Deep learning based on PINN for solving 2 D0F vortex induced

vibration of cylinder with high Reynolds number

Chen Cheng1, Peng-Fei Xu1, Yong-Zheng Li3, Guang-Tao Zhang2,
1. College of Harbor, Coastal and Offshore Engineering, Hohai University, Jiangsu, China
2. College of Mathematics and Informatics, South China Agricultural University, Guangdong,
China
3. School of Naval Architecture and Ocean Engineering, Jiangsu University of Science and
Technology, Jiangsu, China

Abstract

Vortex-induced vibration (VIV) exists widely in natural and industrial fields. The
main approaches for solving VIV problems are numerical simulations and experimental
methods. However, experiment methods are difficult to obtain the whole flow field
information and also high-cost while numerical simulation is extraordinary time-
consuming and limited in low Reynolds number and simple geometric configuration.
In addition, numerical simulations are difficult to handle the moving mesh technique.
In this paper, physics informed neural network (PINN) is proposed to solve the VIV
and wake-induced vibration (WIV) of cylinder with high Reynolds number. Compared
to tradition data-driven neural network, the Reynolds Average Navier-Stokes (RANS)
equation, by implanting an additional turbulent eddy viscosity, coupled with structure’s
dynamic motion equation are also embedded into the loss function. Training and
validation data is obtained by computational fluid dynamic (CFD) technique. Three
scenarios are proposed to validate the performance of PINN in solving VIV and WIV
of cylinders. In the first place, the stiffness parameter and damping parameter are
calculated via limited force data and displacement data; secondly, the flow field and
lifting force/drag force are inferred by scattered velocity information; eventually, the
displacement can be directly predicted only through lifting forces and drag forces based
on LSTM. Results demonstrate that, compared with traditional neural network, PINN
method is more effective in inferring and re-constructing the unknown parameters and
flow field with high Reynolds number under VIV and WIV circumstances.

Keywords: vortex-induced vibration; physics informed neural network; wake-induced

vibration; fully connected neural network; long-short term memory
1. Introduction

Flow induced vibration (FIV) problems are ubiquitous in natural and industrial
processes, such as pipeline in sea mining, cylinder of offshore and wind turbine etc.
Vortex induced vibration (VIV) of bluff bodies, as a typical branch of FIV, will be
happened when the vortex shedding frequency is close to the natural frequency of the
structure. VIV can generate a huge amplitude vibration of structures according to the
specific reduced velocity, the Reynolds number and structural dynamic characteristics
(Blevins, 1990; Williamson and Govardhan, 2004). Sometimes, VIV can even cause
large fatigue damage to the structures that attracts a substantial amount of attention.
Numerical simulation of VIV problems principally relies on solving the RANS
equation and dynamic motion equation in a discretized form through finite element
method (FEM), finite volume method (FVM) or finite difference method (FDM), which
are described as computational fluid dynamic (CFD) method. However, CFD
techniques are cumbersome in computational efficiency, especially for turbulent flow
and complicated geometries. Furthermore, CFD techniques are also limitative in
handling the moving mesh and other particular technical means.
Reduced order modeling (ROM), as one of the system identification, has been
viewed as a strong tool to decrease the complexity and high dimensionality of the
dynamical models and firstly proposed in optimal design, optimal control and inverse
problem application. Proper orthogonal decomposition (POD) and dynamic mode
decomposition (DMD) are two dominant methods of ROM in solving flow dynamics
in lower dimensional representations (Dowell, 1997; Schmid, 2010). Henshawa et al
(2007) utilized POD to construct the non-linear model of the aircraft behavior with low
dimensionality and evaluate the performance on the real aircraft. Jovanovie et al (2014)
developed a sparsity-promoting variant of the standard DMD algorithm to represent the
flow field by numerical simulation and then compared to the experiments. The results
showed that method can well re-construct the fluid model. Hemati et al (2014)
formulated a low-storage approach to perform DMD to simulate the flow past cylinder
and compared with the results from particle image velocimetry experiments. However,
ROM also has limitations in solving complicated unsteady flows due to the information
loss by compressive model. However, ROM makes fluid dynamics into the linear or
weakly nonlinear problems with powerful assumptions which has limitation in
complicated unsteady flow.
Deep learning (DL) technology has extraordinary ability to deal with the strong
nonlinearity and high dimensionality (LeCun et al, 2015). Recently, DL has a
tremendous breakthrough in some fields, such as speech recognition, image processing
and event prediction (Goodfellow et al, 2016; Xiong et al, 2015). More recently, DL
method is proposed for solving fluid dynamics. Ling et al (2016) constructed the deep
learning of RANS turbulence model by embedding Galileo invariant into depth neural
network, and firstly realized the prediction of channel flow vortex and separated flow.
This is considered to be the first combination of deep neural networks and fluid
mechanics (Nathan, 2017). Yeung et al (2017) proposed a deep learning framework for
computing Koopman operators of nonlinear dynamic systems, which provides a new
idea for modeling nonlinear dynamic systems by combining DMD method with deep
neural networks. Miyanawala and Jaiman (2017) predicted the flow characteristics in
the wake region of a two-dimensional cylinder by deep convolution network. Jin et al
(2018) utilized fusion convolutional neural networks (CNNs) to predict the velocity
fields around the circular cylinder by data obtained by pressure fields. Sekar et al (2019)
also adopted CNNs technique combined with Multilayer Perceptron (MLP) to calculate
the incompressible laminar steady flows. Recurrent neural network (RNN) is another
powerful tool to predict temporal features of flow fields. Deng et al (2019) utilized the
Long Short-Term Memory (LSTM) to obtain the time coefficient of the flow field.
Mohan et al (2019) combined the CNNs and LSTM to predict the spatial-temporal
features of turbulence dynamics. However, DL methods require magnanimous data to
ensure the prediction accuracy and generalization ability. In addition, DL methods build
up a surrogate model which is considered as black box and it means that the model
lacks physical interpretation.
Raissi et al (2017) firstly proposed physics informed neural network (PINN) to
solve the partial differential equations (PDE) and inverse problems. PINN modified the
traditional form of the loss function and was embedded with the physical models, with
its important breakthrough featuring that the PINN can predict the variables based on
physical laws. Tartakovsky et al (2018) utilized PINN to construct the constitutive
equations of Decay flow. It demonstrated that PINN has strong performance in solving
inverse problems. Moreover, Yang et al (2020) employed Bayesian and PINN to solve
the PDE with noisy data.
The aim of this paper is to utilize PINN method to solve the VIV and wake-induced
vibration of cylinders. The turbulence eddy viscosity is introduced into the RANS
model and then embedded into the loss function. A fully connected neural network and
LSTM are adopted to construct the structure of the PINN. The whole flow field and
unknown parameters (such as damping coefficient and stiffness coefficient) are
calculated by PINN based on scattered training samples. The structure of paper can be
demonstrated as follow. Section 2 introduces the governing equations of fluid
mechanics and dynamic motion of the cylinders. Section 3 describes the principle of
the FCNN and LSTM, then the scheme of PINN is introduced in Section 4. Section 5
demonstrate the three scenarios and show the performance of PINN in these scenarios.
Conclusion is summarized in section 6.

2. Vortex induced vibration

2.1 Governing equations of fluid mechanics

The incompressible flow is governed by the Navier-Stokes (N-S) equation and

conservation equation which can be shown as:
 u  0 x, t  f ,t ,  d

 u, p   0   u 1 (1)
 t   u   u   p  u  b f  0, x, t  f ,t , 
2 d


where u denotes the velocity field (including u, v, w); p the pressure field;  the
kinematic viscosity; b f the body force.
Incompressible flow can be solved with the proper initial and boundary conditions
by the numerical simulation. However, the turbulent flows are generated with the
increasing Reynolds number, the N-S equation is difficult to be solved directly owing
to not only huge computational expense, but also the illness or stiffness of the algebraic
matrices involved (Durbin, 2018). Reynolds-average N-S (RANS), as a strong tool in
industrial practices, is proposed to solve the turbulent flow. The governing equation can
be demonstrated as follow:
ui
0
xi
(2)
 f ui  f ui u j p  u u 
    2ui 
f i j

t xi xi x j
where:
 u u j  2
  uiuj  t  i     kt ij (3)
 x
 j xi  3
where   uiu j denotes Reynold stress  ij ;   the Reynolds average or the spatial
filtering, and ui  ui  ui . According to the Fick’s law, Reynold stress can be re-
modelled as:
ui
 uiuj   t (4)
x j

where  t denotes the turbulent eddy viscosity. The value of  t is determined by the
flow filed. The parameter  t has been calibrated by various methods for several
decades (Poroseva et al, 2016). It is difficult to obtain the  t in a universal sense due
to the case-by-case dependence. Fortunately, a great amount of practices shows that
modelling eddy viscosity can well establish the fitting between filtered experimental
data and solutions of the RANS. With this ideal, the traditional RANS equation can be
re-modeled as follow (Bai et al, 2021):
u u u p   2u  2u 
u v      t   2  2 
t x y x  x y 
v v v p   2v  2v 
u v      t   2  2  (5)
t x y y  x y 
u v
 0
x y

2.2 Kinematic equation and discrete method

 The oscillation of the cylinder can be described as a typical mass-spring-damper
system. Therefore, the motion equations of cylinder in x-direction and y-direction can
be shown as follow:
d 2 d
msystem 2
c  k  FL  t 
dt dt
(6)
d 2 d
msystem 2
c  k  FD  t 
dt dt
where msystem denotes the inertial mass of the vibration system;  and  the motion
in x-direction and y-direction, respectively; c the damping and k the stiffness
coefficient; FL  t  and FD  t  represent the lift force and drag force, respectively.
The velocities of the oscillatory cylinder can be calculated as follow:
vx  t     t 
(7)
vy t    t 

The initial conditions and boundary conditions of the oscillated cylinder can be
described as:
vx  0   v y  0  =0
(8)
x  0  y  0  0

When the lift force and drag force are calculated by integrating the pressure and
velocity gradients, the Eq. 6 can be discretized by the fourth-order Runge-Kutta method.
The equation in y-direction can be expressed as follow:
t
  tn 1     tn    k1  2k2  2k3  k4 
6
(9)
t 2
  tn 1     tn  +v y  tn  t   k1  k2  k3 
6
where:
FL  tn  c k
k1   v y  tn  t    tn 
msystem msystem msystem
FL  tn  c  t  k  t 
k2  
msystem msystem v y  tn   2 k1   m   tn   2 v y  tn  
system
(10)
F t  c  t  k  t t 2 
k3  L n  v y  tn   2 k2   m   tn   2 v y  tn   4 k1 
msystem msystem system  
FL  tn  c k  t t 2 
k4    v y  tn   k3 t  
   tn   v y  tn  
msystem 
k2 
msystem msystem  2 2 

where k1 、 k 2 、 k 3 and k 4 are the coefficients of the fourth-order Runge-Kutta,  t

denotes the time step; v y the velocity of the cylinder in y-direction. The equation in
x-direction is same as that in y-direction.

3. Deep learning

3.1 Fully connected neural network

Fully connected neural network (FCNN) is a classic structure of neural network.

FCNN consist input layer, hidden layer and output layer which can be viewed in Fig. 1.
Layer 0 is input layer and layer L is output layer, the other layers are hidden layers.
Each layer includes large number of neurons which have weights, biases and activation
functions. The weights and biases are tuned by training the neural network. It is
noteworthy that the activation function plays a significant role in handling the nonlinear
problems. The common activation functions are sigmoids, tanh and rectified linear units.
In the recent year, the adaptive activation function is demonstrated has better
performance in solving strong nonlinearity (Jagtab et al, 2019). The output of a neuron
can be calculated as follow:
z lj  wljk fl 1  zkl 1   blj (11)
l l
where z j denotes the output of neuron j in layer l; w jk the weight between neuron
k in layer l-1 and neuron j in layer l; f   the activation function; b j the bias of
l

neuron j in layer l. The formula can also be written as:

z lj   wljk ykl 1  blj (12)
k

Fig. 1. The structure of the fully-connected neural network

The input data x   x1 , x2 , xn  and output data y   y1 , y 2 , yn  are utilized to

adjust the parameters in neural structure including weights and biases. The approximate
result y L predicted by FCNN is compared to the exact value y , the difference
between approximate result and exact result is defined as cost function which can be
shown as:
w , b  arg min w,b C  y, y L  (13)

where w and b are tuned weights and biases, respectively; C y, y L represents  

the cost function. How to reduce the loss function as much as possible is the premise to
ensure that the neural network can effectively predict the concerned results. Generally,
backpropagation is a standard approach to compute the gradients and can be viewed as
follow:
C
 lj  (14)
z lj
Go a step further, the gradient of the cost function can be computed as another form
which can be demonstrated:
C C
 jL   L  z Lj  ,  ykl 1 lj ,
y j
L
w jk
l

(15)
C
   w   l  z  ,
l l 1 l 1 L
  lj
blj
j kj k j
k

The  -term in Eq. (5) can be expressed as vector form:

 L  z L  ,  l  W l 1   l 1  l  z l 
T
 L = yL C (16)
T
 C C 
 yL C   L ,
, L (17)
 y1 yM 
where is the Hadamard product. The notation C without a subscript the vector
of partial derivatives in respect of the input x   x1 , x2 , xn  .

3.2 Long short-term memory

Long short-term memory (LSTM) is a kind of time series neural network, which
is specially designed to solve the long-term dependence problem of general RNN
(recurrent neural network). The advantage of LSTM is to store and memorize previous
information which can reduce the complexity and number of layers in its structure
(Wang, 2017).
The structure of LSTM includes input gate, forget gate, block input, cell state,
output gate and block output. By controlling to open and close the gates, LSTM is able
to truncate gradients in the neural network. The model of the LSTM can be described
as follow:
 
it   Wxi xt  Whi ht 1  bii 

ft   Wx f xt  Wh f ht 1  bi f 

zt  tanh Wxc xt  bc  (18)
ct  ft ct 1  it zt

ot   Wxo xt  Who ht 1  bio 
ht  ot tanh  ct 
where xt denotes the input vector at time step t; W  and b  the weight matrix and
threshold vector respectively;    the activation function; is the Hadamard
product; LSTM can control the flow of data information via opening and closing the
different gates which can be demonstrated in Fig. 2.

Fig. 2. The structure of the long short-term memory

4. Physics informed neural network

4.1 Physics-constrained deep learning

Conventionally, DL method builds up a surrogate model, such as FCNN or CNN,

for predicting the solution of the fluid flow which are approximately equal to real values
(Zhu et al, 2018).
f  t , x,    f  t , x,   z l  t , x,  ; W,b  (19)

where f is the solution vector including the velocity fields and pressure fields; W and b
denote the weights and biases, respectively. z l  t , x,  ; W,b  the predicted by the
surrogate model; f the locally minimized. The solution of flow dynamics can be cast
into an optimization problem which can be demonstrated as follow:
 1 d
f  t , x,    z l  t , x,  ; W,b 
2
data ( W, b) 
N (20)
W , b  argmin data ( W, b)
w,b

where data ( W, b) denotes the loss function based on data; N the number of training
samples. f d the training data.
However, the traditional DL requires large number of training data, which is too
difficult to achieve from time-consuming CFD simulation. Physics-constrained deep
learning embeds the physical model into the loss function by minimizing the violation
of the solution on the basis of the known partial differential equations for fluid flows
over a domain of interests without the demands of handling these equations for each
parameter with conventional numerical simulations. The residual of N-S equations and
mass conservation equations are computed by FCNN and the specific loss function can
be demonstrated as follow:
2
u
Nf Nf
1 1 1
phy ( W, b )      u   u  p  2u  b f +  u
2

Nf i 1 t  Nf i 1

Structure imposed by N  S equations Mass conservation

 
W , b  argmin phy ( W, b ) (21)
w,b

  x, p, u,    0, t  0, in  f
s.t . 
  t , x, p, u,    0 on  f ,t
where phy (W, b) denotes the physics-based loss; and the initial and boundary
conditions, respectively;
The first and/or second derivative terms of velocity and pressure in the loss function
can be computed by the automatic differentiation approach (AD) (Baydin et al, 2018).
Compared to the traditional differential calculation, such as Manual Differentiation,
Numerical Differentiation and Symbolic Differentiation, the core problem of AD is to
calculate the derivatives, gradients and Hessian matrix values of complex functions,
which are usually multi-layer composite functions at a certain point. The advantage of
the AD is more accurate due to the absence of truncation or round-off errors. Generally,
AD can be directly utilized in deep learning framework such as Tensorflow, Pytorch
and Theano (Paszke et al, 2017; Abadi et al, 2016; Bastien et al, 2012). In order to
reduce the error of the loss function, the Adam optimizer is utilized to optimize the
target function. Adam optimizer can constantly adjust the learning rates with the
situation changes in the learning process (Diederik and Jimmy, 2017). ‘Xavier’ method
is designed to decide the initial weights and biases which can ensure faster convergence
of neural network (Glorot and Bengio, 2010). A residual neural network is added in the
FCNN to avoid gradient explosion and/or gradient disappearance (He et al, 2016).

4.2 Initial and Boundary condition enforcement

 The loss function constrained by the physical equations becomes identically zero,
the predicted values of velocity and pressure fields will precisely satisfy the N-S
equations. Consequently, the solutions driven by FCNN particularly have physical
interpretation through penalizing the PDE residuals. Furthermore, to make the problem
well-posed, the appropriate initial conditions and boundary conditions are required and
imposed as constraints which are dealt with a soft manner by amending the original loss
function with penalty terms (Márquez-Neila et al, 2017). The Eq. (10) can be rewritten
by adding initial loss and boundary loss as follow:
c
phy  W, b, i , b   phy  W, b   i  x, p, u,   f ,t
 b  t , x, p, u,   (22)
f ,t
Equation loss
Initial loss Boundary loss

where i and b are penalty coefficients.

5. PINN for solving vortex induced vibration of 2DOF cylinder

5.1 CFD method for obtaining data

CFD techniques of the 2 DOF of VIV and WIV are carried out and the simulation
results are selected as training data. It is interesting to note that the experimental data
can also be utilized for training neural network. 2D flow field is calculated through the
solver, pimpleDyMFoam, executed in OpenFOAM, which is an open source framework
of FVM. Shear stress transport (SST) k   , as a known turbulence model is employed.
Furthermore, the nested grid technique, which is the latest dynamic grid, is adopted to
handle the moving boundary of cylinder.
The whole computational zone is a rectangle region with the length of 40D and the
width of 20D. D is the diameter of cylinder and located in origin of coordinate. The
inlet flow is enforced on the left part of the computational zone with a Dirichlet
boundary condition u  U  , 0  while the outlet is Neumann boundary (zero-gradient
pressure) at the right part of computational zone. The distance between the inlet and the
center of the cylinder is 10D while the distance between the outlet and the center of the
cylinder is 30D, that can guarantee that the cylinder is not distributed by remoting
boundary. The upper and bottom part of computational zone is slide-wall. The Reynolds
number, Re  U  D  , is 2889. The parameters of cylinder are the damping parameter
c  0.07444 and the stiffness parameter k  17.2589 . The concrete details can be
viewed in Fig. 3.
 Fig. 3. The sketch of the vortex induced vibration of cylinder

5.2 Inferring damping and stiffness parameters from forces and displacement

It is fact that the stiffness and damping parameters of cylinder cannot be measured
directly but the force and displacement of cylinder can be measured by force balance
and laser range finder. Therefore, the PINN method is adopted to infer the damping and
stiffness parameter through the limited force and displacement datasets. The PINN for
solving parameters of cylinder can be viewed in Fig. 4 and the physical law can be
described as follow:
FL : mtt  ct  k
(23)
FD : m tt  c t  k
It is noteworthy that the damping and stiffness parameters are transformed into the
parameter of the resulting PINN. The loss function can be viewed as follow:
Loss    t n    n    t n    n  FL  t n   FL n  FD  t n   FD n (24)

Training dataset (N=120), corresponding to exact damping and stiffness parameters,

is utilized to tune a neural network with 10 hidden layers and 32 neurons in each layer
by minimizing the sum of mean squared error of loss function utilizing the Adam and
L-BGFS optimizer. The PINN is applied to predict the whole solution functions of
forces and displacements in two degrees, as well as the stiffness and damping
parameters. The predicted values for stiffness and damping parameter are k  17.1983
and c  0.07138 . This corresponds to about 0.62% and 0.02% comparative errors in
the predicted results for k and c, respectively. The entire functions of forces and
displacement can also well be inferred by PINN which can be viewed in Fig. 5. It is
obvious that the PINN can obtain the unknown parameters of cylinder effectively from
limited dataset and also achieve the entire force and displacement functions.
 Fig. 4. PINN method for solving damping and stiffness parameters

Fig. 5. PINN method for predicting the force and displacement

5.3 Inferring lift force, drag force and pressure from scatter velocity field

The aim of this section is to reconstruct the whole flow field with high Reynolds
number and infer the forces (including lifting force and drag force) enforcing on the
   
moving cylinder based on scattered information t n , n ,  n and t n , xn , y n , u n , vn .
It is worth recalling that the pressure information and turbulent eddy viscosity are
viewed unknown parameters that also need to be solved. The loss function induced by
partial differential equations includes three parts which are contributed by the x-
component velocity u , y-component velocity v and mass conversation, respectively.
It can be demonstrated as follow:
 u u u p   2u  2u 
e1   u  v     t   2  2    tt
t x y x  x y 
v v v p   2v  2v 
e2   u  v     t   2  2   tt (25)
t x y y  x y 
u v
e3  
x y

For the simplicity of presentation, the over-line symbol for the operator in Eq.2 is
omitted. It should be noted that the horizontal displacement and vertical displacement
are incorporated into e1 and e2 , respectively, so that the fluid flow coordinate system
is attached to the cylinder. The total loss function can be determined as:
Lsum  Lu  Lv  L  L  Le
 u t n , xn , y n   u n  v t n , xn , y n   vn (26)

   t n    n    t n    n   ei
3

i 1

The specific process of PINN to solve this problem can be viewed in Fig. 6. The
fully-connected neural network including 12 hidden layers with 32 neurons in each
layer computes the gradient of loss function via Adam optimizer. The differential
operations are generated by automatic differentiation implemented in Tensorflow. The
adaptive activation function is adopted in each layer to enhance the
nonlinear processing capability. The Latin hypercube sampling (LHS) is adopted to
obtain the training data in different snapshots, the total numbers of training data are
40000. Three snapshots of flow field ( t  30s, 40s,50s ) are selected to validate the
PINN

Fig. 6. PINN method for solving lift force, drag force and entire fluid flow
 Fig. 7 demonstrates the whole flow field simulated by CFD technique and inferred
flow field by PINN at different snapshots. It is obvious that the proposed framework is
able to reconstruct the whole velocity field with high Reynold number accurately which
can be viewed in Fig. 7(a) and Fig. 7(b). A remarkable result originates from PINN’s
ability to infer the whole pressure field accurately in defect of any training samples on
the pressure itself (view Fig. 7(c)). The mean square errors of velocity field and pressure
field are listed in Table 1. A strange phenomenon is that the difference in magnitude
between the predicted pressure and exact one, although the distribution of the pressure
filed is almost same. It is validated by the law of the N-S equation due to the pressure
field is only recognizable up to a fixed value. For the incompressive flow, the absolute
value of the pressure is of no great important.

(a)

(b)
 (c)
Fig. 7. PINN method for reconstructing the flow field with different times ((a) u, (b) v, (c) p)

Table 1. Mean square errors of entire flow fields

Mean square errors
u v p

t  30s 2.83 10-3 0.82 10-3 0.4 100

t  40s 1.62 10-3 1.36 10-3 0.87 100

t  50s 1.42 10-3 0.92 10-3 0.21100

When the pressure and velocity fields are obtained, the lift force and drag force on
the cylinder can be approximately calculated, based on the function of the pressure and
velocity gradients, through trapezoidal law as:

 2 v 1  u v  
FL   


 pn y 
Re y
n y     nx ds
Re  y x  
(27)
 2 u 1  u v  
FD     pnx  nx     ny ds

 Re x Re  y x  

 
where nx , ny denotes the outward normal on the cylinder while ds the arc length on
the surface of the cylinder.
Fig. 8 shows the comparison between the inferred drag and lifting forces with the
exact ones. The blue solid lines represent exact values while the orange dotted line
represent inferred values. The mean square of errors of lifting force and drag force are
0.17 104 and 0.63 104 . PINN can well calculate the forces on cylinder due to the
accurate prediction of velocity and pressure fields and then infer the damp and lifting
coefficients by above setup. Therefore, in practical engineering, we only utilize the
particle image velocimetry (PIV) to obtain the scattered velocity information that can
infer the whole flow field and forces on structures. It is no doubt that it greatly reduces
the difficulty of obtaining the experimental data of vortex induced vibration.
 Fig. 8. PINN method for calculating the lift force and drag force

5.4 PINN for solving wake-induced vibration of the cylinder behind a cylinder

In this part, the wake-induced vibration of a 2-DOF cylinder which is behind a 2-

DOF cylinder is investigated by PINN method. The flow field between two cylinders
and surfaces of cylinders become more complex and this problem can be utilized to
validate the applicability of PINN in more complicated VIV setup.
The boundary conditions are same as above and the distance between two cylinders
is 6D which guarantees that the wake field of upstream structure has sufficient
development space. The Reynolds number, Re  U  D  , is 9000 in this setup and
flow can be viewed as turbulent. The parameters of cylinder are the damping parameter
c  0.5183 and the stiffness parameter k  2530.113 . The concrete details can be
viewed in Fig. 9.

Fig. 9. The sketch of the wake-induced vibration of cylinder behind cylinder

In this section, two freedom degrees of displacements of two cylinders are

incorporated into the loss function and total loss function can be shown as:
u u u p   2u  2u 
e1   u  v     t   2  2    1tt   2tt
t x y x  x y 
v v v p   2v  2v 
e2   u  v     t   2  2   1tt  2tt (28)
t x y y  x y 
u v
e3  
x y
Lsum  Lu  Lv  L  L  Le
 u t n , xn , y n   u n  v t n , xn , y n   vn

 1  t n   1n   1  t n    1n  2  t n   2n   2  t n    2n (29)

3
  ei
i 1

More training samples are selected in calculation region between two cylinders in
order to enhance the predictive performance of PINN in WIV setup. The number of
total training dataset is 45000. The entire flow fields at t  80s,90s,100s are adopted
to validate and the results can be demonstrated in Fig. 10. It is obvious that PINN well
infers the whole flow field (velocity field and pressure field) at different times from
scattered velocity information. The mean square errors of velocity and pressure can be
viewed in Table 2.

Table 2. Mean square errors of entire flow fields of wake-induced vibration

Mean square errors
u v p

t  80s 4.5110-3 2.4110-3 2.1100

t  90s 3.39 10-3 1.86 10-3 3.47 100

t  100 s 3.84 10-3 3.32 10-3 2.85 100

 (a)

(b)

(c)
Fig. 10. PINN method for reconstructing the flow field with different times ((a) u, (b) v, (c) p)

With the cases investigated, the turbulent eddy viscosity is also introduced as an
unknown parameter that need to be inferred. Fig. 11 indicates the inferred  t predicted
by PINN and the reference  t simulated by CFD. The results show that the PINN has
an effective adaptivity to approximate the unknown parameter from turbulence flow
and the magnitude of mean square error at different times reaches to 104 . This
treatment represents that PINN technique could have a transformative effect for
modelling turbulence closure.
Furthermore, drag force and lifting force on two cylinders predicted by PINN are
also considered in this case and can be shown in Fig. 12. The mean square of errors of
FL (cylinder 1), FD (cylinder 1), FL (cylinder 2), FD (cylinder 2), are 2.87 10-4 ,
1.32 10-4 , 5.89 10-5 , 3.34 10-5 , respectively.

Fig. 11. PINN method for inferring eddy viscosity

Fig. 12. PINN method for inferring lifting forces and drag forces of two cylinders

5.5 PINN for solving unknown displacements and force based on recurrent neural
network

In this section, the displacements of cylinders are inferred directly based on lifting
forces and drag forces through physics informed LSTM. The damping coefficient and
stiffness coefficient are assumed as known parameters in this case. Compared to the
traditional LSTM, the initial/boundary conditions (Eq. 8) and 4-th Runge-Kutta
integrations (Eq. 9 and Eq. 10) are embedded into the LSTM cell which can be viewed
in Fig. 13 and the comparison between inferred trajectories of two cylinders and exact
ones is described in Fig. 14. Blue lines represent the exact values while red lines
represent the predicted values. For the upstream cylinder, the trajectory like butterfly
shape can be well inferred by the PINN. A more intriguing result is that the trajectory
of downstream cylinder is more irregular, due to the complexity flow field between two
cylinders, and can also well predicted by the PINN technique. The mean square errors
of two trajectories can be listed in Table 3.

Fig. 13. Physics informed LSTM

Fig. 14. PINN method for inferring trajectories of two cylinders

 Table 3 The mean square error of displacements of two cylinders

1 1 2 2

MSE 2.110-6 2.6 10-6 1.110-6 2.8 10-6

6. Conclusion

In this paper, PINN based on FCNN and LSTM is adopted to solve the 2 DOF
vortex-induced vibration and wake-induced vibration of cylinders under the flow in
high Reynolds number or even turbulence flow. the Reynolds Average Navier-Stokes
(RANS) equation, by implanting an additional turbulent eddy viscosity, coupled with
structure’s dynamic motion equation are also embedded into the loss function. The
training samples are obtained by CFD technique. The main conclusions can be
summarized as follow:
(1) PINN technique can well infer the unknown parameters (stiffness and damping
coefficient) of dynamic motion equation of cylinder based on
a very limited amount of training data, including force samples and
displacement samples (N=120). The error percentages of these parameters are
0.62% and 0.02%, respectively;
(2) PINN technique can well reconstruct the whole flow field at different times
including velocity field and pressure field only from scattered velocity
information and the pressure information is absent. The mean square errors of
flow fields reach to 103 . Furthermore, the lift force and drag force on the
cylinder can be calculated by trapezoidal law based on pressure and velocity
gradients. The mean square errors of lifting force and drag force are 0.17 104
and 0.63 104 , respectively;
(3) PINN technique has a strong applicability for solving more complicated VIV
problem, called wake-induced vibration (WIV) of cylinder behind cylinder. The
whole flow field, lift forces and drag forces on two cylinders can well inferred
by PINN;
(4) The turbulent eddy viscosity, as an important value in turbulence, is also
introduced as an unknown parameter that need to be inferred and the results
show that PINN has an effective adaptivity to obtain the  t , which means PINN
technique could have a transformative effect for modelling the turbulence
closure;
(5) The physics informed LSTM is utilized to infer the trajectories of cylinders
directly based on forces. The initial/boundary conditions and 4-th Runge-Kutta
integrations are embedded into the LSTM cell. The results demonstrate that the
trajectories of two cylinders can be well predicted only by the force dataset.
Reference

Blevins R. D. Flow-induced vibration. Krieger Pub Co, New York, second edition, 1990.

Williamson C. H. K., Govardhan R. 2004. Vortex-induced vibration. Annual Review of

Fluid Mechanics, 36(1): 413-455.

Durbin P. A. 2018. Some recent developments in turbulence closure modeling. Annual

Review of Fluid Mechanics, 50(1): 77-103.

Poroseva S. V., Colmenares J. D., Murman S. M. 2016. On the accuracy of RANS

simulations with DNS data. Physics of Fluids, 28, 115102.

Jagtab A. D., Kawaguchi K., Karniadakis G. E. 2019. Locally adaptive activation

functions with slope recovery term for deep and physics-informed neural networks.
ArXiv: 1909.12228.

Wang Y. 2017. A new concept using LSTM neural networks for dynamic system
identification. IEEE American Control Conference.

Dowell E. H. 1997. Eigen mode analysis in unsteady aerodynamics: Reduced-order

models. Applied Mechanics Review, 50(6), 371.

Schmid P. J. 2010. Dynamic mode decomposition of numerical experimental data.

Journal of Fluid Mechanics, 656(10), 5-28.

Henshawa M. J., Badcock K. J., Vio G. A. 2007. Non-linear aeroelastic prediction for
aircraft applications. Progress in Aerospace Science, 43(4-6), 65-137.

Jovanovie M. R., Schmid P. J., Nichols J. W. 2014. Sparsity-promoting dynamic mode

decomposition. Physics of Fluids, 26(2), 024103.

Hemati M. S., Williams M. O., Rowley C. W. 2014. Dynamic mode decomposition for
large and steaming datasets. Physics of Fluids, 26(11), 111701.

LeCun Y., Bengio Y., Courville A. 2015. Deep learning. Nature. 521(7553), 436-444.

Goodfellow I., Bengio Y., Courville A. Deep learning (MIT press, 2016).

Sainath T. N., Mohamed A., Kingsbury B., Ramabhadran B. Deep convolutional neural
networks for LVCSR, in 2013 IEEE International Conference on Acoustics, Speech and
Signal Processing (ICASSP) (IEEE, 2013), 8614–8618.
Ling J., Andrew, K. 2016. Reynolds averaged turbulence modelling using deep neural
networks with embedded invariance. Journal of Fluid Mechanics, 807: 155-166.

Nathan K. J. 2017. Deep learning in fluid dynamics. Journal of Fluid Mechanics, 814:
1-4.

Yeung E., Kundu S., Hodas N. 2017. Learning deep neural network representations for
Koopman operators of nonlinear dynamical systems. arXiv:1708.06850v2.

Miyanawala T. P., Jaiman R. K. 2017. An efficient deep learning technique for the
Navier-Stokes Equations: application to unsteady wake flow dynamics.
arXiv:1710.09099

Jin X., Cheng P., Chen W., Li H. 2018. Prediction model of velocity field around
circular cylinder over various Reynolds numbers by fusion convolutional neural
networks based on pressure on the cylinder. Physics of Fluids, 30, 047105.

Sekar V., Jiang Q., Shu C., Khoo B. 2019. Fast flow field prediction over airfoils using
deep learning approach. Physics of Fluids, 31, 057103.

Deng Z., Chen Y., Liu Y., Kim K. 2019. Time-resolved turbulent velocity field
reconstruction using a long short-term memory (LSTM)-based artificial intelligence
framework. Physics of Fluids, 31, 075108.

Mohan A. T., Daniel D., Chertkov M., Livescu D. 2019. Compressed convolutional
LSTM: An efficient deep learning framework to model high fidelity 3D turbulence.
arXiv:1903.00033.

Raissi M., Perdikaris P., 2017. Physics-informed neural networks: A deep learning
framework for solving forward and inverse problems involving nonlinear partial
differential equations. arXiv:1808.03398.

Tartakovsky A.M., Marrero C.O., Tartakovsky D., Barajas-Solano D. Learning

parameters and constitutive relationships with physics informed deep neural networks.
arXiv:1808.03398

Yang L., Meng X. H., Karniadakis G. 2020. B-PINN: Bayesian physics-informed neural
networks for forward and inverse PDE problems with noisy data. Journal of
Computational Physics, 425:1-23.

Acknowledgement
This work was supported by the Fundamental Research Fund for the Central
Universities of China, and the Postgraduate Research (B200203073) and Practice
Innovation Program of Jiangsu Province (KYCX20_0483).