Physics-Driven Convolutional Autoencoder Approach for CFD Data Compressions Alberto Olmo

2025-05-02 0 0 3.04MB 6 页 10玖币

侵权投诉

Physics-Driven Convolutional Autoencoder Approach

for CFD Data Compressions

Alberto Olmo

National Renewable Energy Laboratory

Golden, CO 80401

aolmoher@asu.edu

Ahmed Zamzam

National Renewable Energy Laboratory

Golden, CO 80401

ahmed.zamzam@nrel.gov

Andrew Glaws

National Renewable Energy Laboratory

Golden, CO 80401

andrew.glaws@nrel.gov

Ryan King

National Renewable Energy Laboratory

Golden, CO 80401

ryan.king@nrel.gov

Abstract

With the growing size and complexity of turbulent ﬂow models, data compres-

sion approaches are of the utmost importance to analyze, visualize, or restart the

simulations. Recently, in-situ autoencoder-based compression approaches have

been proposed and shown to be effective at producing reduced representations

of turbulent ﬂow data. However, these approaches focus solely on training the

model using point-wise sample reconstruction losses that do not take advantage

of the physical properties of turbulent ﬂows. In this paper, we show that training

autoencoders with additional physics-informed regularizations, e.g., enforcing in-

compressibility and preserving enstrophy, improves the compression model in three

ways: (i) the compressed data better conform to known physics for homogeneous

isotropic turbulence without negatively impacting point-wise reconstruction quality,

(ii) inspection of the gradients of the trained model uncovers changes to the learned

compression mapping that can facilitate the use of explainability techniques, and

(iii) as a performance byproduct, training losses are shown to converge up to 12x

faster than the baseline model.

1 Introduction

With the advancement of high performance computing (HPC) there has been an increase in the interest

of leveraging such systems for computational ﬂuid dynamics (CFD) simulations. These have become

more readily available with larger than ever data sizes and performance ﬁdelity [Sprague et al., 2020,

Fischer et al., 2021, Musser et al., 2022]. Further, recent advances in computational processing

power achieved through heterogeneous architectures that couple traditional processors with graphics

processing units (GPUs) have led to an increase in the gap between processing power and memory

due to bandwidth constraints and memory-access times given by high latency input/output operations.

This can lead to memory-bottleneck scenarios where HPC machines are limited by the need to save,

analyze, visualize, or restore data from massive simulations. Given these factors and the increase

in available datasets, it becomes critically important to develop in-situ data compression techniques

to enable efﬁcient use of the data without sacriﬁcing accuracy. Additionally, a recent report from

the U.S. Department of Energy (DoE), shows the analysis and visualization of CFD simulations as a

central issue for next generation systems [Gerber et al., 2018].

Several lossy compression approaches have been proposed recently [Fukami et al., 2020, Glaws

et al., 2020, Carlberg et al., 2019, Dunton et al., 2020] utilizing singular value decompositions or

Preprint. Under review.

arXiv:2210.09262v1 [physics.flu-dyn] 17 Oct 2022

neural networks. In particular, convolutional autoencoders have proved to be able to obtain better

generalization results [Glaws et al., 2020]. However, these approaches only leverage sample quality

metrics while missing the physical properties inherent in the CFD and the beneﬁts of embedding

them at training time. Therefore, motivated by the big successes of convolutional neural networks in

the processing of CFD data [Guo et al., 2016, Tompson et al., 2017] and in-situ data compression

tasks [Liu et al., 2019] and the increased ability to generate large CFD data sets, we develop a

physics-driven convolutional autoencoder compression approach that builds on previous work [Glaws

et al., 2020].

In this work, we show that using an autoencoder model enhanced with two physical properties of

CFD leads to compression models that are more conformant with the physical characteristics of

the data as measured by known metrics for homogeneous isomorphic turbulent ﬂow, including the

divergence-free condition of incompressible ﬂow ﬁelds and the preservation of both enstrophy and

dissipation ratio [Constantin and Foias, 2020]. In addition, analyzing the model performance shows

a signiﬁcant reduction in training time as well as a reduced amount of training data necessary to

achieve same-quality reconstructions as compared to the baseline. Further, our preliminary analysis

of the learned models shows better explicability compared to models trained using only sample

quality metrics. All of this encourage the use of such networks with physical-losses and illustrate that

gradient-based explainability techniques can be leveraged in the future. 1

2 Approach

While lossless data compression approaches can be options for this problem [Fout and Ma, 2012,

Lindstrom and Isenburg, 2006], they pose memory and execution time burdens for the system.

On the other hand, lossy data compression aims to reduce the memory consumption by incurring

some manageable loss of information after decompression. Hence, in our problem of reducing the

dimensionality of CFD data and their memory expense, lossy compression methods are more suitable.

Thus, a general lossy data compression function with full data space

and compressed data space

can be deﬁned in two parts: the compression step

φ:X → Y

and reconstruction step

ψ:Y → X

where the degree of compression is measured by the compression ratio (CR). To this end, we design

a convolutional autoencoder with compression function

deﬁned by the encoder

, data

and

embedded data

such that

E(x) = z

and decompression function

by the decoder

such that

D(z) = ˆx

. Thus, the compressed data is obtained from

E(x)

which can be stored more easily than

the full data, and the full data reconstruction can be recovered with D(z).

Data

The dataset we use consists of simulated snapshots of ﬂuid velocities from incompressible

decaying isotropic ﬂows with component velocities on the

and

dimensions. Thus, the vector

ﬁelds are comprised of 3-dimensional meshes of

128 ×128 ×128

datapoints generated by the

spectralDNS package [Mortensen and Langtangen, 2016]. To increase robustness of the network, we

introduce turbulences in the simulations as measured by Taylor-scale Reynolds numbers between (65,

105) and gather a total of 1300 snapshots for our training dataset.

Physics-informed loss

The network follows a fully convolutional architecture. In general, the main

goal of a parameterized autoencoder is to minimize the reconstruction error with some pointwise

metric such as the squared 2-norm:

ΘE,ΘD= argmin

ΘE,ΘD

||x−D(E(x; ΘE); ΘD)||2

2.(1)

Recent works have shown the advantages of using the physical properties of the domain during

training. For instance, in Raissi et al. [2019], the authors train shallow neural networks with losses

that include domain-speciﬁc physics laws and show improved generalization performance of the

trained models. Similarly, Cai et al. [2022] show how physics-informed learning improves the

inference performance for CFD domains such as three-dimensional wake ﬂows or supersonic ﬂows.

Thus, in addition to MSE (1), we design the autoencoder loss with two physical laws that are

applicable to the domain in consideration, the divergence-free condition and the preservation of

enstrophy. For the former, due to the incompressibility of the ﬂow ﬁeld, the density of the CFD

remains constant expressed by

∇ · ~v = 0

and therefore is a property that can be enforced. Similarly,

the enstrophy of a ﬂuid measures the kinetic energy in the ﬂow that corresponds to dissipation

1The code of this work is attached as supplementary material and will be made publicly available.

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

Physics-DrivenConvolutionalAutoencoderApproachforCFDDataCompressionsAlbertoOlmoNationalRenewableEnergyLaboratoryGolden,CO80401aolmoher@asu.eduAhmedZamzamNationalRenewableEnergyLaboratoryGolden,CO80401ahmed.zamzam@nrel.govAndrewGlawsNationalRenewableEnergyLaboratoryGolden,CO80401andrew.glaws@nrel.gov...

展开>> 收起<<

Physics-Driven Convolutional Autoencoder Approach for CFD Data Compressions Alberto Olmo.pdf

共6页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Physics-Driven Convolutional Autoencoder Approach for CFD Data Compressions Alberto Olmo

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: