An advanced hybrid deep adversarial autoencoder for parameterized nonlinear fluid flow modelling

https://doi.org/10.1016/j.cma.2020.113375Get rights and content

Highlights

  • A hybrid DAA is developed for parameterized nonlinear fluid flow modelling.

  • The combination of GAN and VAE enables fast and accurate predictive simulations.

  • The use of VAE for model reduction can better represent the nonlinearity of flows.

  • Given varying inputs, DAA can capture the details of spatio-temporal flow features.

  • Using DAA the computational cost is reduced by several orders of magnitude.

Abstract

Considering the high computation cost required in conventional computation fluid dynamic simulations, machine learning methods have been introduced to flow dynamic simulations in years, aiming on reducing CPU time. In this work, we propose a hybrid deep adversarial autoencoder (VAE-GAN) to integrate generative adversarial network (GAN) and variational autoencoder (VAE) for predicting parameterized nonlinear fluid flows in spatial and temporal dimensions. High-dimensional inputs are compressed into the low-dimensional representations by nonlinear functions in a convolutional encoder. In this way, the predictive fluid flows reconstructed in a convolutional decoder contain the dynamic fluid flow physics of high nonlinearity and chaotic nature. In addition, the low-dimensional representations are applied to the adversarial network for model training and parameter optimization, which enables fast computation process. The capability of the hybrid VAE-GAN is illustrated by varying inputs on a flow past a cylinder test case as well as a second case of water column collapse. Numerical results show that this hybrid VAE-GAN has successfully captured the spatio-temporal flow features with CPU speed-up of three orders of magnitude. These promising results suggest that the hybrid VAE-GAN can play a critical role in efficiently and accurately predicting complex flows in future research efforts.

Introduction

Numerical simulations of nonlinear fluid flows are used to describe the complex evolution of many physical processes. Accurate simulations of nonlinear fluid flows are of great importance to many fields such as atmosphere, flood and ocean modelling, which systems often exhibit rich flow dynamics in both space and time [1], [2]. The numerical simulations have benefited from the availability of high-resolution spatio-temporal data due to recent advances in measurement techniques [3], which make the studies of more complex flows possible. However, the computational cost involved in solving complex problems is intensive which still precludes the development in these areas. In order to address the issue of the high computational cost, this paper proposes a hybrid deep adversarial autoencoder to solve fluid problems in an efficient manner.

Recent advances in machine learning technologies are increasingly of interest for efficiently simulating flows in dynamical systems [4], [5], [6]. Machine learning has demonstrated its potential capability in fluid flow applications [7], such as feature extraction [8], [9], [10], dimensionality reduction [5], [11] and superresolution [12]. For example, Kim et al. [13] adopted a convolutional neural network (CNN) to extract flow features and reconstruct fluid fields. Gonzalez and Balajewicz [11] developed a deep convolutional recurrent autoencoder for nonlinear model reduction, which identified the low-dimensional feature dynamics by the high-dimensional datasets and modelled flow dynamics on the underlying manifold space. Xie et al. [12] proposed a temporally coherent generative model to produce the super-resolution fluid flows. Despite the fact that these research efforts demonstrated how to successfully reconstruct flow dynamics, it is noted that they commonly do not take into account the temporal and spatial evolution of inputs or parameters, which is crucial for realistic dynamical systems [14].

Most recently, it has been shown that Generative Adversarial Network (GAN) is an efficient approach for reconstructing the spatial distribution of tracers and flow fields [15]. GAN introduced by Goodfellow et al. [16], has emerged as having a leading role in recreating the distributions of complex data [12], [17]. The key feature of GAN is the adversarial strategy consisting of two modules. GAN defines a learning generative network by transforming a latent variable into a state variable using nonlinear functions. Then GAN drives the learning process by discriminating the observed data from the generated data in a discriminator network. Because of the special adversarial architecture, GAN has demonstrated great capability in producing high-resolution samples, mostly in images, e.g., image synthesis [18], semantic image editing [19], style transfer [20] etc.

For efficient GAN training and accurate spatio-temporal fluid flow prediction, variational autoencoder (VAE) [21] has been introduced to GAN in this study. VAE has been widely used in various research areas, such as text generation [22], facial attribute prediction [23], image generation [24], [25], graph generation [26], music synthesis [27], and speech emotion classification [28], etc. The hybrid deep adversarial autoencoder (VAE-GAN) developed here takes advantage of both GAN and VAE. GAN allows for training on large datasets and is fast to yield visually high-resolution images, but the flexible architecture is easy to result in the model collapse problem and generate unreal results [29]. The VAE is attractive for achieving better log-likelihoods than GAN [30], [31], this suggests a hybrid VAE-GAN to better represent all the training data and discourage model-collapse problem in GAN [29].

The hybrid VAE-GAN developed here is a robust and efficient numerical tool for accurate prediction of parameterized nonlinear fluid flows. The advantages of the hybrid VAE-GAN include:

  • The proposed method exploits spatial features by use of convolutional neural networks. It will be advantageous over the traditional reduced order models (ROMs) based on proper orthogonal decomposition (POD) [32], [33], Galerkin regression [34], [35], cluster-based reduced order model (CROM) [36], [37] and dynamic mode decomposition (DMD) [38], which are used for model reduction of dynamical systems by linear combinations of the original snapshots. In convolutional neural networks, the high-dimensional datasets are compressed into the low-dimensional representations by nonlinearity functions in a convolutional encoder. In this way, the predictive fluid flows containing both high nonlinearity and chaotic nature can be represented by a convolutional decoder.

  • The low-dimensional representations, several orders of magnitude smaller than the dimensional size of the original datasets, are applied into the adversarial network for representation learning and parameter optimization, thus accelerating the computation process.

  • With the trained hybrid VAE-GAN, for given unseen inputs, the spatio-temporal features can be automatically extracted in the encoder. Consequently, accurate predictive nonlinear fluid fields can be further obtained in the decoder in an efficient manner.

This is the first time that the hybrid VAE-GAN is adopted to address parameterized nonlinear fluid flow problems. It is expected to make a breakthrough in predicting accurate nonlinear fluid flows with the high-speed computation.

The remainder of this paper is organized as follows. Section 2 presents the governing equation for nonlinear fluid flows in dynamic systems. Methodologies of VAE and GAN are briefly introduced in Sections 3 Variational autoencoder for model reduction, 4 Generative adversarial network, respectively. The hybrid VAE-GAN for parameterized nonlinear fluid fields is described in detail in Section 5. Section 6 demonstrates the performance of the hybrid VAE-GAN using a flow past a cylinder and water column collapse as test cases. Finally in Section 7, conclusions are presented.

Section snippets

Governing equations

In the computational fluid dynamic models, nonlinear fluid flows are usually simulated by solving nonlinear partial differential equations, such as Navier–Stokes equations. In general, a parameterized partial differential equation for a spatio-temporal fluid flow problem can be written as: M(ħ(x,μ,t),x,μ,t)=F(ħ(x,μ,t),x,μ,t),where M denotes a nonlinear partial differential operator, ħ(x,μ,t) is the state variable vector (for example, pressure, density, velocity, etc.), x represents the spatial

Variational autoencoder for model reduction

VAE was introduced by Kingma and Welling [21], which combines Bayesian inference with deep learning. The VAE is a generative model which aims to produce the desired local variable ħ from the underlying latent variable ζ (as shown in Fig. 1). Mathematically, let p(ζ) be a prior distribution of ζ, and the probability of the local state variable ħ be modelled by ħpθ(ħ|ζ),pθ(ħ)=pθ(ħ|ζ)pθ(ζ)dζ,where pθ(ħ|ζ) is the conditional distribution of the local state variable ħ given ζ, which is modelled by

Generative adversarial network

The GAN is generally implemented with a minimax game in a system of two players [39]. One player is responsible for generating the new samples ħ from a random dataset μ, while another player aims to discriminate the real samples ħd from the generated samples ħ. The former player is called the generator G and the latter is the discriminator D.

In the discriminator, D(ħd)=1 if the real samples ħd are accepted while D(G(μ))=0 if the generated samples ħ (ħ=G(μ)) are rejected. Unlike the

Hybrid deep adversarial autoencoder for nonlinear fluid flow modelling

In this paper, for nonlinear fluid flow modelling, a hybrid deep learning fluid model based on deep adversarial autoencoder (VAE-GAN) is proposed by combining a VAE and a GAN.

In spatio-temporal fluid flow simulations, the state variable vector ħ represents the fluid flow distribution as ħ=ħt11ħt12ħt1Nxħt21ħt22ħt2NxħtNt1ħtNt2ħtNtNx,and the state variable vector ħ in Eq. (11) can be rewritten as: ħtnt=(ħtnt1,,ħtntnx,,ħtntNx),ħnx=(ħt1nx,,ħtntnx,,ħtNtnx)T,where ħtnt is the state variable

Numerical examples

In this section, two examples, flow past a cylinder and water column collapse, are used to illustrate the capabilities of the hybrid VAE-GAN in resolving nonlinear fluid flow problem governed by the Navier–Stokes equations. The input–output datasets in both examples were obtained by running the unstructured mesh finite element fluid model (Fluidity) [44] (referred as the original high fidelity model). In Table 1, the module architecture of the encoder, decoder and discriminator used for two

Conclusions

In this work, a hybrid VAE-GAN method has been, for the first time, used to predict nonlinear fluid flows in varying parameterized space. For any given input parameters μ, the hybrid VAE-GAN is capable of predicting accurately dynamic nonlinear fluid flows and remains highly efficient in simulations since it combines both advantages of VAE and GAN.

The performance of the hybrid VAE-GAN has been demonstrated by a flow past a cylinder test case and a second case of water column collapse. To

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgements

The Authors acknowledge the support of: China Scholarship Council (No. 201806270238) and funding from the EPSRC (MAGIC) (EP/N010221/1) and INHALE (EP/T003189/1), and the Royal Society (IEC/ NS- FC/170563) in the UK, as well as the joint KAUST-Imperial research project (EACPR_P83206).

References (47)

  • LeeK. et al.

    Model reduction of dynamical systems on nonlinear manifolds using deep convolutional autoencoders

    (2018)
  • MurataT. et al.

    Nonlinear mode decomposition for fluid dynamics

    (2019)
  • BruntonS.L. et al.

    Machine learning for fluid mechanics

    Annu. Rev. Fluid Mech.

    (2020)
  • LingJ. et al.

    Reynolds averaged turbulence modelling using deep neural networks with embedded invariance

    J. Fluid Mech.

    (2016)
  • KutzJ.N.

    Deep learning in fluid dynamics

    J. Fluid Mech.

    (2017)
  • GonzalezF.J. et al.

    Deep convolutional recurrent autoencoders for learning low-dimensional feature dynamics of fluid systems

    (2018)
  • XieY. et al.

    Tempogan: A temporally coherent, volumetric gan for super-resolution fluid flow

    ACM Trans. Graph.

    (2018)
  • KimB. et al.

    Deep fluids: A generative network for parameterized fluid simulations

  • ReichsteinM. et al.

    Deep learning and process understanding for data-driven Earth system science

    Nature

    (2019)
  • FarimaniA.B. et al.

    Deep learning the physics of transport phenomena

    (2017)
  • GoodfellowI. et al.

    Generative adversarial nets

  • ReedS. et al.

    Generative adversarial text to image synthesis

    (2016)
  • LiC. et al.

    Precomputed real-time texture synthesis with Markovian generative adversarial networks

  • Cited by (0)

    View full text