Materials Today Communications 35 (2023) 106281

Contents lists available at ScienceDirect

Materials Today Communications

journal homepage: www.elsevier.com/locate/mtcomm

Hybrid LSTM+CNN architecture for unsteady flow prediction

Koldo Portal-Porras a, Unai Fernandez-Gamiz a, *, Ekaitz Zulueta b, Oscar Irigaray a,
Roberto Garcia-Fernandez a, c
a
Nuclear Engineering and Fluid Mechanics Department, University of the Basque Country, UPV/EHU, Nieves Cano 12, Vitoria-Gasteiz, 01006 Araba, Spain
b
System Engineering and Automation Control Department, University of the Basque Country, UPV/EHU, Nieves Cano 12, Vitoria-Gasteiz, 01006 Araba, Spain
c
Sunsundegui S.A., Polígono Ibarrea, s/n, 31800 Altsasu, Navarra, Spain

A R T I C L E I N F O A B S T R A C T

Keywords: Introduction: Data-driven methods are increasingly used for modeling fluid dynamic systems, since traditional
Unsteady flow numerical methods, such as Computational Fluid Dynamics (CFD), have certain limitations, including the
Neural networks required computational resources and user influence. There are many Deep Learning based methods capable of
LSTM
providing very accurate results for stationary problems. However, the prediction of unsteady flows remains being
CNN
a challenge, since with the addition of the time component, these methods lose reliability.
Computational fluid dynamics
Objectives: This paper aims to design a hybrid neural network for unsteady flow prediction, which combines a
Long-Short Term Memory (LSTM) and a Convolutional Neural Network (CNN).
Methods: Unsteady-state RANS-based CFD simulations are conducted to obtain data of flows around cylinders. In
these simulations different inlet velocities and cylinder diameters are considered, to ensure diversity in the
dataset. A hybrid neural network is designed, in which a LSTM predicts the Lift Coefficient for each time step and
then, based on those predictions, a CNN predicts the velocity and pressure fields. For training and testing the
proposed net the conducted CFD simulations are used.
Results: Even if there is a small mismatch between the ground-truth vortex shedding frequency and the predicted
one, the proposed network is able to accurately predict the vortex shedding behind the cylinders, with very low
errors throughout the whole studied range.

Neural Network (CNN) is the most popular for predicting flow phe­
nomena, since it provides the possibility of easily working with multi­
1. Introduction dimensional data; and as stated by Zhang et al [1]., these networks
provide the highest generalization capacity. The pioneering work of Guo
Since their introduction, CFD tools have been the most common et al [2]. demonstrates that Convolutional Neural Networks (CNNs) are
method for solving fluid dynamic problems, due to their simplicity and able to predict flow characteristics around simple geometries accurately
capability to provide detailed characteristics of the flow. However, CFD and extremely fast, reducing computational time by four orders of
tools have certain limitations, which can be prohibitive in cases where magnitude in comparison with CFD methods. Based on this work, other
the analyzed system is complex or an accurate modelling of the turbu­ authors focused their efforts on improving the predictions obtained
lence is required. The main limitation associated with CFD simulations is using the CNN. For example, Ribeiro et al [3]. evaluated the results
the required computational time and resources. In addition to this, there obtained with different encoder and decoder variants. Portal-Porras et al
are other aspects that can negatively influence the accuracy of the ob­ [4]. proposed the addition of a previous stage for pressure and vorticity
tained results, such as the influence of the user when generating the field prediction, for subsequent velocity field prediction, showing that
mesh, selecting the models and establishing the conditions of the predicting additional flow features enhances the performance of the
simulation, and many others. These aspects, added to the exponential CNN. Abucide-Armas et al [5]. proposed a data augmentation method
growth of Artificial Intelligence (AI) in the last few years and the for improving the accuracy of the CNN without requiring additional
irruption of Deep Learning (DL) techniques, have led more and more data. Since its initial implementations, CNNs have been proven to be
authors using data-driven methods for fluid dynamic problem solving. successful for many different fluid dynamics applications, such as, car
Among the different neural network structures, the Convolutional

* Corresponding author.
E-mail address: unai.fernandez@ehu.eus (U. Fernandez-Gamiz).

https://doi.org/10.1016/j.mtcomm.2023.106281
Received 20 April 2023; Received in revised form 11 May 2023; Accepted 22 May 2023
Available online 23 May 2023
2352-4928/© 2023 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
K. Portal-Porras et al. Materials Today Communications 35 (2023) 106281

Nomenclature fs Vortex shedding frequency

Φ Normalized magnitude
AI Artificial Intelligence p Pressure
CFD Computational Fluid Dynamics p Order of accuracy (Richardson Extrapolation)
CNN Convolutional Neural Network ρ Density
ConvLSTM Convolutional LSTM R Convergence condition (Richardson Extrapolation)
DL Deep Learning RE Estimated value (Richardson Extrapolation)
LSTM Long-Short Term Memory Re Reynolds number
RANS Reynolds-Averaged Navier-Stokes St Strouhal number
ReLU Rectifier Linear Unit σ Standard deviation
RNN Recurrent Neural Network u Velocity
SDF Signed Distance Function u∞ Freestream velocity
SST Shear Stress Transport μ Arithmetic mean
* Dimensionless variable v Kinematic viscosity
‘ Variable ranged between 0 and 1 x, y, z Cartesian components
D Cylinder diameter Z Zero-level set
CL Lift Coefficient

aerodynamics [6], airfoils [7], flow control device modelling [8], built
environment [9], and many others.
Nevertheless, although CNN is very accurate for steady cases, as
shown in the study conducted by Abucide-Armas et al [10]., the CNN by
itself is not able to accurately predict the temporal component in un­
steady cases. In that study, it is shown that even if in the first instants the
predictions are accurate, as the predicted time increases the error also
increases. For this reason, authors have attempted to use other neural
network architectures that are more suitable for predicting the temporal
component. For example, Mohan and Gaitonde [11] proposed a
Reduced Order Model (ROM) for turbulent flow prediction based on a
Long-Short Term Memory (LSTM), which is a especial case of a Recur­ Fig. 1. Numerical domain (not to scale).
rent Neural Network (RNN) introduced by Hochreiter and Schmidhuber
[12]. Other authors, such as Fan et al [13]. and Hou et al [14]., added
the LSTM layer to the CNN architecture in order to directly predict the
flow fields. This type of network has been used for several CFD-related
areas, such as aerodynamics [15], chemistry [16] and air pollution [17].
However, in traditional LSTM layers the states are one-dimensional,
which means that when working with higher dimensional data, such as
velocity or pressure fields, these have to be reshaped, leading to the loss
of some flow characteristics. For that reason, Shi et al [18]. introduced
the Convolutional LSTM (ConvLSTM), which allows working with
higher dimensional data, and therefore, retaining spatial information.
Mohan et al [19]. proposed a convolutional autoencoder with a
ConvLSTM to predict three-dimensional flows. Han et al [20]. designed
a encoder-decoder network with a ConvLSTM layer for unsteady flow
prediction under laminar and turbulent regimes, showing good agree­ Fig. 2. Example of the mesh generated.
ments between the results obtained with the neural network and the
ground-truth CFD data. Following this study, Han et al [21]. developed a 2. Methodology
similar network structure for unsteady flow prediction with moving
geometries. 2.1. CFD setup
In the present paper, a hybrid neural network architecture is pro­
posed for predicting the vortex shedding on the wake behind different With the objective of obtaining data for training and testing the
cylinders at various Reynolds numbers. The proposed network combines proposed network, several CFD simulations were conducted. In those
a LSTM network, for predicting the Lift Coefficient (CL ) in each time simulations, a channel is considered, with different geometries and input
step; and a CNN, for predicting the velocity and pressure fields based on velocities. For running the simulations Star-CCM+ v2019.1 [22] com­
the predictions of the LSTM. Hence, this combination allows performing mercial code was used.
predictions of unsteady-state flows by means of two simple neural The numerical domain consists of a two-dimensional
network structures. 128 mm × 256 mm plate. Left and right sides of the plate are set as
The following of the manuscript is divided as follows: Section 2 inlet and outlet, respectively; and top and bottom sides as walls with slip
provides a detailed explanation of the methodology followed to set up conditions. On the horizontal symmetry axis, there is a cylinder located
and conduct the CFD simulations, prepare the data for the neural at 50 mm from the inlet. Cylinders of different diameters, between
network, design and train the network; Section 3 shows the obtained 10 mm and 20 mm, are considered. No-slip wall conditions are set on
results, with an in-depth analysis of the main sources of error; and the cylinder. Fig. 1 provides a detailed view of the numerical domain
Section 4, summarizes the main findings of this study.

2
K. Portal-Porras et al. Materials Today Communications 35 (2023) 106281

Table 1
Mesh verification and comparison with experimental data.
Case Mesh Resolution Richardson Extrapolation Experimental

u∞ D(mm) Re Coarse Medium Fine RE p R

5 0.01 3200 0.813 0.871 0.902 0.952 2.178 0.534 0.94

10 0.02 12,800 1.062 1.131 1.158 1.17 1.805 0.391 1.18

and the boundary conditions.

With this domain, an unstructured polygonal mesh is generated, with
around 25000 cells. This mesh contains a mesh refinement in the near-
geometry region and on the wake behind the geometry, since these areas
are considered the most conflictive for carrying out the simulations.
Fig. 2 provides an example of the designed mesh.
Regarding the fluid, incompressible turbulent unsteady air is
considered. The density (ρ) of the fluid is equal to 1.18415 kg/m3, and
its dynamic viscosity (µ) is equal to 1.85508⋅10− 5 Pa⋅s. These values are
assumed to be constant. For turbulence modelling Menter’s [23]
RANS-based k-ω Shear Stress Transport (SST) turbulence model is
selected. In many studies, such as in Rajani et al [24]. and Rahman et al Fig. 3. Example of the SDF of a cylinder, the contour of the geometry is drawn
[25]., this model is proven to be appropriate for simulations similar to in white.
the ones carried out in this study. Each case was simulated for 1 s, with a
time step of 10− 4 s, which is considered to be small enough for obtaining testing the network. For training and testing the networks MATLAB
a good residual convergence and properly capturing the vortex shedding 2022b [31] commercial code, with its Deep Learning Toolbox [32] was
and the changes in the lift coefficient. Upwind scheme [26] was used to used.
discretize the convective terms, ensuring the robustness of the solution;
and a second-order temporal discretization was selected. After each time 2.2.1. Long short-term memory
step the instantaneous data required for training and testing the network Recurrent Neural Networks (RNN) are the most suitable option for
was extracted. sequential data handling. However, due to the simplicity of their
Different inlet velocities have been considered, between 5 m/s and structure, basic RNNs are not capable of predicting long-term de­
10 m/s; which means that the Reynolds number, according to Expres­ pendencies, see Yu et al [33]. For this reason, Hochreiter and Schmid­
sion (1) is between 3200 and 12,800. On the contrary, in the second huber [12] introduced the Long Short-Term Memory (LSTM). In this
group the input velocity is set at 5 m/s in all cases, so the Reynolds network, a gate is added to remember long-term significant data.
number is 3200 in all cases. In the present study, a LSTM is used to predict the Lift Coefficient
(CL ). In order to simplify and improve the LSTM training process, the
u∞ ⋅D⋅ρ
Re = (1) data of the CL has been standardized in accordance with Expression (2).
μ
CL − μCL
CL std = (2)
where D represents the diameter of the cylinder. σ CL
In order to verify sufficient mesh resolution of the meshes, the
The proposed network consists of a single LSTM cell, with 128 hid­
General Richardson Extrapolation Method [27] is applied to the mean
den units. For training the network Adam optimizer [34] was selected,
Drag Coefficient (CD ). For this, two different cases are considered, with
with a descending learning rate starting from 0.001. Training was con­
Re = 3200 and Re = 12800. This method consists of estimating the
ducted until a good training-loss convergence was achieved.
value of the analyzed parameter when the size of the cells tend to zero.
For estimating this value, a minimum of three meshes is required.
2.2.2. Convolutional neural network
Therefore, in this case a coarse mesh (of around 15,000 cells), a medium
A CNN is used to predict the velocity and pressure fields. This type of
mesh (of around 20,000 cells) and a fine mesh (the previously-explained
network has been proven to be suitable for solving different fluid dy­
mesh, around 25,000 cells) are considered. As in this case the mesh
namic problems in many studies, such as in [3,7,8,35,36].
refinement is not equal to 2, the procedure detailed by Almohammadi
Four different input layers are considered for this network: a Signed
et al. is followed [28]. As summarized in Table 1, the convergence
Distance Function (SDF) layer, to represent the geometry; an input ve­
condition (R), which should be between 0 and 1 to ensure a monotonic
locity layer, only for the cylinders, since in the other geometry-set it is
convergence, is fulfilled, and the estimated values (RE) are close to the
constant; and two layers for the CL , one for the current instant and
ones obtained with the fine mesh. Therefore, the mesh is suitable for the
another one for the previous instant. The last three layers are constant
conducted simulations. Additionally, the obtained results are compared
on the whole domain.
with the experimental ones obtained by Roshko et al [29]., showing very
The SDF layer provides the shortest distance between each cell and
close values. The same procedure for mesh validation of a similar case is
the contour of the geometry. As demonstrated by Guo et al [2]., in
followed in Aramendia et al [30].
comparison with binary representation, SDF layers improve signifi­
cantly the performance of the CNN. Fig. 3 provides an example of the
2.2. Neural networks SDF layer.
To create SDF layers, firstly the contour of the geometry is defined as
In this study a combination of two different networks is proposed, the zero-level set (Z), following Expression (1). In those cells the SDF
with the aim of predicting unsteady flow fields. On the one hand, a LSTM value is equal to zero.
is designed to predict the CL ; and on the other hand, a CNN is designed to /
Z = {(X, Y) ∈ R2 SDF(x, y) = 0} (3)
predict the velocity and pressure fields, based on the CL calculated by the
LSTM. With both networks the same dataset is considered, where 60 %
where (X, Y) are the coordinates of the cells of the geometry contour.
of the geometries is used for training, 30 % for validation and 10 % for

3
K. Portal-Porras et al. Materials Today Communications 35 (2023) 106281

Fig. 4. CNN architecture.

Then, the sign of SDF is defined depending on the location of the cell.
If the cell is on the geometry contour, SDF(x, y) = 0; if the cell is inside
the geometry, SDF(x, y) < 0; and if the cell is outside the geometry,
SDF(x, y) > 0.
Finally, after defining the sign, the SDF value of each cell is calcu­
lated by Expression (2).
SDF(x, y) = min |(x, y) − (X, Y) |⋅sign (4)
(X,Y)∈Z

Three different output layers are considered, one for each magnitude
to be predicted. To prepare these layers, the velocity and pressure fields
are first interpolated to fit into a 128 × 256 mesh, in order to adapt the
CFD data to the CNN. Subsequently, data is made dimensionless Fig. 5. Hybrid LSTM+CNN network.
following Expressions (3), (4) and (5).
and is doubled in each of the following blocks. The decoding blocks, by
ux
u∗x = (5) means of deconvolutions and Unpooling layers, perform the reverse
u∞
process of their symmetrical blocks of the encoding phase. Encoding and
uy decoding blocks are connected by concatenation layers. Fig. 4 provides a
u∗y = (6)
u∞ schematic view of the proposed CNN.
With regards to the network training progress, Adam optimizer [34]
p∗ =
p
(7) is selected, with a batch size of 64, a weight decay of 0.005 and a
ρ⋅u2∞ learning rate that starts at 0.001 and decreases as training progresses.
The validation is performed after each epoch. For training the network
where u∗x , u∗y and p∗ are dimensionless magnitudes. 200 time-steps of each case are considered, after the flow is fully
Finally, dimensionless magnitudes are normalized between 0 and 1 developed. It is estimated that this number of time-steps is enough to
according to Expression (6). This last step is also applied in the input capture several vortex shedding periods.
layers, with the purpose of speeding up and improving the training of the
network. 2.2.3. Hybrid network
Ф − Фmin In the present study a hybrid network is proposed, which combines
Фnorm = (8)
Фmax − Фmin

where Ф represents each dimensionless magnitude.

Regarding the architecture, in this study a U-Net architecture [37] is
proposed, which is a special case of an encoder-decoder network. In this
kind of networks, the input layers are compressed by means of convo­
lutions, obtaining a Latent Geometry Representation (LGR); and then,
the LGR is expanded by means of deconvolutions, getting the desired
output. This network has been successfully applied for flow field pre­
diction by Portal-Porras et al [4,8].
The proposed network consists of four encoding/decoding blocks.
Each encoding block contains two convolutional layers, the first one
followed by a ReLU (Rectifier Linear Unit) activation layer, and the
second one followed by a ReLU layer and a 2-by-2 Max Pooling layer. In
the first two blocks, the kernel size in the convolutions is equal to 5.
Additionally, strided convolutions are performed on those blocks, in
order to reduce the handled data. Conversely, in the last two blocks the Fig. 6. Vorticity contour lines in four different instants for Re = 12800. (a) t =
kernel size is equal to 3. The number of filters is set to 8 in the first block, 0.25T; (b) t = 0.5T; (c) t = 0.75T; (d) t = T.

4
K. Portal-Porras et al. Materials Today Communications 35 (2023) 106281

Fig. 7. Comparison of the CL obtained with both CFD and LSTM for a sample of the test-set.

both the LSTM and CNN networks. The LSTM predicts the lift coefficient predicting the instantaneous CL based on the previous values. As the
in each time step. Based on the predictions of the LSTM, the CNN pre­ predictions of the CNN are based on the predictions of the LSTM, a
dicts the velocity and pressure fields. As demonstrated by Portal-Porras correct CL prediction is essential. Fig. 7 shows a comparison of the
et al [4]., including a previous stage for predicting characteristics of the ground-truth CL obtained by means of CFD and the ones obtained with
flow improves the final predictions. Hence, the predictions of the ve­ the LSTM.
locity and pressure fields should improve taking the CL as input of the The results show that the LSTM is able to correctly predict the CL
CNN. Fig. 5 provides a schematic view of the proposed hybrid network. amplitude. However, a small difference between vortex-shedding fre­
quencies obtained by CFD and LSTM is noticeable.
3. Results In order to obtain a quantitative view of the obtained predictions, the
vortex shedding frequency (fs ) CL amplitude and St is evaluated. For this,
3.1. Vortex shedding the cases corresponding to the test-set are considered. These values are
provided in Table 2.
When fluid flows across a bluff body, vortices are generated on the The results show that the relative errors are low for all the analyzed
back of the body, and detach periodically from one side of the body to cases. The LSTM performs uniform predictions for amplitude among
the other, leading to the von Karman vortex street. When analyzing cases, underpredicting cases with high amplitude and overpredicting
vortex shedding the Lift Coefficient (CL ) is a very relevant parameter, those with low amplitude. Regarding the vortex shedding frequency and
since, due to the behavior of the vortices, the CL quantifies the vortex the Strouhal number, which are directly related, the LSTM predicts
shedding frequency (fs ). Fig. 6 shows the vorticity contour and instan­ lower frequencies for low-Re cases, and higher frequencies for cases with
taneous CL at four different moments of a single vortex shedding period higher Re. Nonetheless, the errors can be attributed to the fact of
for the case Re = 12800. In this Figure the generation, convection and considering different input velocities, since the diversity among the
diffusion of the vortexes can be clearly observed. samples can make the training process more complicated.
For oscillating flow analysis, such as vortex shedding, the Strouhal
number (St) is a commonly used parameter. The Strouhal number is a
dimensionless number, which represents the ratio between the inertial 3.3. Velocity and pressure field prediction
forces caused by the unsteadiness of the flow or the local acceleration,
and the inertial forces caused by the changes in velocity from one point Velocity and pressure fields obtained with the CNN are compared
of the flow field to another. St is calculated according to Expression (9). with the ground-truth values obtained by means of CFD. For this com­
parison, four different instants are considered, with the vortex moving
fs ⋅D
St = (9) upward, the vortex at the top, the vortex moving downward and the
u∞
vortex at the bottom. Fig. 8 provides the mentioned fields at each
moment.
3.2. Lift coefficient prediction The results show that the CNN is able to accurately predict the ve­
locity and pressure fields, modeling the vortex shedding behind the
As mentioned, the recurrent part of the proposed network consists of geometries. In all the cases, the larger errors appear on the contour of the

Table 2
Comparison of the CL and St obtained with both CFD and LSTM.
Case CFD LSTM Relative error

Re u_∞ [m/s] D [mm] CL amplitude fs [Hz] St CL amplitude fs [Hz] St CL amplitude St

4400 5 14 1.6659 77.2 0.216 1.4952 72.9 0.204 10.2 % 5.5 %

4400 7 10 1.3676 150.5 0.215 1.4108 139.3 0.199 3.16 % 7.4 %
6200 8 12 1.3685 137.8 0.207 1.41 144.2 0.216 3.03 % 4.3 %
9200 9 16 1.4767 113.1 0.201 1.4105 119.7 0.213 4.48 % 5.9 %

5
K. Portal-Porras et al. Materials Today Communications 35 (2023) 106281

Fig. 8. Comparison of the velocity and pressure fields obtained with both CFD and CNN+LSTM in four different instants around a cylinder. (a) Upward vortex; (b)
Vortex at top; (c) Downward vortex; (d) Vortex at bottom.,.

geometry and on the wake behind the cylinder. These areas are assumed experiences in that area, the network struggles to make accurate pre­
to be the most troubleshooting ones. The boundary layer is on the dictions. With respect to the wake behind the cylinder, this zone is the
contour of the geometry. Because of the large velocity changes the flow one in which most differences appear between the cases analyzed. For

6
K. Portal-Porras et al. Materials Today Communications 35 (2023) 106281

Fig. 8. (continued).

this reason, it is the area in which the network has more difficulties to mismatch between the CFD and CNN results, especially as the pre­
properly learn all the characteristics of the flow. However, these errors dictions become longer in time. Fig. 9 shows an instant where the CL
are very low in all the analyzed range. obtained with CFD and the one obtained with the neural network
Nevertheless, as the results obtained with the LSTM are used as input mismatch, causing the vortices to be generated in opposite directions. In
in the CNN and the LSTM predicts the Strouhal number with a small the analyzed case, the ground-truth CFD CL is in its maximum value,
error, as shown in Fig. 7, there are some instances where there is a while the CL predicted by the LSTM is in its minimum value. However,

7
K. Portal-Porras et al. Materials Today Communications 35 (2023) 106281

Fig. 9. Velocity and pressure fields around a cylinder at an instant with phase mismatch between CFD and CNN+LSTM results.

despite the difference in frequency, the evolution of the vortices is the 4. Conclusions
same throughout the whole analyzed time.
Introduction: Traditional numerical methods for fluid dynamics
3.4. Error analysis system analysis have some limitations, such as the required computa­
tional resources and the influence of the user for defining the case. For
In order to obtain a detailed view of the error along a vortex shedding that reason, data-driven methods, especially Deep Learning-based
period, the error in a period where the CL obtained with CFD and LSTM methods, are becoming increasingly popular. This kind of methods
match is analyzed for all the test-set cases. Fig. 10 provides the mean have been proven to be very effective and accurate for steady-state
error evolution of all the analyzed magnitudes throughout a vortex problems. However, the prediction of unsteady flows remains being a
shedding period for all the cases of the test-set. challenge, since with the addition of the time component, these methods
The mean error evolution of velocity fields show that the larger er­ lose reliability.
rors appear when the vortex is located at the top or at the bottom. In the Objectives: The objective of this study is to propose a combination of
case of pressure, the results are more diverse. In this case the error does neural networks for the prediction of unsteady flows. The proposed
not follow a clear trend with the position of the vortex, so it depends on network combines a LSTM network, for predicting the CL in each time
the sample. In all the analyzed cases, the mean errors show a direct step; and a CNN, for predicting the velocity and pressure fields based on
correlation with the error in CL amplitude explained in Section 3.2. As the predictions of the LSTM.
CNN predictions are made taking the LSTM predictions as input, larger Methods: In order to collect data of unsteady flows, different vortex
errors in CL lead to larger errors in the predicted velocity and pressure shedding cases are simulated. Unsteady-state RANS-based CFD simula­
fields. tions of different flows around cylinders are conducted. In those simu­
lations, various inlet velocities and cylinder diameters are considered to
3.5. Performance comparison ensure a diverse dataset. The collected CFD simulation data is used for
training and testing the proposed neural network.
The main advantage of data-driven methods over traditional nu­ Results: The LSTM is able to accurately predict the CL throughout the
merical methods is the speed with which results can be obtained. For whole analyzed range of time. The CNN, which takes as input the CL
that reason, the computational time required by each method is predicted by the LSTM, is also able to accurately predict the velocity and
compared. Table 3 shows the computational time required by each pressure fields. Nevertheless, the LSTM predicts the CL with a slight error
method to perform a single simulation using a single Intel Xeon Gold on the frequency. This leads to a mismatch between the position of the
5120 CPU core. vortexes predicted by the CNN and the ones obtained by CFD at
As shown in the comparison, CFD methods require an average of advanced time-steps. However, despite the error on the vortex shedding
around 121 h to perform a single simulation of 1 s, while the proposed frequency, the CNN is able to correctly predict the evolution of the
LSTM+CNN network only requires about 37 min to predict all the time- vortexes.
steps of the same simulation. Therefore, the proposed network is 192.4 The error analysis shows that the most troubleshooting areas are the
times faster than CFD. Around 30 min were required to train the LSTM, contour of the cylinder and the wake behind the cylinder, where the
and around 12 h to train the CNN. vortexes appear. For velocity fields the larger mean errors appear when
the vortex is located at the top or at the bottom, while for pressure no
clear trend can be observed. The error in the CNN predictions is directly
related to the error in the CL predictions of the LSTM.

8
K. Portal-Porras et al. Materials Today Communications 35 (2023) 106281

Fig. 10. Mean error of the dimensionless magnitudes during a vortex shedding period. (a) u∗x error; (b) u∗y error; (c) p∗ error.

9
K. Portal-Porras et al. Materials Today Communications 35 (2023) 106281

Table 3 [9] G. Calzolari, W. Liu, Deep learning to replace, improve, or aid CFD analysis in built
Computational time required by each method to obtain the predictions of the lift environment applications: a review, Build. Environ. 206 (2021), 108315, https://
doi.org/10.1016/j.buildenv.2021.108315.
coefficient, velocity and pressure fields. [10] A. Abucide-Armas, K. Portal-Porras, U. Fernandez-Gamiz, E. Zulueta, A. Teso-Fz-
Method Time [s] Speedup Betoño, Convolutional neural network predictions for unsteady Reynolds-averaged
navier–stokes-based numerical simulations, J. Mar. Sci. Eng. 11 (2023) 239,
CFD 435,543.74 s - https://doi.org/10.3390/jmse11020239.
LSTM+CNN 2263.96 s 192.4 [11] Mohan, A.T.; Gaitonde, D.V. , A deep learning based approach to reduced order
modeling for turbulent flow control using LSTM neural networks, ArXiv180409269
Phys., 2018.
With regards to the computational resources, the proposed model [12] S. Hochreiter, J. Schmidhuber, Long short-term memory, Neural Comput. 9 (1997)
1735–1780, https://doi.org/10.1162/neco.1997.9.8.1735.
clearly outperforms the CFD simulations, reducing 192.4 times the [13] Fan, S.; Fei, J.; Guo, X.-W.; Yang, C.; Revell, A. , CNN+LSTM accelerated turbulent
computational time required. flow simulation with link-wise artificial compressibility method, in: Proceedings of
the Fiftieth International Conference on Parallel Processing, ACM, Lemont IL USA,
August 9, 2021, 2021, 1–10.
Funding [14] Y. Hou, H. Li, H. Chen, W. Wei, J. Wang, Y. Huang, A novel deep U-Net-LSTM
framework for time-sequenced hydrodynamics prediction of the SUBOFF AFF-8,
Eng. Appl. Comput. Fluid Mech. 16 (2022) 630–645, https://doi.org/10.1080/
The authors are thankful to the Government of the Basque Country 19942060.2022.2030802.
for the financial support of ELKARTEK21/10 KK-2021/00014 and [15] T. Li, T. Wu, Z. Liu, Nonlinear unsteady bridge aerodynamics: reduced-order
ELKARTEK20/78 KK-2020/00114 research programs, respectively. modeling based on deep LSTM networks, J. Wind Eng. Ind. Aerodyn. 198 (2020),
104116, https://doi.org/10.1016/j.jweia.2020.104116.
[16] H. Kim, M. Park, C.W. Kim, D. Shin, Source localization for hazardous material
CRediT authorship contribution statement release in an outdoor chemical plant via a combination of LSTM-RNN and CFD
simulation, Comput. Chem. Eng. 125 (2019) 476–489, https://doi.org/10.1016/j.
compchemeng.2019.03.012.
Conceptualization, K.P.-P. and U.F.-G.; Methodology, E.Z.; Soft­ [17] Quilodrán-Casas, C.; Arcucci, R.; Mottet, L.; Guo, Y., C. Pain, Adversarial
ware, E.Z. and R.G-F.; Validation, K.P.-P. and U.F.G.; Formal analysis, Autoencoders and Adversarial LSTM for Improved Forecasts of Urban Air Pollution
U.F-G. and O.I.; Investigation, K.P.-P.; Resources, U.F.-G.; Data cura­ Simulations, 2021.
[18] Shi, X.; Chen, Z.; Wang, H.; Yeung, D.-Y.; Wong, W.; WOO, W. , Convolutional
tion, K.P.-P.; Writing – original draft, K.P.-P.; Writing – review & LSTM network: a machine learning approach for precipitation nowcasting, in:
editing, E.Z. and O.I.; Visualization, E.Z. and R.G-F.; Supervision, U.F.- Proceedings of the Advances in Neural Information Processing Systems, Curran
G.; Project administration, E.Z. and R.G-F.; Funding acquisition, U.F.- Associates, Inc., 28, 2015.
[19] Mohan, A.; Daniel, D.; Chertkov, M.; Livescu, D. , Compressed convolutional LSTM:
G. All authors have read and agreed to the published version of the an efficient deep learning framework to model high fidelity 3D turbulence,
manuscript. ArXiv190300033 Nlin Physicsphysics, 2019.
[20] R. Han, Y. Wang, Y. Zhang, G. Chen, A new prediction method of unsteady wake
flow by the hybrid deep neural network, Phys. Fluids 31 (2019), 127101, https://
doi.org/10.1063/1.5127247.
Declaration of Competing Interest [21] R.-K. Han, Z. Zhang, Y.-X. Wang, Z.-Y. Liu, Y. Zhang, G. Chen, Hybrid deep neural
network based prediction method for unsteady flows with moving boundary, Acta
Mech. Sin. 37 (2021) 1557–1566, https://doi.org/10.1007/s10409-021-01129-4.
The authors declare that they have no known competing financial [22] STAR-CCM+ V2019.1. https://www.plm.automation.siemens.com/. (Accessed 2
interests or personal relationships that could have appeared to influence June 2020). 2020.
the work reported in this paper. [23] Menter, F. , Zonal two equation K-w turbulence models for aerodynamic flows, in:
Proceedings of the Twenty Third Fluid Dynamics, Plasmadynamics, and Lasers
Conference; American Institute of Aeronautics and Astronautics, Orlando,FL,U.S.
Data availability A., July 6, 1993. 1993.
[24] B.N. Rajani, A. Kandasamy, S. Majumdar, Numerical simulation of laminar flow
Data will be made available on request. past a circular cylinder, Appl. Math. Model. 33 (2009) 1228–1247, https://doi.org/
10.1016/j.apm.2008.01.017.
[25] M.M. Rahman, Md.M. Karim, M.A. Alim, Numerical investigation of unsteady flow
Acknowledgements past a circular cylinder using 2-D finite volume method, J. Nav. Archit. Mar. Eng. 4
(1970) 27–42, https://doi.org/10.3329/jname.v4i1.914.
[26] S. Osher, S. Chakravarthy, Upwind schemes and boundary conditions with
The authors are grateful for the support provided by the SGIker of applications to Euler equations in general geometries, J. Comput. Phys. 50 (1983)
UPV/EHU. 447–481, https://doi.org/10.1016/0021-9991(83)90106-7.
[27] Richardson, L.F.; Gaunt, J.A. , The deferred approach to the limit, Philos. Trans. R.
Soc. Lond. Ser. Contain. Pap. Math. Phys. Charact., 1927, 226, 299–361. doi:
References 10.1098/rsta.1927.0008.
[28] K.M. Almohammadi, D.B. Ingham, L. Ma, M. Pourkashan, Computational fluid
dynamics (CFD) mesh independency techniques for a straight blade vertical axis
[1] Zhang, Y.; Sung, W.-J.; Mavris, D. ,Application of convolutional neural network to
wind turbine, Energy 58 (2013) 483–493, https://doi.org/10.1016/j.
predict airfoil lift coefficient, ArXiv171210082 Cs Stat, 2018.
energy.2013.06.012.
[2] Guo, X.; Li, W.; Iorio, F. ,Convolutional neural networks for steady flow
[29] Roshko, A. , Vortex shedding from circular cylinder at low Reynolds number, in:
approximation, in: Proceedings of the Twenty Second ACM SIGKDD International
Proceedings of the Cambridge University Press, Cambridge, UK, 1954.
Conference on Knowledge Discovery and Data Mining, ACM, San Francisco
[30] I. Aramendia, U. Fernandez-Gamiz, E. Zulueta Guerrero, J. Lopez-Guede,
California USA, August 13, 2016, 2016, 481–490.
J. Sancho, Power control optimization of an underwater piezoelectric energy
[3] Ribeiro, M.D.; Rehman, A.; Ahmed, S.; Dengel, A. , DeepCFD: efficient steady-state
harvester, Appl. Sci. 8 (2018) 389, https://doi.org/10.3390/app8030389.
laminar flow approximation with deep convolutional neural networks,
[31] MATLAB .https://es.mathworks.com/products/matlab.html. (Accessed 9 June
ArXiv200408826 Phys., 2020.
2021). 2021.
[4] K. Portal-Porras, U. Fernandez-Gamiz, A. Ugarte-Anero, E. Zulueta, A. Zulueta,
[32] Deep Learning Toolbox. https://es.mathworks.com/products/deep-learning.html.
Alternative artificial neural network structures for turbulent flow velocity field
(Accessed 3 July 2021). 2021.
prediction, Mathematics 9 (2021) 1939, https://doi.org/10.3390/math9161939.
[33] Y. Yu, X. Si, C. Hu, J. Zhang, A review of recurrent neural networks: LSTM cells and
[5] A. Abucide-Armas, K. Portal-Porras, U. Fernandez-Gamiz, E. Zulueta, A. Teso-Fz-
network architectures, Neural Comput. 31 (2019) 1235–1270, https://doi.org/
Betoño, A data augmentation-based technique for deep learning applied to CFD
10.1162/neco_a_01199.
simulations, Mathematics 9 (2021) 1843, https://doi.org/10.3390/math9161843.
[34] Kingma, D.P.; Ba, J. , Adam: a method for stochastic optimization, ArXiv14126980
[6] S.J. Jacob, M. Mrosek, C. Othmer, H. Köstler, Deep learning for real-time
Cs, 2017.
aerodynamic evaluations of arbitrary vehicle shapes, SAE Int. J. Passeng. Veh. Syst.
[35] L. Liu, A.K. Padthe, P.P. Friedmann, Computational study of microflaps with
15 (2022) 15–15-02–0006, https://doi.org/10.4271/15-15-02-0006.
application to vibration reduction in helicopter rotors, AIAA J. 49 (2011)
[7] N. Thuerey, K. Weißenow, L. Prantl, X. Hu, Deep learning methods for reynolds-
1450–1465, https://doi.org/10.2514/1.J050829.
averaged navier–stokes simulations of airfoil flows, AIAA J. 58 (2020) 25–36,
[36] L. Deng, Y. Wang, Y. Liu, F. Wang, S. Li, J. Liu, A CNN-based vortex identification
https://doi.org/10.2514/1.J058291.
method, J. Vis. 22 (2019) 65–78, https://doi.org/10.1007/s12650-018-0523-1.
[8] K. Portal-Porras, U. Fernandez-Gamiz, E. Zulueta, A. Ballesteros-Coll, A. Zulueta,
[37] Ronneberger, O.; Fischer, P.; Brox, T. U-Net: convolutional networks for
CNN-based flow control device modelling on aerodynamic airfoils, Sci. Rep. 12
biomedical image segmentation, in: N. Navab, J. Hornegger, W.M. Wells, A.F.
(2022) 8205, https://doi.org/10.1038/s41598-022-12157-w.

10
K. Portal-Porras et al. Materials Today Communications 35 (2023) 106281

Frangi (Eds.), Proceedings of the Medical Image Computing and Computer-Assisted Intervention – MICCAI, 2015, 2015, Springer International Publishing, Cham,
234–241.

11