需求的CT图像降噪综述[J]. 数据与计算发展前沿, 2021, 3(6): 35-49.
[4]	何涛,王桂芳,马廷灿. 基于词嵌入语义异常的跨学科研究内容发现方法[J]. 数据与计算发展前沿, 2021, 3(6): 50-59.
[5]	鹿旭东,宋伟凤,郭伟,崔立真,林岳,姜涛. 大数据驱动的创新方法论与创新服务平台[J]. 数据与计算发展前沿, 2021, 3(5): 141-155.
[6]	张怡宁,何洪波,王闰强. 热门数字音频预测技术综述[J]. 数据与计算发展前沿, 2021, 3(4): 81-92.
[7]	陈子健,李俊,岳兆娟,赵泽方. 基于自编码器与属性信息的混合推荐模型[J]. 数据与计算发展前沿, 2021, 3(3): 148-155.
[8]	肖建平,龙春,赵静,魏金侠,胡安磊,杜冠瑶. 基于深度学习的网络入侵检测研究综述[J]. 数据与计算发展前沿, 2021, 3(3): 59-74.
[9]	李序,连一峰,张海霞,黄克振. 网络安全知识图谱关键技术[J]. 数据与计算发展前沿, 2021, 3(3): 9-18.
[10]	郭佳龙,王宗国,王彦棡,赵旭山,宿彦京,刘志威. 基于计算机技术的材料研发方法概述[J]. 数据与计算发展前沿, 2021, 3(2): 120-132.
[11]	赵伟昱,张宏海,仲波. 基于深度学习的遥感影像地块分割方法[J]. 数据与计算发展前沿, 2021, 3(2): 133-141.
[12]	于建军,廖方宇,周小军,孙健英. 新一代ARP对管理应用创新的初探[J]. 数据与计算发展前沿, 2021, 3(2): 39-49.
[13]	沈飙,陈扬,杨琛,刘博文. 海洋科学中尺度涡的计算机视觉检测和分析方法[J]. 数据与计算发展前沿, 2020, 2(6): 30-41.
[14]	张圣林,林潇霏,孙永谦,张玉志,裴丹. 基于深度学习的无监督KPI异常检测[J]. 数据与计算发展前沿, 2020, 2(3): 87-100.
[15]	陈雷,袁媛. 基于深度迁移学习的农业病害图像识别[J]. 数据与计算发展前沿, 2020, 2(2): 111-119.