RESEARCH ARTICLE | DECEMBER 22 2022

Physics-constrained deep learning of nonlinear normal

modes of spatiotemporal fluid flow dynamics 
Abdolvahhab Rostamijavanani ; Shanwu Li (黎善武) ; Yongchao Yang 

Physics of Fluids 34, 127121 (2022)

https://doi.org/10.1063/5.0124455

CrossMark

 
View Export
Online Citation

20 November 2023 07:32:15

 Physics of Fluids ARTICLE scitation.org/journal/phf

Physics-constrained deep learning of nonlinear

normal modes of spatiotemporal fluid
flow dynamics
Cite as: Phys. Fluids 34, 127121 (2022); doi: 10.1063/5.0124455
Submitted: 6 September 2022 . Accepted: 16 November 2022 .
Published Online: 22 December 2022

Abdolvahhab Rostamijavanani, Shanwu Li (黎善武),a) and Yongchao Yangb)

AFFILIATIONS
Department of Mechanical Engineering-Engineering Mechanics, Michigan Technological University, Houghton,
Michigan 49931, USA

a)
Electronic mail: [email protected]
b)
Author to whom correspondence should be addressed: [email protected]

ABSTRACT
In this study, we present a physics-constrained deep learning method to discover and visualize from data the invariant nonlinear normal

20 November 2023 07:32:15

modes (NNMs) which contain the spatiotemporal dynamics of the fluid flow potentially containing strong nonlinearity. Specifically, we
develop a NNM-physics-constrained convolutional autoencoder (NNM-CNN-AE) integrated with a multi-temporal-step dynamics predic-
tion block to learn the nonlinear modal transformation, the NNMs containing the spatiotemporal dynamics of the flow, and reduced-order
reconstruction and long-time future-state prediction of the flow fields, simultaneously. In test cases, we apply the developed method to ana-
lyze different flow regimes past a cylinder, including laminar flows with low Reynolds number in transient and steady states (RD ¼ 100) and
high Reynolds number flow (RD ¼ 1000), respectively. The results indicate that the identified NNMs are able to reveal the nonlinear spatio-
temporal dynamics of these flows, and the NNMs-based reduced-order modeling consistently achieves better accuracy with orders of magni-
tudes smaller errors in construction and prediction of the nonlinear velocity and vorticity fields, compared to the linear proper orthogonal
decomposition (POD) method and the Koopman-constrained-CNN-AE using the same number or dimension of modes. We perform an
analysis of the modal energy distribution of NNMs and find that compared to POD modes, the few fundamental NNMs capture a very high
level of total energy of the flow, which is advantageous for reduced-order modeling and representation of the complex flows. Finally, we dis-
cuss the potentials and limitations of the presented method.
Published under an exclusive license by AIP Publishing. https://doi.org/10.1063/5.0124455

I. INTRODUCTION POD, perhaps the most widely studied approach for mode
The existence of large-scale coherent structures in fluid flows has decomposition of fluid dynamics,5–9 performs a linear transformation
been observed despite fluids being highly complex, high-dimensional of the physical coordinates based on the energy norm, and the result-
dynamical systems.1 Therefore, discovering proper coordinates or ing modes are ranked according to the energy content. However, it
modes enables understanding and characterization of complex dynamic becomes ineffective when considerable nonlinearity is present such as
features of fluid flows exhibiting highly nonlinear phenomena, which is a flow with a high Reynolds number (HRN).10 Also, POD only
essential for many applications in system identification,2 modal analy- involves a spatial transformation, which captures spatial patterns in
sis,3 and system control.4 Specifically, these characteristic modes are able the original flow fields without explicitly considering the temporal
to (i) reveal the hidden structures of flow fields for an improved under- dynamics. As an example, if the temporal order of snapshots used for
standing of the flow physics and (ii) represent the flow in a reduced- deriving POD modes is not preserved, it will not affect derived POD
dimensional modal space, enabling the construction of a reduced-order modes as POD operation only takes snapshots as samples and does
model. To perform mode decomposition and identification of fluid not take into account their temporal order. Additionally, since this
dynamics, many methods have been developed, most representative method uses the Singular Value Decomposition (SVD) method, it
being proper orthogonal decomposition (POD) and Koopman operator cannot efficiently handle invariances in data, such as transitional or
or its approximation, e.g., dynamic mode decomposition (DMD). rotational invariances of low-rank objects embedded in fluid flow.1

Phys. Fluids 34, 127121 (2022); doi: 10.1063/5.0124455 34, 127121-1

Published under an exclusive license by AIP Publishing
 Physics of Fluids ARTICLE scitation.org/journal/phf

As a result, the POD modes potentially lose temporal dynamic features In this study, we explore the invariance properties of NNMs and
of the flow. POD–Galerkin is a reduced order modeling (ROM) for non- present a holistic physics-constrained deep learning framework to (1)
linear dynamical systems, which approximates the state vector with the discover and visualize the NNMs from data to reveal the spatiotempo-
finite sum of POD modes, but it is subject to some limitations including ral dynamics of the fluid flow potentially containing strong nonlinear-
instability and computational efficiency.1,11 Recently, by leveraging the ity and (2) leverage the identified NNMs subspace for predictive
unprecedented nonlinear modeling capability of machine/deep learning modeling of the nonlinear flow fields. Specifically, we devise a NNM-
techniques for flow fields5,10,12–22 presented a nonlinear mode decompo- physics-constrained convolutional autoencoder (CNN-AE) integrated
sition of a flow around a circular cylinder at RD ¼ 100 by using a convo- with a multi-temporal-step dynamics prediction block to learn the
lutional autoencoder (CNN-AE), where latent fields contain series of nonlinear modal transformation functions, the NNMs containing the
POD modes to ensure mode decomposition. However, as with POD, spatiotemporal dynamics of the flow, and long-time future-state pre-
the temporal dynamics of the flow was not explicitly considered therein. diction of the flow fields, simultaneously. We investigate the perfor-
In this work, we aim to retain both spatial patterns and the associated mance of the presented method, with comparisons with POD and
temporal dynamics in the discovered modes of the flow. Koopman-based method, in studying the flows around circular cylin-
Koopman operator23–25 is able to discover both modes and the der for different regimes including low and high Reynolds numbers
associated modal dynamics of flow fields.3,25–34 With a proper mea- streams.
surement (coordinate transformation) function which can be nonlin- Distinctions are observed between the presented method and rel-
ear, the Koopman operator enables linearization of any nonlinear evant work. One approach to discovering dynamical systems from
systems in a new infinite-dimensional space, where the nonlinear data is to use the sparse identification of the nonlinear dynamics
dynamics can be represented by a linear superposition of infinite- (SINDy) algorithm. The SINDy algorithm is capable of identifying a
dimensional modes. The challenge of this method, however, is obtain- governing equation from time-discretized data as well as identifying
ing the measurement function that transforms nonlinear dynamics to dominant terms from a large set of possibilities.2,59 SINDy-based
a new coordinate space where the underlying dynamics can be approx- methods (in particular, Refs. 60 and 61) are closely related to our
imately represented linearly. Also, the required dimension of the new NNMs-embedded deep learning (DL) framework as both learn/
coordinate space is difficult to determine in advance because it should identify a mathematical model from data to represent the observed
be infinite in theory. DMD was proposed35 as a numerical approxima- dynamics. However, two significant differences are observed. First,
tion to the Koopman spectral analysis and successfully applied to these two frameworks have different levels of requirements on prior

20 November 2023 07:32:15

mode decomposition of fluid flows.1 Recent advances have shown that knowledge and, thus, yield different complexity of resulting models.
the new coordinates (also known as Koopman eigenfunctions) that Specifically, SINDy requires sufficient prior knowledge of the observed
enable the intrinsic linear representation of nonlinear dynamical sys- system to determine a library of all possible functions, a few of which
tem can be learned from the flow field data using the strong nonlinear will be identified to be active terms and compose a parsimonious
modeling power of deep learning.36–42 Notably,36 introduced a low- model. Our NNMs-embedded DL framework does not require any
dimensional autoencoder as deep neural networks (DNNs) to capture prior knowledge of the possible functions for the observed system
Koopman eigenfunctions for dynamical systems with continuous thanks to the universal approximation capabilities of deep neural net-
spectra. These Koopman operator-based methods provide the unique works (DNNs), but, the complexity of the resulting DNN model is
advantage of representing nonlinear systems using the transformed generally higher than that of a SINDy model. Second, regarding the
linear dynamics, which significantly facilitates system identification incorporation of physical constraints, the Ref. 61 proposed an effective
and control. Nevertheless, they may lack accuracy when applied to way to incorporate physical constraints. However, this constraint is
highly nonlinear systems; in particular, they require a large number of limited to the pre-defined library of possible functions, e.g., an inter-
modes to represent the dynamics of flow containing significant non- dependence between two possible functions. Unlike the SINDy model,
linearity, and such a required dimension for the operator is difficult to a DNN is flexible to incorporate physical constraints because of the
specify as a priori. In this work, we aim to identify low-dimensional nature of the network structure and the way to update the network
modes to represent the potential highly nonlinear fluid dynamics with- parameters in training (e.g., backpropagation). Additionally, the com-
out the requirement for linearization of dynamics. bination of POD and SINDy60 provides an effective method for data-
Nonlinear normal modes (NNMs), introduced by Ref. 43 and driven reduced-order modeling, but still requiring sufficient prior
studied by many researchers,44–49 are natural extensions of linear normal knowledge on the observed fluid system to pre-define the required
modes (LNMs) for understanding and characterizing nonlinear dynami- library of all possible functions.
cal systems.44,50–52 Unlike above-mentioned methods with a target of II. METHODS
linearization of nonlinear dynamics, NNMs, extending the concept of
LNMs, generalize the intrinsic invariance properties of a nonlinear sys- We specifically consider a two-dimensional flow field over a cyl-
tem while retaining nonlinear dynamics in the modal coordinates. The inder, which is a typical example used in a number of existing
invariant manifolds, introduced in the general NNMs framework by works,5,20 to study the performance of our method, while the concep-
Shaw and Pierre,45,46 enforce that any motion initiated on such a mani- tual idea of the presented method is also applicable to three-
dimensional flow in theory as discussed in Sec. V.
fold remains there during the time, naturally extending the invariance
property of LNMs to nonlinear systems.53–56 Importantly, the invariance
property of NNMs allows spanning the generally lowest-dimensional A. Proper orthogonal decomposition (POD)
subspace to capture nonlinear dynamics57,58 and is leveraged in this POD performs a linear transformation of the fluid flows.62 The
work to discover nonlinear dynamic modes of fluid flow fields. inputs are snapshots [zðf; tÞ] of the flow such as temperature,

Phys. Fluids 34, 127121 (2022); doi: 10.1063/5.0124455 34, 127121-2

Published under an exclusive license by AIP Publishing
 Physics of Fluids ARTICLE scitation.org/journal/phf

pressure, and velocity, and the outputs are orthogonal modes repre- evolution and the system state remains on that manifold all the time).
senting the dominant spatial features of the flow: Derivation of the function G in our case can be treated as an invariant
X manifold-based reduced-order modeling approach since we identify
zðf; tÞ 
z ðfÞ ¼ aj ðtÞ/j ðfÞ; (1) this function in the latent/modal space, where a few coordinates
j contain/capture the most pertinent information about the dynamical
where
z ðfÞ is the temporal mean of the flow field and /j ðf; tÞ and aj system. By its invariance properties45,46 and nonlinear mapping,
are modes and expansion coefficients, respectively.3 POD has shown NNMs are able to (1) capture and reveal the highly nonlinear spatio-
good performance for fluid systems that are not highly nonlinear and temporal dynamics in the fluid flow and (2) span the generally lowest-
allows reconstruction of the flow fields with a linear superposition of dimensional subspace to capture nonlinear dynamics.57,58
only a few dominant modes. However, the POD modes do not explic- Finding the analytical functions for both Koopman operator and
NNMs method is extremely challenging, if not impossible, especially
itly retain the temporal dynamics of the fluid flow because POD
when the governing equation is (partially) unknown. In Sec. III, we
mainly involves a spatial transformation with capturing spatial pat-
introduce a physics-constrained deep learning framework to discover/
terns in the original flow fields. Its performance will be investigated as
identify NNMs and the associated intrinsic modal coordinates that
a comparison with the proposed method in Sec. IV.
capture the underlying nonlinear dynamics of the fluid flow from data
only. The framework is flexible to allow integration of the physics con-
B. Koopman operator and nonlinear normal modes straints of Koopman operator, and the corresponding method for
(NNMs) learning Koopman modes is developed in the same manner and com-
The observed dynamics of flow fields can be expressed in the dis- pared with NNMs in Sec. IV.
crete space and time as follows:
III. PHYSICS-CONSTRAINED DEEP LEARNING
z kþ1
¼ Fðz Þ;k
(2) FRAMEWORK TO DISCOVER NONLINEAR NORMAL
MODES (NNMs) OF FLUID FLOWS
where z 2 R denotes the vector field measurements, superscript k
n
We devise NNMs-physics-constrained CNN-AE which (1)
stands for the time step, and F represents the dynamics of the system approximates nonlinear normal modal transformation through deep
which maps current field to next field forward in time. In this work, neural networks with loss functions about NNMs to enforce them to fol-

20 November 2023 07:32:15

we consider both velocity and vorticity as the studied variable z. The low the constraints related to NNMs and (2) embeds the multi-step pre-
evolution function F corresponds to the Navier–Stokes (NS) equation, diction in a temporal dynamics block, thereby discovering the NNMs
which is usually (partially) unknown for the observed fluid fields due and the associated temporal dynamics of the flow. The overall NNM-
to the, e.g., insufficient knowledge of the boundary condition. CNN-AE framework is described in Sec. III A, followed by the
In theory, the Koopman operator linearly advances current mea- NNM-physics-constrained CNN-AE in Sec. III B. The same architecture
surements of a state to approximate the next in time with a measure- integrating the loss functions about Koopman is presented in Sec. III C.
ment function as u ¼ /koop ðzÞ, where /koop denotes the Koopman
measurement function. Rigorously, the measurements u must be infi- A. Overall CNN-AE framework
nite dimensional to guarantee the linear dynamics in the measurement
coordinates: The architecture of our CNN-AE is presented in Fig. 2. Original
flow field coordinates are transformed to intrinsic coordinates (NNMs
ukþ1 ¼ Kuk ; (3) or Koopman modal coordinates) with encoder block, u : Rn ! Rs ,
where n and s are the dimension of original coordinates (the number of
where K is the Koopman operator and u is the vector of Koopman grid points of flow field) and latent coordinates, respectively. We use
measurement coordinates or eigenfunctions.36 This linear representa- four modal coordinates (s ¼ 4) of NNMs to compare with those
tion of nonlinear dynamics allows using well-developed linear theories obtained with (1) the same number of modal coordinates for POD and
for controls and performing both reconstruction by inverse of /koop (2) for Koopman operator, different number of latent coordinates are
and dynamics prediction by K. However, discovering the measurement implemented (s  4). Thus, the high-dimensional flow field is mapped
function /koop with the infinite-dimensional coordinates u is intracta- to low-dimensional space by passing through the encoder layers, while
ble although some existing approaches have been proposed to approxi- the most important features of flow are captured in low-dimensional
mate the Koopman operator with a finite dimension, e.g., DMD. space. After passing through encoder block and having intrinsic coordi-
In contrast to Koopman operator, with the nonlinear modal nates, the network predicts the evolution of dynamics of fluid system
transformation u ¼ /NNM ðzÞ, the dynamics of the derived modal by the dynamics block, and finally, the evolved modal coordinates are
coordinates in the nonlinear normal modes (NNMs) framework transferred back to original flow coordinates by the decoder block.
remain nonlinear as follows:
B. NNM-physics-constrained CNN-AE
ukþ1 ¼ Gðuk Þ; (4)
The proposed NNM-physics-constrained CNN-AE integrates
where G is the nonlinear temporal evolution function. In the phase the constraints related to NNMs (Sec. II B) into the deep learning. The
space of a system model, NNMs can be defined as invariant manifolds. overall loss function is as following:
If it is possible to express such manifolds63 explicitly as some function
_ lNNM ¼ arec lrec þ acorr lcorr þ aevolNNM levolNNM þ aprd lprd ;
mðuÞ ¼ c with c as a constant, then, mðuÞ ¼ 0 (see Fig. 1, where
the portion of general manifold does not change over temporal (5)

Phys. Fluids 34, 127121 (2022); doi: 10.1063/5.0124455 34, 127121-3

Published under an exclusive license by AIP Publishing
 Physics of Fluids ARTICLE scitation.org/journal/phf

FIG. 1. Invariance properties of NNMs:

the vorticity of three physical points fol-
lows a projection for each mode. ti repre-
sents the vorticity value for the selected
physical points. Projection for each mode
is a part of the general nonlinear manifold
of the flow identified by NNMs neural
networks.

where lNNM is the overall loss function for NNM-CNN-AE and latent coordinates, the decoder should transform them back to
lrec ; lcorr ; levol , and lprd are reconstruction in original coordi- the original coordinates by minimizing jjzkþ1  #1 ðGð#ðzk ÞÞÞjj,
nates, independence between modal coordinates, evolution (dynamics) or generally for multiple-time-step prediction, lprd ¼ jjzkþm
in latent space, and prediction in original coordinates, respectively, #1 ðGðGðG…ð#ðzk ÞÞÞÞÞjj.
which have been expressed in detail as below:
During the implementation, different weights (hyper-parameters in
1. Encoder block to identify forward and inverse nonlinear coordi- Table I) are assigned to these loss terms comprising the overall loss
nate transformation functions (lrec ). The first loss is reconstruc- function to be minimized.
tion to enforce our NNM-CNN-AE to reconstruct the original
coordinates by the transformed latent intrinsic coordinates; this
C. Koopman-physics-constrained CNN-AE
is implemented by minimizing lrec ¼ jjzk  #1 ð#ðzk ÞÞjj. z is
the input coordinates (velocity or vorticity) and the superscript k The performance of the proposed NNM-CNN-AE is compared

20 November 2023 07:32:15

indicates the time. to a similar Koopman-CNN-AE. Koopman operator requires infinite
2. Independence between NNMs coordinates (lcorr ). This loss dimensions of modes (latent coordinates) to transform nonlinear
function is the physics-integrated loss function because it enfor- dynamics to linear dynamics. Therefore, the dimension of the latent
ces the generalized invariance of NNMs. It makes the NNMs’ coordinates are not constant in the neural network designed for this
modal coordinates as independent as possible by minimizing the operator. The other main feature of Koopman operator is the embed-
modal-uncorrelated loss. In practice, the independence require- ded linear dynamics. To ensure the dynamics block of the network
ment generally yields that any two generated modal coordinates makes linear transformation, we use linear activation function for neu-
either have different frequency components or have the same rons in hidden layers of this block. The overall loss function used for
frequency but with 90 phase difference (conjugate modes). Koopman-CNN-AE is similar to NNM-CNN-AE [see Eq. (5)] with
With this way, the so-called nonlinear invariant manifolds (con- different evolution loss function:
figurations) which is a key feature of NNMs method can be
achieved (see Figs. 1 and 3): lKPM ¼ arec lrec þ acorr lcorr þ aevolKPM levolKPM þ aprd lprd ;
(7)
lcorr ¼ jjCorrðui ; uj Þ  Iss jj; i ¼ 1; …; s; j ¼ 1; …; s; (6)
where lKPM is the overall loss function for Koopman-CNN-AE and
where I and Corr are identity matrix and correlation matrix, lrec ; lcorr , and lprd correspond to the same loss functions stated in
respectively and s is the number of degrees of freedom of the sys- Eq. (5) and levolKPM is expressed as follow:
tem. This correlation loss task is modal coordinate decomposi- levolKPM : Evolution in latent subspace (linear dynamics). This
tion the latent space. is similar to that for NNMs except that in this dynamics block we
3. Dynamics block to identify temporal evolution (dynamic) function enforce the network to make a linear transformation for the evolution
(levolNNM ): evolution in latent subspace (nonlinear dynamic). In of latent coordinates by minimizing: jj#ðzkþ1 Þ  K#ðzk Þjj. K is the
the dynamics block, the network uses the initial time response of Koopman operator which can be modeled as linear embedded dynam-
each training example to predict the evolution of system recursively. ics with linear activation functions. For multiple-time step prediction,
This is enforced by minimizing jj#ðzkþ1 Þ  Gð#ðzk ÞÞjj. G repre- it becomes levolKPM ¼ jj#ðzkþm Þ  K m #ðzk Þjj.
sents the dynamics block which is modeled as a nonlinear embed-
ded dynamics with nonlinear activation functions (Relu function is
used). For multiple-time-step prediction, the loss becomes D. Network architecture and training
levolNNM ¼ jj#ðzkþm Þ  GðGðG…ð#ðzk ÞÞÞÞjj, where sate spaces In the proposed DNN, there are three models: (1) encoder, (2)
at time k are enforced to pass m (a hyper-parameter) times through decoder, and (3) dynamics block. Encoder and decoder are CNN mod-
the nonlinear dynamics block (G). els, while the dynamics blocks (G and K) are multilayer perception
4. Future-state prediction by orderly incorporating the autoencoder models. The information, including their number of layers, filters, and
and the dynamics block (lprd ). After evolution prediction in kernal size, is presented in Fig. 2. The dynamics block is composed of

Phys. Fluids 34, 127121 (2022); doi: 10.1063/5.0124455 34, 127121-4

Published under an exclusive license by AIP Publishing
 Physics of Fluids ARTICLE scitation.org/journal/phf

20 November 2023 07:32:15

FIG. 2. Architecture of our physics-constrained CNN-AE. (a) The overall framework consists of a CNN-AE that encodes original coordinates of flow Zt into intrinsic coordinates
ut using / ¼ #ðzÞ and then decodes them back to original coordinates by z ¼ #1 ð/Þ. There are additional physics-based constraints that can be applied to the intrinsic
coordinates / to enable them to be translated to desired modal coordinates. (b) In addition, we implement a dynamics block (G/K), which advances intrinsic coordinates for-
ward in time and enforces the equivalence between encoding the next original coordinates and advancing current intrinsic coordinates forward. This allows us to ensure the
dynamics of the system remain in the identified intrinsic coordinates. (c) By combining encoder, dynamics block, and decoder orderly, we are able to determine intrinsic coordi-
nates for enabling future flow fields prediction. It should be noted that decoder is not exactly the inverse function of encoder, but, we try to approximate it as much as possible
through reconstruction loss function. (d) Encoder block (blue) contains convolutional layers with kernel size 3  3 and maxpooling layers with a pool size 2  2. (e) Decoder
block (green) contains convolutional layers with a kernel size 3  3 and upsampling layers with size 2  2.

Phys. Fluids 34, 127121 (2022); doi: 10.1063/5.0124455 34, 127121-5

Published under an exclusive license by AIP Publishing
 Physics of Fluids ARTICLE scitation.org/journal/phf

FIG. 3. Phase portraits of the modal coor-

dinates of the POD modes and NNMs,
respectively, for streamwise velocity of (a)
low Reynolds number laminar regime and
(b) high Reynolds number stream. NNMs
curves are more distorted (nonlinear)
compared to POD modes which are
circular.

four perception layers, each with 256 neurons. Each model performs mapping), and these intrinsic coordinates represent the Koopman
the following tasks. modal coordinates, which are independent of each other (using
lcorr ).
1. Encoder It should be noted that all of the three models are trained simul-
taneously, meaning that all weights are shared and modified during
The objective of this model is to convert the original coordi- the training phase. The activation functions used for the layers in
nates into modal coordinates (forward modal transformation). The NNM-CNN-AE are all nonlinear as we are seeking nonlinear modal
output of the encoder is the latent modal coordinates, and these transformations as well as nonlinear mapping for the flow dynamics
coordinates will then be passed through the dynamics block/ potentially containing strong nonlinearity. Among different nonlinear
decoder. As presented in Figs. 2(a)–2(c), the convolutional layer activation functions, we use the “Tanh” activation function for convo-
extracts the spatiotemporal features of the flow field, while the pool- lutional layers in encoder and decoder blocks for both NNMs and

20 November 2023 07:32:15

ing layer decreases the degree of freedom of the flow in the encoder Koopman operators. In the dynamics block of Koopman-CNN-AE,
block. Thus, we have more compressed fields containing the physics linear activation functions are used to ensure that the evolution of the
features about the flow when we pass through the encoder by using latent spaces is pure linear pertaining Koopman’s theory. The nonlin-
such layers. Therefore, the loss functions associated with this model ear activation functions (Relu) are used for the dynamics blocks of
are lrec ; lprd , and lcorr . NNM-CNN-AE because it is assumed that dynamics and evolution
are both nonlinear.
2. Decoder All laminar flow fields (velocity and vorticity fields) are trained
on a NVIDIA RTX4000 GPU. Four thousand iterations were used for
The purpose of this model is to transfer the modal coordinates in training, which took nearly 28 h. Both NNM-CNN-AE and Koopman-
the latent space back to their original coordinates. Once the intrinsic CNN-AE use Adam’s optimizer with a learning rate, a ¼ 0:0001.
physics of flow has been extracted from the latent layers, the original Initial weights are randomly selected, then hyper-parameter tuning are
flow field is reconstructed by using upsampling and convolutional performed, and the results are based on the hyper-parameters that had
layers within the decoder block. We use fully connected layers, the fewest validation errors (Table I). We use the grid search approach
reshape, and flatten layers in order to preserve the size of the recon- within a certain range to determine the hyper-parameters including
structed flow field. To train this model, lrec and lprd should be the loss weights and architecture of neural networks. Regarding the
considered. loss weights, the correlation loss and the evolution loss are critical
because they are the key to identifying NNMs and associated modal
3. Dynamics block dynamics. These two losses are also relatively difficult to converge in
This model aims to represent the dynamics of systems by map- contrast to the reconstruction loss and prediction loss. Therefore, we
ping intrinsic modal coordinates to some specified time steps in
advance. This can be achieved by training two loss functions in the
overall proposed frameworks: levolNNM (levolKPM for Koopman TABLE I. Hyper-parameters used in the networks for velocity and vorticity fields.
DNN) and lprd . This block is given the initial state and is expected to
predict the future states recursively (multi-time steps prediction). For arec aevolNNM aprd acorr
NNMs, as the dynamics block (G) uses nonlinear activation functions,
this mapping (dynamics) is nonlinear. Further, since each intrinsic LRN streamwise velocity 1 5 1 0.001
coordinate is independent (using lcorr ), we can ensure that the intrin- HRN streamwise velocity 1 10 1 0.001
sic coordinates represent the NNMs coordinates and the invarience Transient streamwise velocity 1 10 1 0.005
properties of NNMs are preserved (see Figs. 1 and 3). The same Transverse velocity 1 5 1 0.015
approach applies to the Koopman dynamics block (K), where the acti- Vorticity 1 5 1 0.005
vation functions are linear, as we seek linear dynamics (linear

Phys. Fluids 34, 127121 (2022); doi: 10.1063/5.0124455 34, 127121-6

Published under an exclusive license by AIP Publishing
 Physics of Fluids ARTICLE scitation.org/journal/phf

suggest fixing the reconstruction loss weight and prediction loss weight (nonlinear), while the trajectory plot of POD modes is circular [Fig. 3].
and tune the correlation loss weight and evolution loss weight using a These results suggest that the identified NNMs are able to reveal the
from-coarse-to-fine grid search. As for the architecture, a more nonlinear physics behind the flow in the laminar regime better com-
strongly nonlinear dynamic regime generally needs a larger architec- pared to the linear method POD. These nonlinear spatiotemporal fea-
ture of neural network (with more layers and units in each layer) with tures captured by NNMs are also beneficial to the reduced-order
more capacity to represent and identify the stronger nonlinearity. For reconstruction and prediction of the flow field potentially containing
flow with high Reynolds number and transient flow regimes, the nonlinearity, as detailed in the following.
model training is conducted on a supercomputer of 24 computation We further evaluate the efficiency and accuracy of NNM-CNN-
nodes (ThetaGPU) at Argonne Leadership Computing Facility. Each AE compared with the linear method POD and Koopman operator in
node has two AMD Rome 64-core CPUs and eight NVIDIA A100 terms of the reconstruction and prediction of flow fields (i.e., velocity
GPUs with 320 GB GPU Memory. and vorticity) using their identified modes, respectively. To compare
POD and NNMs methods, the same number or dimension of the
IV. RESULTS AND DISCUSSION identified modal coordinates is used for (1) reconstruction of the flow
We study the performance of the proposed method on the flow field and (2) prediction for a time window which covers over six cycles.
past a circular cylinder for the laminar steady flow with low Reynolds The presented results are from using four identified modal coordinates
number (LRN) (RD ¼ 100), transient flow with Reynolds number for each approach (NNMs/POD), respectively.
ranging from 100 to 110, and a high Reynolds number (RD ¼ 1000), The reconstruction using a linear superposition [Eq. (1)] of four
respectively. identified POD modes [Fig. 4(b)] and a nonlinear combination of four
identified NNMs [Fig. 4(c)] are presented in Figs. 5(a) and 5(b),
A. Laminar regime: Low Reynolds number (LRN) respectively. The spatial least squares errors (L2 errors) between recon-
We first consider a flow past the cylinder with a Reynolds num- structed and real flow fields are depicted in the middle column five,
ber RD ¼ 100 in steady state. It produces vortex shedding in the wake whereas the reconstructed time history of a physical point in the flow
after the cylinder known as Karman vortex street. The governing equa- field compared with real data is in the right column. The mean
tions are the incompressible continuity and Navier–Stokes (NS) squared error (MSE) of the spatial flow is in order 105 , whereas this
equations: error has an order of 104 in the POD method (Table II). Also, time-
history reconstruction of the selected point with NNMs has better

20 November 2023 07:32:15

r  U ¼ 0; accuracy compared with POD as reported in Table II. It is seen that
@U 1 2 (8) the reconstructed velocity field with NNMs has significantly smaller
¼ r  ðUUÞ  rp þ r U;
@t RD error in comparison with POD with the same number of modal
coordinates.
where U and p are the velocity vector and pressure, respectively. Unlike POD modes that represent symmetric spatial patterns in
Streamwise and transverse velocity are assigned as u and v correspond- the flow, NNMs discover a nonsymmetric twisted (nonlinear) coher-
ingly. No-slip boundary condition is applied, and the time interval ent spatial structures of flow10 which evolve with mono frequency.
between each snapshot of flow field is Dt ¼ 0:4 s. The channel is Furthermore, prediction over 40 s by NNM-CNN-AE using four
divided into 67  150 evenly spaced grid points, and 15 000 snapshots NNMs is conducted. There are 100 time steps for prediction in this
are the samples used in total for training, validation, and testing range. NNM-CNN-AE receives the initial flow field (velocity in this
phases. We compare the NNM-CNN-AE results with POD in both case) and predicts the future 99 time steps recursively. Since each step
spatial and time domains and also present the decomposed flow fields is the prediction based on the estimated previous time step, it requires
acquired by modal coordinates in each method for velocity fields a precise prediction at each time step to avoid error accumulation over
(streamwise and transverse directions) and vorticity cyclic field, time. As indicated in the accurately predicted velocity field in Fig. 5(c)
respectively, as described in the following. and the corresponding small errors both in spatial domain and time
domain for the selected point in the flow (Table II), the proposed
1. Streamwise velocity NNM-CNN-AE shows a considerable ability for the prediction of a
wide range of time spans recursively.
We first present the mode decomposition of the streamwise veloc-
ity field by the proposed NNM-CNN-AE compared with POD.
2. Transverse velocity
Nonlinear normal modal decomposition enabled by the NNM-CNN-
AE is observed [Fig. 4(c)] in revealing the nonlinear spatial and tempo- Transverse velocity of the flow over a cylinder is also studied.
ral features of the flow. In more details, the original velocity field with Mode decomposition of the transverse velocity field with POD and
two dominant modes of different spatial patterns and temporal fre- NNM-CNN-AE is presented in Figs. 6(b) and 6(c), respectively. While
quencies [Fig. 4(a)] are both decomposed into two pairs of conjugated the original velocity field contains multiple modes [Fig. 6(a)], only the
POD modes [Fig. 4(b)] and NNMs [Fig. 4(c) for single-mode- first mode is dominant. Similar to the streamwise flow case, the identi-
reconstructions of NNMs], each with distinguished temporal fre- fied NNMs capture more latent nonlinear spatial patterns and temporal
quencies (0.16 and 0.31 Hz, respectively). On the other hand, as is seen dynamics in the flow field than the identified POD modes. Such cap-
comparing Figs. 4(b) and 4(c), the identified NNMs have distorted tured spatiotemporal features in the NNMs also benefit both the time
spatial patterns, capturing the nonlinear nature of the flow. This is also and spatial domain flow reconstruction with NNM-CNN-AE using
be seen in the phase portrait of the trajectory of the first and second four NNMs [Fig. 7(b)], yielding better accuracy (2 orders of magnitude
modes of POD and NNMs, where the NNMs curve is more distorted smaller errors) compared with the reconstruction using four POD

Phys. Fluids 34, 127121 (2022); doi: 10.1063/5.0124455 34, 127121-7

Published under an exclusive license by AIP Publishing
 Physics of Fluids ARTICLE scitation.org/journal/phf

20 November 2023 07:32:15

FIG. 4. Mode decomposition of the streamwise velocity field with POD and NNM-CNN-AE. (a) Original streamwise velocity spatial- and time-domain flow field. Mid- and right
columns: instantaneous frequency and time history of a selected spatial point in the flow, respectively. (b) Identified POD modes (shown the first four) with instantaneous fre-
quency and time history. (c) Identified NNMs (shown the first four) with instantaneous frequency and time history by the presented NNM-CNN-AE.

Phys. Fluids 34, 127121 (2022); doi: 10.1063/5.0124455 34, 127121-8

Published under an exclusive license by AIP Publishing
 Physics of Fluids ARTICLE scitation.org/journal/phf

20 November 2023 07:32:15

FIG. 5. Streamwise velocity field reconstruction and prediction. (a) POD reconstructed field using the first four POD modes; mid-column: the corresponding L2 error in spatial
domain for an instantaneous snapshot; right column: time-domain reconstruction for a selected point in the flow. (b) NNM-CNN-AE reconstruction using the first four NNMs in
spatial and time domain. (c) NNM-CNN-AE prediction in spatial and time domain. The intensity colorbar is at the bottom of (a).

modes [Fig. 7(a), as also reported in Table III]. As seen in Fig. 7(c) and fv D
Su ¼ ; (9)
Table III, the proposed NNM-CNN-AE also shows accurate recursive U1
prediction of both the spatial- and time-transverse velocity field.
where fv, U1 , and D are the vorticity main frequency, free stream
velocity, and diameter of cylinder, respectively. However, the vorticity
3. Vorticity
contains other higher-frequency components (modes) which should
We apply the proposed NNM-CNN-AE to study the vortex be considered for modeling this phenomenon accurately. In this stud-
shedding which takes place after the cylinder wake. The dimensionless ied case, the vorticity has multiple temporal frequencies, as illustrated
Strouhal number, Sr is commonly used as an indicator of the domi- in Fig. 8(a). Since a vorticity is a curl of velocity vectors, its two domi-
nant shedding frequency: nant modes have the same frequencies as within a velocity field. The
modal decomposition by POD and NNM-CNN-AE is illustrated in
Figs. 8(b) and 8(c), respectively. Both methods individually decompose
TABLE II. MSE of reconstruction and prediction (NNM only) of streamwise velocity the vorticity into two pairs of conjugated modes. Higher modes can be
field in both spatial and time domains by POD and NNM-CNN-AE.
visualized both in time domain and spatial domain where there are
larger wavenumbers for the higher-frequency modes. In Fig. 8(c), we
Spatial domain Time domain also observe the distorted nonlinear spatial and temporal fields in the
Reconstruction Prediction Reconstruction Prediction identified NNM modes, indicating that more nonlinear dynamics of
the vorticity is captured by the proposed NNM-CNN-AE method.
POD 2.12  104  3.52  104  Similarly, to study the benefit of the captured nonlinear dynamics
NNM 2.92  105 2.84  105 8.72  106 1.37  105 of the vorticity in the identified NNMs in representing the vorticity
field, in the same fashion with the velocity fields, we compare the

Phys. Fluids 34, 127121 (2022); doi: 10.1063/5.0124455 34, 127121-9

Published under an exclusive license by AIP Publishing
 Physics of Fluids ARTICLE scitation.org/journal/phf

20 November 2023 07:32:15

FIG. 6. Mode decomposition of the transverse velocity field with POD and NNM-CNN-AE. (a) Original transverse velocity spatial- and time-domain flow field. Mid- and right col-
umns: instantaneous frequency and time history of a selected spatial point in the flow, respectively. (b) Identified POD modes (shown the first four) with instantaneous fre-
quency and time history. (c) Single-mode reconstruction of NNM modes (shown the first four) with instantaneous frequency and time history.

Phys. Fluids 34, 127121 (2022); doi: 10.1063/5.0124455 34, 127121-10

Published under an exclusive license by AIP Publishing
 Physics of Fluids ARTICLE scitation.org/journal/phf

20 November 2023 07:32:15

FIG. 7. Transverse velocity field reconstruction and prediction. (a) POD reconstruction field using the first four POD modes; mid-column: the corresponding L2 error in spatial
domain for an instantaneous snapshot; right column: time-domain reconstruction for a selected point in the flow. (b) NNM-CNN-AE reconstruction using the first four NNMs in
spatial and time domains. (c) NNM-CNN-AE prediction in spatial and time domain. The intensity colorbar is at the bottom of (a).

performance of the POD and NNM-CNN-AE in reconstructing vor- consistent with the reconstruction error, suggesting the robustness of
ticity fields using the same number of modes (four POD modes and NNM-CNN-AE in capturing the nonlinear dynamics of the vorticity.
four NNMs, respectively). The POD- and NNMs-reconstructed vor-
ticity fields are shown in Figs. 9(a) and 9(b), respectively, and the 4. Koopman analysis and comparisons with NNMs
reconstruction errors are reported in Table IV. Clearly, NNMs’ recon-
struction of the vortex shedding is more accurate than POD: in spatial We investigate and compare the performance of the Koopman
domain, the error is in the order of 102 with the POD method com- operator in the CNN-AE framework and present the vorticity field
pared to order of 104 with NNMs, whereas in time domain, it is 101 case in this section, in terms of reconstruction and prediction of the
with POD compared to 104 with NNMs. Accurate prediction of the vorticity using the identified Koopman modes. Figure 10 and Table V
vorticity field both in the spatial and time domains using NNM-CNN- present the performance for three different numbers of Koopman
AE is also observed in Fig. 9(c). Interestingly, the prediction error is modes, compared with POD and NNM-CNN-AE methods. It is seen
that Koopman-CNN-AE has similar accuracy to POD in the spatial
domain with the same number of latent space (four modal coordi-
TABLE III. MSE of reconstruction and prediction (NNM only) of transverse velocity nates). As the number of Koopman modes increases to 6, the accuracy
field in both spatial and time domains for POD and NNM-CNN-AE. increases in both time and spatial domains, whereas eight modal coor-
dinates reduce the reconstruction error significantly. However, NNM-
Spatial domain Time domain CNN-AE using only four NNM modes achieves the best accuracy
Reconstruction Prediction Reconstruction Prediction compared to POD and Koopman methods. This is because the nonlin-
earity is captured in all the NNM-CNN-AE blocks with the invariance
POD 0.0011  0.0043  properties of NNMs embedded in the loss function: the encoder block
NNM 9.30  106 1.48  105 1.003  105 2.69  105 is nonlinear, ensuring nonlinear modal transformation; the dynamics
block is also nonlinear, embedded the physics of the system; and the

Phys. Fluids 34, 127121 (2022); doi: 10.1063/5.0124455 34, 127121-11

Published under an exclusive license by AIP Publishing
 Physics of Fluids ARTICLE scitation.org/journal/phf

20 November 2023 07:32:15

FIG. 8. Mode decomposition of the vorticity velocity field with POD and NNM-CNN-AE. (a) Original vorticity spatial- and time-domain flow field. Mid- and right columns: instan-
taneous frequency and time history of a selected spatial point in the flow, respectively. (b) Identified POD modes (shown the first four) with instantaneous frequency and time
history. (c) Single-mode reconstruction of NNM modes (shown the first four) with instantaneous frequency and time history.

Phys. Fluids 34, 127121 (2022); doi: 10.1063/5.0124455 34, 127121-12

Published under an exclusive license by AIP Publishing
 Physics of Fluids ARTICLE scitation.org/journal/phf

20 November 2023 07:32:15

FIG. 9. Vorticity reconstruction and prediction. (a) POD reconstruction field using the first four POD modes; mid-column: the corresponding L2 error in spatial domain for an
instantaneous snapshot; right column: time-domain reconstruction for a selected point in the flow. (b) NNM-CNN-AE reconstruction using the first four NNM modes in spatial-
and time-domain. (c) NNM-CNN-AE prediction in spatial and time domains. The intensity colorbar is at the bottom of (a).

nonlinear decoder block transfers the nonlinear modal space back to The reconstruction, prediction, and identified NNM modes of
the physical flow space. the flow are presented in Fig. 11. As seen in Fig. 11(d), the NNMs
contain the nonlinear spatial and temporal features of the HRN flow.
B. High Reynolds number (HRN) stream This is also visualized in the phase portrait in the modal coordinates
of the first and second modes of POD and NNMs, respectively, where
We also study the performance of the NNM-CNN-AE method
the NNMs curve is more distorted, whereas the trajectory curve of
applied to a high Reynolds number flow past a cylinder, containing
POD modes is circular [Fig. 3]. Figure 11(a) shows the reconstruction
irregular fluctuations and patterns. The spatial domain of the flow
using the first four POD modes. As it can be seen the error is
consists of n ¼ 154 by m ¼ 78 grid points and the cylinder center is
significant in both spatial and time domains, indicating that
located in ðm=3; m=2Þ. The Reynolds number is RD ¼ 1000, and the
POD has failed in capturing the flow with this number of modes (see
steady state regime is considered with 10 000 snapshots and a 0.4-s
Table VI). On contrast, NNM-CNN-AE using the same number of
time steps. The numerical simulation is done using Julia,64 and we pre-
modes, i.e., four NNM modes, has represented this HRN stream
sent the streamwise flow case only.
accurately with quiet small errors in both reconstruction and predic-
tion. The results indicate that the proposed NNM-CNN-AE has
TABLE IV. MSE of reconstruction and prediction (NNM only) of the vorticity field in promising ability to capture and represent the HRN flow, while POD
both spatial and time domains for POD and NNM-CNN-AE. performance deteriorates while increasing Reynolds number or being
in transient regime.
Spatial domain Time domain
C. Transient flow
Reconstruction Prediction Reconstruction Prediction
We apply the NNM-CNN-AE method to study transient flow
POD 0.0275  0.1092  over a cylinder in the range of ½100; 110, which is considered to be
NNM 5.104  104 5.05  104 1.73  104 2.91  104 laminar flow, to understand the nonlinear representation of the
flow for the given regime. It is generated on a 50-s time horizon.

Phys. Fluids 34, 127121 (2022); doi: 10.1063/5.0124455 34, 127121-13

Published under an exclusive license by AIP Publishing
 Physics of Fluids ARTICLE scitation.org/journal/phf

20 November 2023 07:32:15

FIG. 10. Reconstruction of vorticity in both spatial and time domains using different numbers of Koopman modes (KPMs) and fixed number of POD and NNM modes (four
modes). The intensity colorbar is at the bottom of top sub-figures.

TABLE V. MSE of reconstruction and prediction of the vorticity field in both spatial and time domains for POD, Koopman (KPM)-CNN-AE, and NNM-CNN-AE.

Spatial domain Time domain

Reconstruction Prediction Reconstruction Prediction

POD: 4 modes 0.0275  0.1092 

KPM: 4 modes 0.0263 0.0502 0.0401 0.1661
KPM: 6 modes 0.0147 0.0489 0.0328 0.1367
KPM: 8 modes 0.0022 0.0023 0.0038 0.0035
NNM: 4 modes 5.104  104 5.05  104 1.73  104 2.91  104

Phys. Fluids 34, 127121 (2022); doi: 10.1063/5.0124455 34, 127121-14

Published under an exclusive license by AIP Publishing
 Physics of Fluids ARTICLE scitation.org/journal/phf

20 November 2023 07:32:15

FIG. 11. High Reynolds number stream streamwise velocity reconstruction and prediction and the identified NNMs. (a) POD reconstruction using the first four POD modes;
mid-column: the corresponding L2 error in spatial domain for an instantaneous snapshot; right column: time-domain reconstruction for a selected point in the flow. (b)
NNM-CNN-AE reconstruction using the first four NNM modes in spatial- and time domain. (c) NNM-CNN-AE prediction in spatial and time domains. (d) Identified NNM modes
(single-mode reconstruction, shown the first four) with instantaneous frequency and time history first four NNM modes. The intensity colorbar is at the bottom of (a).

We found that across different spatial locations of the whole spatial frequency. Therefore, all the latent modal coordinates have the same
domain, the flow mostly oscillated with its fundamental frequency as temporal frequency.
shown in its wavelet spectra [Fig. 12(e)] where the higher frequency The identified spatial NNMs of the transient flow are seen to con-
is not dominant for the spatial points which oscillate with this tain considerable nonlinear features, as illustrated in Fig. 12(d).

Phys. Fluids 34, 127121 (2022); doi: 10.1063/5.0124455 34, 127121-15

Published under an exclusive license by AIP Publishing
 Physics of Fluids ARTICLE scitation.org/journal/phf

20 November 2023 07:32:15

FIG. 12. Transient streamwise velocity reconstruction and prediction and the identified NNMs. (a) POD reconstruction using the first four POD modes; mid-column: the corre-
sponding L2 error in spatial domain for an instantaneous snapshot; right column: time-domain reconstruction for a selected point in the flow. (b) NNM-CNN-AE reconstruction
using the first four NNM modes in spatial- and time-domain. (c) NNM-CNN-AE prediction in spatial and time domains. (d) Identified NNM modes (single-mode reconstruction,
shown the first four). (e) Instantaneous temporal frequency of two selected spatial points in the flow. The intensity colorbar is at the bottom of (a).

In terms of reconstruction performance, the POD reconstruction required for reconstruction for different flow regimes and flow fields,
using four POD modes [Fig. 12(a)] fails in capturing the physics respectively. The percentage of the total energy that the first two
behind this transient flow. Using the same number of four NNM modes of POD and NNMs, respectively, contain for different flows is
modes, on the other hand, NNM-CNN-AE provide both accurate presented in Table VIII, indicating that NNMs’ fundamental modes
reconstruction and prediction with considerably less error than that of contain a quite high level of total energy, while POD’s fundamental
POD [see Figs. 12(b), 12(c), and Table VII]. modes contribute considerably less energy as the flow complexity
increases from LRN steady state flow to transient or HRN steady state
D. Energy distribution analysis of POD modes flow conditions. This is consistent with the results in Sec. IV A 1–3, B,
and NNMs and C that the identified NNMs contain significant more nonlinear
In this section, we further analyze the energy distribution of POD dynamics features of the flow than POD. On contrast, analyzing the
modes and NNMs, and the number of POD modes and NNMs normalized energy of various flow fields vs POD modes in Fig. 13,

Phys. Fluids 34, 127121 (2022); doi: 10.1063/5.0124455 34, 127121-16

Published under an exclusive license by AIP Publishing
 Physics of Fluids ARTICLE scitation.org/journal/phf

TABLE VI. MSE of reconstruction and prediction (NNM only) of the HRL streamwise V. CONCLUSIONS
velocity field in both spatial and time domains for POD and NNM-CNN-AE.
In this study, we explore the physics of nonlinear normal modes
Spatial domain Time domain (NNMs) and develop a holistic physics-constrained deep learning
framework for mode decomposition and data-driven predictive
Reconstruction Prediction Reconstruction Prediction modeling of fluid flow fields. We construct an NNMs-physics-
constrained convolutional autoencoder (NNM-CNN-AE) which (1)
POD 0.0067  0.86  approximates the nonlinear normal modal transformation through
NNM 2.71  108 2.85  108 3.57  105 5.64  105 deep neural networks with loss functions about the constraints related
to NNMs to discover and visualize the NNMs and the associated
modal coordinates of the flow potentially containing strong nonlinear-
TABLE VII. MSE of reconstruction and prediction (NNM only) of the transient stream- ity and (2) leverages the identified low-dimensional NNMs subspace
wise velocity field in both spatial and time domains for POD and NNM-CNN-AE. through embedding the multi-step prediction in a temporal dynamics
block for data-driven predictive modeling of the nonlinear flow fields.
Spatial domain Time domain For testing cases, we apply the developed method to analyze dif-
Reconstruction Prediction Reconstruction Prediction ferent flow regimes over a cylinder, including laminar flows with low
Reynolds number in transient and steady (RD ¼ 100) states and a
POD 0.0019  0.309  HRN flow (RD ¼ 1000), respectively. The results indicate that the iden-
NNM 1.09  108 1.24  108 2.26  104 3.96  104 tified NNMs are able to reveal the nonlinear spatiotemporal dynamics
of these flows more acurrately and efficiently compared to the linear
POD modes. Using the identified NNM modes (only four modes),
TABLE VIII. First two POD/NNMs modes contribution in terms of energy in recon- NNM-CNN-AE achieves the best accuracy with orders of magnitude
struction of various flow fields and regimes. smaller errors for reduced-order reconstruction and long-time futur-
e-state prediction of these velocity and vorticity flow fields, compared to
LRN LRN HRN the POD and Koopman-CNN-AE methods using the same number or
streamwise transverse LRN LRN streamwise dimension of the modes, suggesting the benefits of the nonlinear spatio-
velocity velocity vorticity transient flow temporal dynamics features captured by NNMs. An analysis of the

20 November 2023 07:32:15

energy distribution of the modes also indicates that NNMs’ fundamental
POD 70% 72% 53% 35% 54% modes capture a quite high level of total energy of the flow, which is ben-
NNMs 94% 99% 96% 99% 66% eficial for reduced-order modeling and representation of the complex
flows, whereas POD’s fundamental modes contribute considerably less
clearly, as the complexity of the flow increases from LRN laminar energy as the flow complexity increases from laminar steady state flow to
to HRN and from steady state to transient, much more POD transient or higher Reynolds number flow conditions.
modes are required to capture the essential characteristics of the flow. Advantages and disadvantages are observed for both NNM and
This reasonably explains the better performance of NNMs over POD POD. In spite of the fact that NNMs are more accurate or more effi-
modes in representing (reconstructing) transient or HRN flow. cient than POD modes for capturing intrinsically nonlinear systems,
the computational cost of NNMs is greater than that of PODs. On the
one hand, as a nonlinear optimization problem, the deep learning of
NNM is more computationally expensive than POD which only
involves a linear optimization. On the other hand, the deep learning of
NNMs requires hyperparameter tuning, whereas POD is almost a
hyperparameter-free method for dimension reduction.
Some limitations are identified in the presented framework which
require future study. First, the computational cost is high as the input
flow data are time series of dense spatial grid (2D flow fields), requir-
ing a high amount of GPU memory, and the training time of NNM-
CNN-AE may exceed even one day if not using supercomputers. In
addition, it is a challenge to identify and decompose the weekly present
modes to reveal subtle features in the flow, as we were not successful
in the transient laminar flow. It will potentially require either much
larger-scale deep neural networks or carefully specified loss functions
to discover the minute features in the flow. Relevantly, it is also found
that the high Reynolds number flow requires much more hyperpara-
meter tuning and training time because its nature is far more complex
FIG. 13. Mode-energy distribution of POD modes for different flow fields and
than the laminar-regime flow with low Reynolds numbers. To mitigate
regimes. Fundamental modes cover the highest portion of total energy of the flow. this challenge, it may require NNM-CNN-AE to incorporate more
More complex flow fields require more POD modes to capture the physics of the latent spaces (higher-dimensional modes) to accurately represent the
flow. complex spatiotemporal dynamics of HRN flows.

Phys. Fluids 34, 127121 (2022); doi: 10.1063/5.0124455 34, 127121-17

Published under an exclusive license by AIP Publishing
 Physics of Fluids ARTICLE scitation.org/journal/phf

ACKNOWLEDGMENTS 10
T. Murata, K. Fukami, and K. Fukagata, “Nonlinear mode decomposition with
convolutional neural networks for fluid dynamics,” J. Fluid Mech. 882, A13
This research uses the supercomputing resources of the Argonne (2020).
Leadership Computing Facility, which is a U.S. Department of Energy 11
M. Rathinam and L. R. Petzold, “A new look at proper orthogonal decomposi-
(DOE) Office of Science User Facility supported under Contract No. tion,” SIAM J. Numer. Anal. 41, 1893–1925 (2003).
DE-AC02-06CH11357. We would also like to gratefully acknowledge
12
S. L. Brunton and B. R. Noack, “Closed-loop turbulence control: Progress and
challenges,” Appl. Mech. Rev. 67, 050801 (2015).
the support and help from Dr. Bethany Lusch of Argonne National 13
J. N. Kutz, “Deep learning in fluid dynamics,” J. Fluid Mech. 814, 1–4 (2017).
Laboratory, and the discussions with Dr. Pin Lyv. 14
M. Raissi, P. Perdikaris, and G. E. Karniadakis, “Physics-informed neural net-
This research was partially funded by the Physics of Artificial works: A deep learning framework for solving forward and inverse problems
Intelligence Program of U.S. Defense Advanced Research Projects involving nonlinear partial differential equations,” J. Comput. Phys. 378,
Agency (DARPA) and the Michigan Technological University faculty 686–707 (2019).
startup fund.
15
J. Ling, A. Kurzawski, and J. Templeton, “Reynolds averaged turbulence model-
ling using deep neural networks with embedded invariance,” J. Fluid Mech.
807, 155–166 (2016).
AUTHOR DECLARATIONS 16
R. Maulik and O. San, “A neural network approach for the blind deconvolution
Conflict of Interest of turbulent flows,” J. Fluid Mech. 831, 151–181 (2017).
17
X. Jin, P. Cheng, W.-L. Chen, and H. Li, “Prediction model of velocity field
The authors have no conflicts to disclose. around circular cylinder over various Reynolds numbers by fusion convolu-
tional neural networks based on pressure on the cylinder,” Phys. Fluids 30,
Author Contributions 047105 (2018).
18
K. Fukami, K. Fukagata, and K. Taira, “Super-resolution reconstruction of tur-
Abdolvahhab Rostamijavanani: Investigation (lead); Methodology bulent flows with machine learning,” J. Fluid Mech. 870, 106–120 (2019).
(equal); Software (lead); Validation (lead); Writing – original draft (lead); 19
S. Lee and D. You, “Data-driven prediction of unsteady flow over a circular cyl-
Writing – review & editing (equal). Shanwu Li: Conceptualization inder using deep learning,” J. Fluid Mech. 879, 217–254 (2019).
(equal); Investigation (equal); Methodology (equal); Project administra-
20
M. Raissi, A. Yazdani, and G. E. Karniadakis, “Hidden fluid mechanics:
tion (equal); Resources (equal); Software (equal); Supervision (equal); Learning velocity and pressure fields from flow visualizations,” Science 367,
1026–1030 (2020).
Validation (equal); Writing – original draft (equal); Writing – review & 21
X. Jin, S. Laima, W.-L. Chen, and H. Li, “Time-resolved reconstruction of flow
editing (equal). Yongchao Yang: Conceptualization (equal); Funding field around a circular cylinder by recurrent neural networks based on non-

20 November 2023 07:32:15

acquisition (equal); Investigation (equal); Methodology (equal); Project time-resolved particle image velocimetry measurements,” Exp. Fluids 61, 114
administration (equal); Resources (equal); Software (equal); Supervision (2020).
(equal); Validation (equal); Writing – original draft (equal); Writing – 22
C. Jiang, J. Mi, S. Laima, and H. Li, “A novel algebraic stress model with
review & editing (equal). machine-learning-assisted parameterization,” Energies 13, 258 (2020).
23
B. Koopman and J. von Neumann, “Dynamical systems of continuous spectra,”
Proc. Natl. Acad. Sci. U. S. A. 18, 255 (1932).
DATA AVAILABILITY 24
I. Mezic, “Spectral properties of dynamical systems, model reduction and
The data that support the findings of this study are available decompositions,” Nonlinear Dyn. 41, 309–325 (2005).
25
I. Mezic, “Analysis of fluid flows via spectral properties of the Koopman opera-
from the corresponding author upon reasonable request. tor,” Annu. Rev. Fluid Mech. 45, 357–378 (2013).
26
H. Arbabi and I. Mezic, “Study of dynamics in post-transient flows using
REFERENCES Koopman mode decomposition,” Phys. Rev. Fluids 2, 124402 (2017).
1
27
S. Bagheri, “Koopman-mode decomposition of the cylinder wake,” J. Fluid
J. N. Kutz, S. L. Brunton, B. W. Brunton, and J. L. Proctor, Dynamic Mode
Mech. 726, 596–623 (2013).
Decomposition: Data-Driven Modeling of Complex Systems (SIAM, 2016). 28
C. W. Rowley, I. Mezic, S. Bagheri, P. Schlatter, and D. S. Henningson,
2
S. L. Brunton, J. L. Proctor, and J. N. Kutz, “Discovering governing equations
from data by sparse identification of nonlinear dynamical systems,” Proc. Natl. “Reduced-order models for flow control: Balanced models and Koopman
Acad. Sci. 113, 3932–3937 (2016). modes,” in Seventh IUTAM Symposium on Laminar-Turbulent Transition
3
K. Taira, S. L. Brunton, S. T. Dawson, C. W. Rowley, T. Colonius, B. J. (Springer, 2010), pp. 43–50.
McKeon, O. T. Schmidt, S. Gordeyev, V. Theofilis, and L. S. Ukeiley, “Modal
29
O. Semeraro, G. Bellani, and F. Lundell, “Analysis of time-resolved PIV mea-
analysis of fluid flows: An overview,” AIAA J. 55, 4013–4041 (2017). surements of a confined turbulent jet using POD and Koopman modes,” Exp.
4
Q. Lu, Y. Sun, and S. Mei, Nonlinear Control Systems and Power System Fluids 53, 1203–1220 (2012).
Dynamics (Springer Science & Business Media, 2013), Vol. 10.
30
S. L. Brunton, B. R. Noack, and P. Koumoutsakos, “Machine learning for fluid
5
R. Maulik, B. Lusch, and P. Balaprakash, “Reduced-order modeling of mechanics,” Annu. Rev. Fluid Mech. 52, 477–508 (2020).
advection-dominated systems with recurrent neural networks and convolu-
31
S. Pan, N. Arnold-Medabalimi, and K. Duraisamy, “Sparsity-promoting algo-
tional autoencoders,” arXiv:2002.00470 (2020). rithms for the discovery of informative Koopman-invariant subspaces,” J. Fluid
6
K. Kunisch and S. Volkwein, “Galerkin proper orthogonal decomposition Mech. 917, A18 (2021).
methods for a general equation in fluid dynamics,” SIAM J. Numer. Anal. 40,
32
H. Arbabi and I. Mezic, “Ergodic theory, dynamic mode decomposition, and
492–515 (2002). computation of spectral properties of the Koopman operator,” SIAM J. Appl.
7
S. Ravindran, “Reduced-order adaptive controllers for fluid flows using POD,” Dyn. Syst. 16, 2096–2126 (2017).
J. Sci. Comput. 15, 457–478 (2000).
33
M. S. Hemati, C. W. Rowley, E. A. Deem, and L. N. Cattafesta, “De-biasing the
8
G. Ortali, N. Demo, and G. Rozza, “A Gaussian process regression approach dynamic mode decomposition for applied Koopman spectral analysis of noisy
within a data-driven POD framework for engineering problems in fluid datasets,” Theor. Comput. Fluid Dyn. 31, 349–368 (2017).
34
dynamics,” Math. Eng. 4, 1–16 (2022). C. Y. Li, Z. Chen, T. K. Tse, A. U. Weerasuriya, X. Zhang, Y. Fu, and X. Lin,
9
D. Baroli, C. M. Cova, S. Perotto, L. Sala, and A. Veneziani, “Hi-POD solution “Establishing direct phenomenological connections between fluid and structure
of parametrized fluid dynamics problems: Preliminary results,” in Model by the Koopman-linearly time-invariant analysis,” Phys. Fluids 33, 121707
Reduction of Parametrized Systems (Springer, 2017), pp. 235–254. (2021).

Phys. Fluids 34, 127121 (2022); doi: 10.1063/5.0124455 34, 127121-18

Published under an exclusive license by AIP Publishing
 Physics of Fluids ARTICLE scitation.org/journal/phf

35
P. J. Schmid, “Dynamic mode decomposition of numerical and experimental 51
M. Peeters, G. Kerschen, and J.-C. Golinval, “Modal testing of nonlinear vibrat-
data,” J. Fluid Mech. 656, 5–28 (2010). ing structures based on nonlinear normal modes: Experimental demonstra-
36
B. Lusch, J. N. Kutz, and S. L. Brunton, “Deep learning for universal linear tion,” Mech. Syst. Signal Process. 25, 1227–1247 (2011).
embeddings of nonlinear dynamics,” Nat. Commun. 9, 4950 (2018). 52
C. Touze and M. Amabili, “Nonlinear normal modes for damped geometrically
37
E. Yeung, S. Kundu, and N. Hodas, “Learning deep neural network representa- nonlinear systems: Application to reduced-order modelling of harmonically
tions for Koopman operators of nonlinear dynamical systems,” in 2019 forced structures,” J. Sound Vib. 298, 958–981 (2006).
American Control Conference (ACC) (IEEE, 2019), pp. 4832–4839. 53
G. I. Cirillo, A. Mauroy, L. Renson, G. Kerschen, and R. Sepulchre, “A spectral
38
N. Takeishi, Y. Kawahara, and T. Yairi, “Learning Koopman invariant subspa- characterization of nonlinear normal modes,” J. Sound Vib. 377, 284–301 (2016).
ces for dynamic mode decomposition,” in Advances in Neural Information 54
K. Worden and P. Green, “A machine learning approach to nonlinear modal
Processing Systems (NeurIPS Proceedings, 2017), pp. 1130–1140. analysis,” Mech. Syst. Signal Process. 84, 34–53 (2017).
39
S. E. Otto and C. W. Rowley, “Linearly recurrent autoencoder networks for 55
N. Dervilis, T. E. Simpson, D. J. Wagg, and K. Worden, “Nonlinear
learning dynamics,” SIAM J. Appl. Dyn. Syst. 18, 558–593 (2019). modal analysis via non-parametric machine learning tools,” Strain 55, e12297
40
I. Mezic, “Koopman operator, geometry, and learning of dynamical systems,” (2019).
Not. Am. Math. Soc. 68, 1087–1105 (2021). 56
S. Li and Y. Yang, “Data-driven identification of nonlinear normal modes via
41
S. Leask, V. McDonell, and S. Samuelsen, “Modal extraction of spatiotemporal physics-integrated deep learning,” Nonlinear Dyn. 106, 3231–3246 (2021).
atomization data using a deep convolutional Koopman network,” Phys. Fluids 57
M. Amabili and C. Touze, “Reduced-order models for nonlinear vibrations of
33, 033323 (2021). fluid-filled circular cylindrical shells: Comparison of POD and asymptotic non-
42
M. Li and L. Jiang, “Deep learning nonlinear multiscale dynamic problems linear normal modes methods,” J. Fluids Struct. 23, 885–903 (2007).
using Koopman operator,” J. Comput. Phys. 446, 110660 (2021). 58
S. Li and Y. Yang, “Hierarchical deep learning for data-driven identification of
43
R. M. Rosenberg, “Normal modes of nonlinear dual-mode systems,” J. Appl. reduced-order models of nonlinear dynamical systems,” Nonlinear Dyn. 105,
Mech. 27, 263 (1960). 3409–3422 (2021).
44
G. Kerschen, M. Peeters, J.-C. Golinval, and A. F. Vakakis, “Nonlinear normal 59
K. Fukami, T. Murata, K. Zhang, and K. Fukagata, “Sparse identification of
modes, Part I: A useful framework for the structural dynamicist,” Mech. Syst. nonlinear dynamics with low-dimensionalized flow representations,” J. Fluid
Signal Process. 23, 170–194 (2009). Mech. 926, A10 (2021).
45
S. Shaw and C. Pierre, “Non-linear normal modes and invariant manifolds,” 60
J.-C. Loiseau and S. L. Brunton, “Constrained sparse Galerkin regression,”
J. Sound Vib. 150, 170 (1991). J. Fluid Mech. 838, 42–67 (2018).
46
S. W. Shaw and C. Pierre, “Normal modes for non-linear vibratory systems,” 61
K. Champion, P. Zheng, A. Y. Aravkin, S. L. Brunton, and J. N. Kutz, “A unified
J. Sound Vib. 164, 85–124 (1993). sparse optimization framework to learn parsimonious physics-informed models
47
R. H. Rand, “A higher order approximation for non-linear normal modes in from data,” IEEE Access 8, 169259–169271 (2020).
62
two degree of freedom systems,” Int. J. Non-Linear Mech. 6, 545–547 (1971). P. Holmes, J. Lumley, and G. Berkooz, Turbulence, Coherent Structures,

20 November 2023 07:32:15

48
L. Manevich and I. V. Mikhlin, “On periodic solutions close to rectilinear nor- Dynamical Systems and Symmetry (Cambridge University, 1998).
mal vibration modes,” J. Appl. Mech. 36, 988 (1972). 63
N. E. Sujovolsky and P. D. Mininni, “Extraction of invariant manifolds and
49
A. F. Vakakis, Analysis and Identification of Linear and Nonlinear Normal application to turbulence with a passive scalar,” Phys. Rev. E 103, 063107
Modes in Vibrating Systems (California Institute of Technology, 1991). (2021).
50
L. Renson, A. Gonzalez-Buelga, D. Barton, and S. Neild, “Robust identification 64
B. Font, G. D. Weymouth, V.-T. Nguyen, and O. R. Tutty, “Deep learning of
of backbone curves using control-based continuation,” J. Sound Vib. 367, the spanwise-averaged Navier–Stokes equations,” J. Comput. Phys. 434,
145–158 (2016). 110199 (2021).

Phys. Fluids 34, 127121 (2022); doi: 10.1063/5.0124455 34, 127121-19

Published under an exclusive license by AIP Publishing