Computers and Fluids 236 (2022) 105312

Contents lists available at ScienceDirect

Computers and Fluids

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

A deep learning approach for the transonic flow field predictions around
airfoils
Cihat Duru a ,∗, Hande Alemdar b , Ozgur Ugras Baran a
a Department of Mechanical Engineering, Middle East Technical University, Ankara, Turkey
b
Department of Computer Engineering, Middle East Technical University, Ankara, Turkey

ARTICLE INFO ABSTRACT

Keywords: Learning from data offers new opportunities for developing computational methods in research fields, such
Deep learning as fluid dynamics, which constantly accumulate a large amount of data. This study presents a deep learning
Transonic flow approach for the transonic flow field predictions around airfoils. The physics of transonic flow is integrated
Airfoil aerodynamics
into the neural network model by utilizing Reynolds-averaged Navier–Stokes (RANS) simulations. A detailed
Shock waves
investigation on the performance of the model is made both qualitatively and quantitatively. The flow features
associated with the transonic effects and the angle of attack variation, such as the shock waves and the flow
separation, are well predicted. Furthermore, predicted flowfield data are used to compute the aerodynamic
coefficients. The findings indicate that the presented model may allow avoiding time-consuming computational
fluid dynamics (CFD) simulations, especially in the design optimization studies with a slight loss of accuracy.

1. Introduction hand, computational fluid dynamics (CFD) simulations require signifi-

cant computational power, and they are not feasible for experimenting
The design of airfoils is a lengthy process in which calls for ex- with many airfoils.
perimental, computational, and theoretical efforts. Various families of Recent advances in machine learning show promise for a fast alter-
airfoils [1] have been developed for different applications, e.g., aircraft, native approach instead of expensive experiments and time-consuming
sailplanes, propellers. The starting point of the airfoil selection for simulations. Learning from data offers new opportunities for developing
a new application design such as a wing, rotorcraft, wind turbine,
computational methods in research fields, such as fluid dynamics,
and others is usually selecting the best appropriate airfoil from those
which constantly accumulate a large amount of data. Therefore, a
families. This decision is made by considering the mission requirements
machine learning approach may be utilized to relieve the massive
based on, but not limited to, flight regime, lift requirements, stall limits,
drag limitations, structural limits and maneuvering [2]. However, it amount of work from CFD simulations, at least in the initial design
might be challenging to fulfill all the requirements by selecting an stages of airfoils. Several works have been carried out and showed
available airfoil since those requirements are strongly dependent on whether and to what extent machine learning methods can contribute
the airfoil geometry, Reynolds number, Mach number and the angle of to aerodynamic studies. Faller and Schreck [4] have addressed possible
attack. The airfoil shape has an exhaustive list of shape parameters such aeronautics problems in which neural networks can provide a practical
as thickness, camber, location of maximum thickness and maximum approach. Initially, neural networks have been applied to inverse airfoil
camber and the leading edge radius. All these parameters must be taken design problems [5,6] and aerodynamic coefficient prediction [7,8].
into account [3] since they determine the airfoil performance. There- However, as computing resources have been developed and new neural
fore, the design engineer should experiment with many configurations network algorithms have also been introduced, such as convolutional
to find an optimum airfoil design that meets all the requirements. As a neural networks (CNN) [9], more complex network models have been
part of the solution and selection procedure, wind tunnel tests might be generated. These new developments have provoked the investigation of
costly and often time-consuming, especially in the early design stages.
the prediction capability of network models for flow field prediction.
It is preferable to conduct experiments at later stages of the design in
Several studies have been conducted to predict flow fields over
order to fully and accurately assess the aerodynamic performance of the
airfoils and various objects using deep learning. A CNN model was
aircraft performance. Recent numerical solution techniques allow very
proposed by Jin et al. [10] that predicts velocity fields around a circular
accurate results for both viscous and compressible flows. On the other

∗ Corresponding author.
E-mail address: [email protected] (C. Duru).

https://doi.org/10.1016/j.compfluid.2022.105312
Received 15 July 2021; Received in revised form 12 December 2021; Accepted 5 January 2022
Available online 20 January 2022
0045-7930/© 2022 Elsevier Ltd. All rights reserved.
C. Duru et al. Computers and Fluids 236 (2022) 105312

cylinder for low Reynolds number flow. Fukami et al. [11] developed a or shock conditions, either. The neural network model is trained with a
machine learning approach to reconstruct a high-resolution flow field database of steady, compressible flow solutions of 204 different airfoil
from low-resolution input images for laminar and turbulent flows. Vi- shapes for a wide range of angles of attack from −10◦ to 20◦ at a free-
querat and Hachem [12] presented a CNN model for the drag prediction stream Mach number, 𝑀 = 0.7 and free-stream Reynolds number, 𝑅𝑒 =
of arbitrary 2D shapes at Re = 10. Prediction of unsteady flow fields 6 × 106 . Contrary to previous studies, we feed each airfoil shape into
around a circular cylinder was studied using several models based on the neural network model by rotating the airfoil around the imposed
generative adversarial networks and CNNs by Lee and You [13]. We center of rotation with the corresponding angle of attack in our study.
use a discriminative approach instead of a generative one in this study. Therefore, each angle of attack cases is considered as a new shape in the
Also, the above-mentioned studies do not consider airfoils. Chen et al. flow field. Then, we evaluate the performance of our model with unseen
[14] studied to predict velocity and pressure fields around arbitrary airfoil shapes during the training process. The new model can predict
2D shapes, including airfoils in laminar flow, using a CNN-based U-net an extensive range of flow conditions around the airfoils from very
architecture. They only present results for four hand-picked airfoils. streamlined flows to flows involving separations, shocks and combined
In one of the first studies considering only airfoils, Guo et al. [15] stall-shock conditions. Our model provides accurate answers in three
developed a CNN model to predict velocity fields over 2D and 3D orders faster than CFD simulations with high reliability regardless of
domains. They achieved significant speedup with their neural network the airfoil shape.
model compared to CFD simulations. Singh et al. [16] developed a Our study serves a dual purpose. First, we provide insight into
machine learning-based methodology that assists the Spalart–Allmaras the transonic flow. Second, we present a proof of concept for the
turbulence model. They have utilized their augmented model to predict implementation of deep learning methodology into fluid flow problems,
aerodynamic coefficients. They do not use CNN, and they present including discontinuities such as shock waves and flow separation. The
results only on three different airfoils. A CNN model was presented to selection of Mach number is important in this regard. At this Mach
predict discretized pressure coefficients on sampling points of airfoil number, flow is mostly shock and stall-free at low angle of attack
surface by Yılmaz and German [17]. Their study reports a validation conditions. Then, for some airfoils, shocks are formed at either side
accuracy of 80%, and they consider zero-angle of attack only. Zhang of the airfoil by changing the attack angle. Stall conditions are met at
et al. [18] used both CNN and MLP based models to predict only the some flow angles, with varying vortex sizes. Moreover, Shock-Wave-
lift coefficient of airfoils for diverse flow conditions. In this study, we Boundary Layer Interaction can be observed in some cases. Therefore,
present a method that can estimate the lift, drag and pitching moment. this Mach number is more challenging than incompressible or purely
A convolutional encoder–decoder model similar to [15] was de- transonic cases in terms of Machine Learning performance since the
veloped by Bhatnagar et al. [19] for the prediction of pressure and variation of the flow field between configurations is extensive.
velocity fields around airfoils. They experimented with a separated The rest of the paper is organized as follows. The structure of
decoder approach for the output fields as well as a shared one. They the neural network model and the data preparation procedure are
conclude that the shared encoder–decoder is computationally more described in detail in Section 2. The neural network model training
feasible. They use only three different airfoil shapes in their study. In and the discussion of the results are presented in Section 3 with
this study, we present a deeper network and obtain better results in qualitative and quantitative analysis. Finally, conclusions and prospects
terms of error. Sekar et al. [20] presented an approach to predict only are presented in Section 4.
laminar flow fields around airfoils. In this study, CNN was applied to
parameterize airfoil shapes, while multilayer perceptron was utilized 2. Methodology
for the approximation model.
A U-net-based network architecture was applied to predict velocity 2.1. Neural network modeling
and pressure fields around airfoils for incompressible flow by Thuerey
et al. [21]. Their model takes a 3-channel input. One channel is the A convolutional encoder–decoder neural network model (CNN-
binary representation of the airfoil mask. The other two are the 𝑥 FOIL), which the authors have previously proposed [24], is imple-
and 𝑦 component of the velocity under freestream conditions, respec- mented on PyTorch [25] to realize this study. The architecture of the
tively, with a zero velocity inside the airfoil shape. In their model, neural network is illustrated in Fig. 1. The model takes the airfoil shape
the inputs contain highly redundant information, and they encode the (with distance map as the shape descriptor) as input and provides the
airfoil shape in all three input fields. This makes the network learn pressure field and the Mach number field around the corresponding
more model parameters than it actually needs. In our model, we only airfoil shape. The model aims to build a mapping function between the
provide the airfoil shape in a single channel using the distance field input and the output, such as 𝑓 (𝐗) = 𝐘, where 𝐗 is a set of 𝑁 points
representation, thus eliminating redundancies. Hui et al. [22] utilized on the field where each point represents the minimum distance to the
deep learning in order to predict just the surface pressure distribution airfoil shape. 𝐘 denotes {𝐶𝑝 , 𝑀} where 𝐶𝑝 is the pressure coefficient
over airfoils. field, and 𝑀 is the Mach number field that corresponds to the input
Wang et al. [23] use a variational autoencoder together with MLP airfoil shape representation.
to predict the flow fields of a given airfoil shape. MLP is used to The model consists of eight convolutional blocks and eight trans-
encode the airfoil shapes, and the variational decoder generates the posed convolution blocks in the encoder and decoder parts, respec-
flow fields. The flow fields of all the airfoils are evaluated at a fixed lift tively, as shown in Fig. 1. The encoder and the decoder parts are shared
coefficient of 0.7 with a Mach number of 0.76 and a Reynolds number for both tasks, i.e., Mach number and pressure field predictions, rather
of 𝑅𝑒 = 5 × 106 . The study considers only supercritical airfoils, and the than two separate encoder and/or decoder. Thus, we benefit from
results are evaluated on the same test dataset that is used for fine-tuning shared-encoding and decoding with reducing the number of learnable
the hyperparameters. In this study, we cover a more comprehensive parameters and saving memory size.
range of airfoils in terms of airfoil geometry, and we present results on The input of the network is the airfoil geometry. The airfoil geom-
airfoil shapes that are out of our original dataset as well. etry is represented by the distance field map of itself. The distance
In this study, we propose a very comprehensive solution for CNN- field (DF) proved itself as a successful shape learning approach for
based predictions of flow around airfoils. The new model is extended neural network training [15,19,22]. The distance field allows compen-
from the previous study [24] in a few but significant areas. The new sating the model errors and resolution related inaccuracies. The airfoil
model allows airfoil predictions in very challenging flow conditions, shapes are constructed using the coordinates provided by the UIUC
including very high angles of attack, stall conditions and both weak database [26], as explained in the next section. All shapes are placed
and strong shocks. The new model is not trained for the isolated stall on the 256 × 256 Cartesian grid, and the distance fields are constructed

2
C. Duru et al. Computers and Fluids 236 (2022) 105312

Fig. 1. CNNFOIL architecture.

by measuring the Euclidean distance of each grid to the nearest point 2.2. Data preparation
on the airfoil boundary, as defined by Eq. (1). The inner elements of
the airfoil are assigned to zero. The airfoil shapes are generated with great attention using coordi-
|(𝑖, 𝑗) − (𝑖∗ , 𝑗 ∗ )| nates provided from the UIUC database [26] and also airfoil profile
𝐷𝐹 (𝑖, 𝑗) = ∗min | | (1)
∗
(𝑖 ,𝑗 )∈𝛺 formulas when available. The airfoil shapes are not selected from a
Here, (𝑖, 𝑗) is the set of points in the Cartesian grid, and 𝛺 denotes particular family or considering the performance requirement. Besides,
the airfoil boundary. The distance field for an arbitrary airfoil shape the selected airfoils are not necessarily designed for the transonic flow
is shown in Fig. 1. regime. The airfoil shapes are chosen randomly since this study is
The encoder part of the network architecture generates a latent interested in the variation of geometrical characteristics of an air-
representation from the distance field of the airfoil shapes. It has eight foil shape. All airfoil curves are constructed with splines in order to
blocks each of which is composed of a convolutional layer, batch nor- prevent local errors. Then, surface shapes and the flow simulation
malization layer and an activation layer. Convolutional layers perform results are controlled carefully to ensure high-quality simulations. The
a particular type of linear operation called the convolution for feature dataset of the flow fields around 204 different airfoils for the range
extraction. A matrix of weights, called the convolutional filter, slides of angles of attack from −10◦ to 20◦ at 𝑀 = 0.7 and 𝑅𝑒 = 6 × 106
over the input tensor and performs an element-wise product. Then, is obtained through CFD simulations. A finite-volume in-house CFD
the sum of the element-wise product is assigned to the output tensor, solver is employed to simulate the flow field around airfoils. We have
which is called the feature map. An arbitrary number of feature maps solved compressible Reynolds-averaged Navier–Stokes (RANS) equa-
can be extracted by repeating the same procedure with different filters tions. The viscous flow solution is necessary to resolve flow separation
in order to obtain different characteristics of the input tensor. The at stall conditions. We have utilized the Spalart–Allmaras turbulence
filter behaves as a feature extractor and highlights specific features. model [29] for this purpose. A RANS-based solution for turbulence
The neural network learns the weights in the filter matrices during the and flow separation is sufficient for our neural network model since
training process. The numbers of feature maps after each convolutional the high Reynolds number indicates less prominent viscous forces. For
layer are indicated in Fig. 1. Convolutions are performed with a 2 × 2 complete viscous simulation, more advanced methods and simulations
filter size and stride of 2, which results in a 1 × 1 latent representation at different Reynolds numbers might be necessary. We have employed
at the end of the encoder stage. Each convolutional layer is followed a very accurate second-order HLLC flux scheme [30] for inviscid fluxes,
by a batch normalization layer in order to stabilize the learning process which is an approximate Riemann solver. The second-order accuracy is
and accelerate the training of the model by reducing the number of achieved by the application of the Venkatakrishnan limiter. Venkatakr-
training epochs [27]. Also, a nonlinear activation function is applied ishnan is an adjustable limiter, where a parameter allows to adjustment
after the batch normalization layer. Except for the last layer prior to of the limiter for the shock sharpness or robustness. We have set the
the decoding stage, the exponential linear unit (ELU) [28] is used as adjustment parameter to ensure robust solutions with good fidelity
the nonlinear activation function. ELU is defined as follows: around the shocks.
{
𝜙𝜙 > 0 A validation study has been carried out for 𝑀 = 0.734, 𝑅𝑒 = 6.5×106
𝛹 (𝜙) = (2)
0.1(𝑒𝜙 − 1) 𝜙 ≤ 0 and angle of attack of 𝛼 = 2.54◦ flow over RAE 2822 airfoil. The
Each feature map of the previous layer is computed as results are compared with the experimental results [31]. It is important
to note that the flow parameters reported in [31] have been updated
𝐲𝑖 = 𝛹 (𝐖𝑖 ∗ 𝐱𝑖−1 + 𝐛𝑖 ) (3) later by wind tunnel corrections applied in the EUROVAL validation
where 𝐱 is the input feature map. At the very beginning of the feature project [32]. Fig. 2 shows the pressure coefficient 𝐶𝑝 distribution over
extraction process, 𝐱0 corresponds to the distance field denoted as 𝐗. the airfoil surface. The numerical results agree well with the experi-
𝐖 and 𝐛 are the weight matrix (convolutional filter) and the additive mental results given in [31]. Please observe that the sharp gradients in
bias, respectively. 𝐲 is the output feature map. Fig. 2 and the results of Section 3.3 prove our careful selection of the
Symmetrically, the decoder part has eight blocks consisting of the flux scheme, limiter and the mesh.
transposed convolutional layers, which are applied to the latent rep- All numerical simulations are conducted on very high-quality and
resentations extracted from the encoder part to construct the model high-orthogonality structured meshes. Several structured mesh domains
output, 𝐘 corresponding to the input, 𝐗. First, feature maps are up- involving different multi-block mesh topologies, external domain sizes,
sampled to be adapted for convolution operations. Then, convolution and grid density distributions are investigated. Corresponding mesh
operations described before are performed using 2 × 2 filter size and independence and solution assessment studies are completed. Finally, a
stride of 2, which result in 256 × 256 flow fields at the end of the standard and proven computational grid topology to be shared among
decoder stage. Similarly, the exponential linear unit (ELU) without the all airfoils is decided. The meshing algorithm generates a single block
batch normalization is also applied after each transposed convolution O-topology grid with a predetermined number of cells in both chord-
excluding the last layer. A single decoder is used to generate 𝐶𝑝 and 𝑀 wise and normal directions around the airfoils in the database. The
fields. cell spacing in the normal direction is chosen appropriately in order

3
C. Duru et al. Computers and Fluids 236 (2022) 105312

Fig. 2. 𝐶𝑝 distribution over RAE 2822 airfoil surface.

Fig. 3. Close-up view of the computational domain.

to ensure that 𝑦+ is below 1. The farfield is fixed at about 500 times the 905 data are used for the test. The pressure coefficient, 𝐶𝑝 , is computed
chord length in order to minimize the effect of boundary conditions as
( )
on calculated flow variables [33]. Fig. 3 shows the general view and 2 𝑝
details of the computational domain near the body. 𝐶𝑝 = −1 (4)
2
𝛾𝑀∞ 𝑝∞
In this equation, the heat capacity ratio is taken for air as 𝛾 = 1.4. The
2.3. Training of the model
free-stream Mach number is selected as 𝑀∞ = 0.7. Variables 𝑝 and 𝑝∞
are static pressures at the point the pressure coefficient is calculated
Pressure and Mach number field data are collected for 204 airfoil
and at the free stream, respectively. We also standardize data sets by
shapes and angle of attack range from −10◦ to 20◦ at 𝑀 = 0.7. The data
dividing each flow variable by its standard deviation combined with
of 6324 cases are divided randomly into 85% for training and validation mean-centering as expressed in Eq. (5) with respect to training set.
sets and 15% unseen data during the training for the test set. Therefore,
𝑥−𝜇
4614 data are used for training, 805 data are used for validation, and 𝑥𝑛 = (5)
𝜎

4
C. Duru et al. Computers and Fluids 236 (2022) 105312

Fig. 4. Training history of the network.

Fig. 5. The comparison of the 𝐶𝑝 and 𝑀 fields of Eppler 547 for different angles of attack. (a) 𝛼 = −8◦ , (b) 𝛼 = −4◦ , (c) 𝛼 = 0◦ , (d) 𝛼 = 4◦ , (e) 𝛼 = 8◦ , (f) 𝛼 = 14◦ , (g) 𝛼 = 16◦ .

Here, 𝑥 and 𝑥𝑛 represent the flow variable and the normalized flow 1 × 10−4 . Adam is a first-order gradient-based optimization algorithm of
variable, respectively. 𝜇 and 𝜎 are the mean and standard deviation stochastic objective functions. The training aims to minimize the output
of the flow variable, respectively. Original flow field solution data of the loss function on the training set. The loss function measures
are interpolated from the CFD solution on structured mesh using a the dissimilarity between the output flow field of the model and the
linear triangulation-based method to generate ground truth data onto ground truth. In this study, we use mean squared error (MSE) as the
loss function. MSE is expressed as
a 256 × 256 Cartesian grid, which corresponds to 𝐷𝐹 .
1 ∑
𝑁
The neural network is first trained by employing Adam (Adaptive
= [(𝐶 − 𝐶̃𝑝 𝑘 )2 + (𝑀𝑘 − 𝑀
̃ 𝑘 )2 ] (6)
moment estimation) optimizer [34] with an initial learning rate of 2𝑁 𝑘=1 𝑝 𝑘

5
C. Duru et al. Computers and Fluids 236 (2022) 105312

Fig. 6. The surface pressure coefficient distribution on Eppler 547 at (a) 𝛼 = −8◦ , (b) 𝛼 = −4◦ , (c) 𝛼 = 0◦ , (d) 𝛼 = 4◦ , (e) 𝛼 = 8◦ , (f) 𝛼 = 14◦ , (g) 𝛼 = 16◦ .

where 𝐶̃𝑝 and 𝑀 ̃ are the model outputs for the pressure coefficient and set allows early stopping and tuning hyperparameters of the model,
Mach fields, respectively. 𝑁 is the number of data points of the flow e.g., the filter size, the number of feature maps, the number of layers,
field and subscript 𝑘 represents the pixel value. The airfoil shapes are
the activation function and the learning rate. Eventually, the network
masked through the training process, and the boundary data are not
provided to the model. Therefore the loss values for those regions are is evolved to construct flow fields at a minimum loss while keeping the
always 0. number of weights to a minimum in order to reduce training time. The
During the training stage, the model hyperparameters are optimally hyperparameters of the model are also calibrated in order to prevent
determined by considering the model performance on the validation overfitting of the model. The training process is performed until the
set. Therefore, the convergence history of the loss function is monitored
loss reached the steady-state for the validation set, as shown in Fig. 4.
for both training and validation sets, as shown in Fig. 4. The validation
set is used to observe sudden responses of the model on the ongoing The training is terminated after 4800 epochs when the loss no longer
optimization process. Monitoring the model responses on the validation reduces considerably for training and validation sets, respectively.

6
C. Duru et al. Computers and Fluids 236 (2022) 105312

are observed at the separation location and the shock wave. In fact, to
overcome this issue, we can increase the resolution of the flow field;
however, this will add extra cost to the training time.
Table 1 presents the model performance evaluations for the cases
given in Fig. 5. The table is organized as 𝐶𝑝 evaluations in three
columns followed by Mach number evaluations in the same order. For
both 𝐶𝑝 and 𝑀 evaluations, the first two columns correspond to the
percentage of the number of data points that satisfy the given metric,
𝐀𝐖𝐓 in Eq. (8), and the third column corresponds to the relation
Fig. 7. Pressure and shear stress on an airfoil surface element.
100−𝐌𝐀𝐏𝐄. The threshold is chosen (𝛿 = 1.1 and 𝛿 = 1.2) for 𝐀𝐖𝐓. The
threshold allows monitoring a certain closeness of the predicted data
to the ground truth. As the threshold approaches 1, the 𝐀𝐖𝐓 indicates
3. Results and discussion more accurate data points. The data points having more than 100%
error are treated as outliers and excluded from the 𝐌𝐀𝐏𝐄 calculations.
3.1. Evaluation metrics At most 2% of data in a predicted field treated as outliers. Even though
the outliers are excluded from the 𝐌𝐀𝐏𝐄 calculations, they are still
After the training is completed, the prediction performance of the present and can be observed in 𝐶𝑝 and 𝑀 fields given in Fig. 5. Table 1
model is evaluated by performing experiments utilizing the unseen shows that our model achieves high performance for both 𝐶𝑝 and
cases in the test set. Several evaluation metrics are used. The model 𝑀 field predictions. Much better performance is achieved in the 𝑀
outputs are compared with the ground truth flow fields qualitatively by predictions compared to the 𝐶𝑝 predictions. This difference stems from
the absolute difference of the field quantity between the model output the fact that, the 𝐶𝑝 predictions are more sensitive to higher gradient
and the ground truth. The mean absolute percentage error (𝐌𝐀𝐏𝐄) is regions, especially shocks, than 𝑀 predictions as shown in Fig. 5. This
investigated for each case. 𝐌𝐀𝐏𝐄 is defined as is also consistent with pressure — Mach relation both for isentropic
expansion and for (normal and oblique) shocks for an ideal gas. Close to
1 ∑ || 𝑦𝑘 − 𝑦̃𝑘 ||
𝑁
𝐌𝐀𝐏𝐄 = | | × 100 (7) the sonic flow, pressure changes faster than the Mach number as it can
𝑁 𝑘=0 || 𝑦𝑘 ||
be observed from compressible flow tables. Therefore larger gradients,
where 𝑁 is the number of data points, 𝑦 and 𝑦̃ are the ground truth hence larger errors, on pressure coefficient are expected. Even with a
and the model output, respectively. Accuracy can be measured as (100− slight shift in the shock location can result in larger errors in 𝐶𝑝 fields
𝐌𝐀𝐏𝐄). than 𝑀 fields.
Accuracy with the threshold(𝐀𝐖𝐓) is also commonly used in prior Fig. 6 shows the predicted 𝐶𝑝 distributions on the surface of airfoils
studies [35–37]. 𝐀𝐖𝐓 is defined as given in Fig. 5. Note that the airfoil shapes are masked through the
( ) training process, and the boundary data are not provided to the model.
| 𝑦 | | 𝑦̃ |
max || || , || || < 𝛿 (8) Therefore, we extract the 𝐶𝑝 data from the immediate vicinity of the
| 𝑦̃ | | 𝑦 |
surface (approximately 1% of the chord) in the normal direction. This
Here, 𝑦 and 𝑦̃ are the ground truth and the model output, respectively. is a valid assumption that the static pressure can be considered constant
𝛿 is the threshold. Since 𝐶𝑝 field is clustered around 0, (that indicates through the normal direction in the boundary layer based on the bound-
pressures close to the free stream pressure), the 𝐀𝐖𝐓 error estimations ary layer theory [38]. Fig. 6 shows that the model predictions agree
can be calculated too large to assess the quality of the estimation. well with the ground truth. For the negative angles of attack, a shock
Errors estimated with this method can be misleading, since division by wave develops on the lower surface. As the angle of attack increases,
a near-zero 𝐶𝑝 may produce large errors even if the actual error on the the shock wave moves to the leading edge and becomes stronger.
pressure is small. This issue can be easily addressed by applying relative The evolution of 𝐶𝑝 distribution, including the shock waves, is well
absolute error for the data points where |𝑦| < 0.1 by adding a bias to the predicted. The predicted shock strength and sharpness are generally
|𝑦−𝑦|
̃
denominator as 1+|𝑦| . It is known that the closeness to zero and the bias predicted well. The location of the shocks are estimated within 2%
are not invariant and scale-dependent. This new error definition is more of the chord (two pixels) compared to the ground truth. Note that,
reliable and improves interpretability of the prediction performance of the shift of the shock location results in large error estimations at
the model. these locations. In fact, both location and the strength of the shocks
are estimated well. Therefore, the error plots may be misleading and
3.2. The angle of attack variation for a single airfoil shape the prediction accuracy exceeds the expectations. The data points for
the 𝐶𝑝 distribution actually results in a small accuracy problem for
Fig. 5 shows a comparison between the ground truth and the model the integration of the forces. This problem will be addressed in future
output for the 𝐶𝑝 and 𝑀 fields of an example for several angles of attack studies.
cases from the test set. The absolute differences between the ground Furthermore, aerodynamic coefficients are extracted from the pres-
truth and the model output are also included in Fig. 5. Our model sure coefficient distribution on the airfoil surface. The net aerodynamic
performs well in capturing flow field evolution as the angle of attack force of an airfoil immersed in the fluid flow arises from the pressure
changes. The model simulates well the movement of the low-pressure and the shear stress distributions over the airfoil surface as depicted
region from the pressure to the suction side of the airfoil as the angle on a surface element in Fig. 7. The normal and axial components of
of attack changes from negative to positive even though almost 2/3 of aerodynamic force, 𝐹 , can be computed by
the data consists of positive angle of attack results.
Flow separation starts from the trailing edge and moves toward the 𝐹𝑁 = − 𝑝𝑛𝑦 𝑑𝑠 − 𝜏𝑛𝑦 𝑑𝑠 (9)
∮𝑆𝐴 ∮𝑆𝐴
leading edge with the increasing angle of attack until the recirculation
zone expands through the chord length. As the angle of attack is further 𝐹𝐴 = − 𝑝𝑛𝑥 𝑑𝑠 + 𝜏𝑛𝑥 𝑑𝑠 (10)
∮𝑆𝐴 ∮𝑆𝐴
increased, the flow will eventually separate up to the leading edge.
The shock location and strength change with the boundary layer/ where 𝑆𝐴 is the airfoil surface area, 𝜏 is the shear stress and 𝑛𝑥 , and 𝑛𝑦
separation location is also captured very well. Both 𝐶𝑝 and 𝑀 field are the components of the normal vector of the surface element 𝑑𝑠. It
predictions in Fig. 5 show good agreement with all the characteristic should be noted that only the pressure contribution on the net force is
features for the angle of attack variations. As expected, maximum errors available due to the reduced resolution of the boundary layer. For the

7
C. Duru et al. Computers and Fluids 236 (2022) 105312

Table 1
Accuracy results of 𝐶𝑝 and 𝑀 predictions of Eppler 547 for different angles of attack.
AoA 𝐶𝑝 𝑀
𝐴𝑊 𝑇 (𝛿 = 1.1)(%) 𝐴𝑊 𝑇 (𝛿 = 1.2)(%) 100 − 𝐌𝐀𝐏𝐄 𝐴𝑊 𝑇 (𝛿 = 1.1)(%) 𝐴𝑊 𝑇 (𝛿 = 1.2)(%) 100 − 𝐌𝐀𝐏𝐄
−8◦ 98.077 99.656 91.932 99.622 99.829 99.385
−4◦ 99.473 99.851 95.659 99.864 99.956 99.685
0◦ 98.257 99.727 95.286 99.704 99.893 99.605
4◦ 99.400 99.861 93.655 99.672 99.930 99.472
8◦ 99.227 99.937 95.055 99.757 99.937 98.238
14◦ 99.135 99.898 97.711 99.801 99.932 99.119
16◦ 99.628 99.885 97.854 99.528 99.877 98.768
Test set 97.424 99.275 94.045 99.180 99.716 99.012

Fig. 8. The comparison of the 𝐶𝑝 and 𝑀 fields of various airfoil shapes at 𝛼 = 10◦ . (a) S2055, (b) NACA 663 − 418, (c) R1080, (d) OAF 102, (e) RAE 101, (f) Eppler 417, (g)
NACA 23024.

resolution of the wall shear stress, much finer data representation than center. Note that the shear stress contribution on the moment is ne-
the current 256 × 256 Cartesian grid is required. Besides, the results glected due to above-mentioned reasons.
of our CFD analysis reveal that the net aerodynamic forces are mainly
1
dominated by the pressure forces. This is observed clearly for the lift 𝐶𝑀 = − 𝐶 𝑛⃗ × (𝑟 − 𝑟𝑐𝑎 )𝑑𝑠 (13)
𝐴𝑟𝑒𝑓 𝑙𝑟𝑒𝑓 ∮𝑆𝐴 𝑝
and also for the drag at transonic speeds in more than a few degrees of
angle of attack. Therefore, we present only the pressure contribution in where 𝑟 is the non-dimensional location along the chord length, 𝑟𝑐𝑎
Eqs. (9) and (10). We observe that, the exclusion of the friction drag denotes the aerodynamic center, which is the quarter-chord location.
brings no significant error. Note that, if desired this contribution can The reference area (𝐴𝑟𝑒𝑓 ) and the reference length (𝑙𝑟𝑒𝑓 ) are the chord
be calculated approximately and added to the drag calculation as a length of the airfoil. Table 2 shows the calculated aerodynamic coef-
further fix. This procedure is rather simple and accurate enough. In ficients for the given cases. It is seen that there is a good agreement
this study, we will keep our attention for the pressure-based drag and between the model predictions and the ground truth within a few
lift. As mentioned before, for the training we have used airfoil shapes percent, whereas some deviation is observed in the aerodynamic coeffi-
rotated by 𝛼. In other words, the 𝑥-direction is the wind axis. Therefore, cients when the angle of attack decreases. The reason for this deviation
the force calculations in 𝑥 directions yield pressure drag, and similarly, is due to the increasing contribution of the viscous forces as the angle of
the normal component of the force corresponds to the lift. When we attack decreases. The pressure difference, particularly around the shock
integrate the pressure coefficient distribution along the airfoil surface, wave as indicated in Fig. 5, causes the variation in the aerodynamic
we can obtain the lift and the pressure drag coefficients as follows coefficients.
1
𝐶𝐿 = − 𝐶 𝑛 𝑑𝑠 (11) 3.3. The airfoil shape variation at an angle of attack
𝐴𝑟𝑒𝑓 ∮𝑆𝐴 𝑝 𝑦
1 In order to investigate the sensitivity of the model predictions to
𝐶𝐷 𝑃 = − 𝐶 𝑛 𝑑𝑠 (12)
𝐴𝑟𝑒𝑓 ∮𝑆𝐴 𝑝 𝑥
handle the airfoil shape variation, different airfoil shapes from the
Also, the moment coefficient can be computed by integrating the test set at the same angle of attack of 𝛼 = 10◦ are examined. Fig. 8
moment of the pressure coefficient distribution about the aerodynamic shows a comparison between the ground truth and the model output for

8
C. Duru et al. Computers and Fluids 236 (2022) 105312

Fig. 9. The surface pressure coefficient distribution on (a) S2055, (b) NACA 663 − 418, (c) R1080, (d) OAF 102, (e) RAE 101, (f) Eppler 417, (g) NACA 23024 at 𝛼 = 10◦ .

Table 2
The comparison of aerodynamic coefficients between the model output and the ground truth with respect to the angle of attack variation for
Eppler 547 airfoil.
AoA 𝐶𝐿,𝑡𝑟𝑢𝑒 𝐶𝐿,𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛 |𝐸𝑟𝑟𝑜𝑟|𝐶𝐿 (%) 𝐶𝐷𝑃 ,𝑡𝑟𝑢𝑒 𝐶𝐷𝑃 ,𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛 |𝐸𝑟𝑟𝑜𝑟|𝐶𝐷 (%) 𝐶𝑀,𝑡𝑟𝑢𝑒 𝐶𝑀,𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛 |𝐸𝑟𝑟𝑜𝑟|𝐶𝑀 (%)
𝑃

−8◦ −0.537 −0.549 2.2 0.069 0.071 3.1 −0.078 −0.080 2.8
−4◦ −0.304 −0.296 2.6 0.025 0.026 2.1 −0.049 −0.053 8.4
0◦ 0.306 0.302 1.2 0.025 0.027 8.2 −0.058 −0.056 3.4
4◦ 0.474 0.481 1.3 0.057 0.055 3.3 −0.048 −0.048 0.0
8◦ 0.628 0.638 1.6 0.106 0.105 1.3 −0.056 −0.056 1.2
14◦ 0.808 0.809 0.2 0.219 0.218 0.8 −0.093 −0.093 0.5
16◦ 0.855 0.853 0.3 0.267 0.264 1.2 −0.110 −0.108 1.8

9
C. Duru et al. Computers and Fluids 236 (2022) 105312

Fig. 10. Network model predictions on NACA 0012 at 𝛼 = 8◦ .

Fig. 11. Control volume around the airfoil geometry. 𝑝𝐶𝑠 is the pressure distribution on the Control surface and 𝜌𝐕𝐬 (𝐕𝐬 ⋅ 𝐧) is the momentum flux.

the 𝐶𝑝 and 𝑀 fields, including the corresponding absolute difference 3.4. Out-of-dataset generalization performance
for airfoils of different shapes. Shape characteristics of the airfoils
compared in Fig. 8 vary in the amount of camber and thickness-to-
chord ratio. Thickness distributions also change along the chord of the In order to give an insight on the generalization performance of the
airfoil. The effects of the shape parameters on the airfoil performance neural network model, independent data (first seen airfoil shape) is also
appear on separation point, shock location, and pressure distribution
experimented. This test is employed to ensure our model has not been
on the airfoil surface. The neural network model captures these flow
characteristics on the flow field remarkably well. Even though the overfitted. Our database of 204 airfoils does not involve any NACA-00
magnitude of the error is considerable on the high gradient zones series symmetrical airfoils. Therefore, our network did not learn about
of 𝐶𝑝 and 𝑀 data along the flow field, the accuracy of the model this family. We have tested our model on NACA 0012 airfoil from the
is high enough to reproduce the overall structure of the flow field. NACA-00 series at 𝛼 = 8◦ for this purpose. The predictions of 𝐶𝑝 and 𝑀
Table 3 demonstrates the accuracy of the predicted flow fields with fields with surface 𝐶𝑝 distribution are given in Fig. 10. The results agree
different metrics mentioned before. The table quantitatively validates
well with the ground truth, with correct shock location, shock strength
the visual accuracy observed on both 𝐶𝑝 and 𝑀 field predictions in
and separation features. As expected, a large margin of error has
Fig. 8. Moreover, 𝐶𝑝 distributions on the airfoil surfaces are plotted in
Fig. 9. Although there are some fluctuations in the network predictions, occurred in the high gradient regions, similar to previous experiments.
a similar trend with the ground truth is observed for the 𝐶𝑝 distribution The predictions of out-of-dataset geometry exhibits similar behavior
on the airfoil surfaces. with the previous results and confirms our model is not overfitted.

10
C. Duru et al. Computers and Fluids 236 (2022) 105312

Fig. 12. Comparison of the surface pressure coefficient distribution on Eppler 547 surface.

Fig. 13. Comparison of the surface pressure coefficient distribution on Eppler 547 surface.

3.5. Discussion on drag predictions This deviation can be fixed using two different approaches. The first
one is to employ a Control Volume (CV) approach. A control volume is
Although our drag force predictions with the explained CNN are defined around the airfoil geometry. This CV consists of 𝑚 × 𝑛 pixels on
mostly satisfactory, for thin airfoils at low angle of attack conditions, the prediction image and requires mapping to physical space. The setup
the exclusion of skin friction may result in significant deviations in is shown in Fig. 11. In this figure dark square is the boundary of the CV.
drag predictions. The method described above may not be extended for Forces 𝐹𝑥 and 𝐹𝑧 are axial and normal components of the forces applied
skin friction drag since the skin friction is calculated from a very thin by the airfoil on the aircraft axis. These forces can be converted to lift
computational cell adjacent to the airfoil wall, and the image provided and drag forces. The momentum and force distribution are available at
for the CNN model is too coarse for this purpose. On top of that, the the CV boundary. n represents the unit normal of the control surface.
training is not conducted for velocities. The shear stress is denoted as 𝜏. For the small angles, only 𝐹𝑥 and 𝜏𝑥𝑥

11
C. Duru et al. Computers and Fluids 236 (2022) 105312

Fig. 14. Comparison of the surface pressure coefficient distribution on Eppler 547 surface.

Table 3
Accuracy measurements for 𝐶𝑝 and Ma predictions of various airfoil shapes at 𝛼 = 10◦ .
Airfoil 𝐶𝑝 Ma
𝐴𝑊 𝑇 (𝛿 = 1.1)(%) 𝐴𝑊 𝑇 (𝛿 = 1.2)(%) 100 − 𝐌𝐀𝐏𝐄(%) 𝐴𝑊 𝑇 (𝛿 = 1.1)(%) 𝐴𝑊 𝑇 (𝛿 = 1.2)(%) 100 − 𝐌𝐀𝐏𝐄(%)
S2055 99.055 99.664 96.987 99.524 99.873 99.101
NACA 663 − 418 98.527 99.592 94.719 99.376 99.807 99.048
R1080 99.288 99.937 97.001 99.831 99.908 99.353
OAF 102 98.760 99.412 97.558 98.738 99.640 98.828
RAE 101 96.703 99.660 94.272 99.480 99.762 98.924
Eppler 417 98.854 99.673 95.555 99.635 99.872 98.995
NACA 23024 94.948 99.266 95.666 98.396 99.738 98.231

are required for drag calculations, but of course, this discussion can be before. For example, [39] for techniques and the application of the
extended to 𝑧 for more elaborate drag calculations. method.
Applying the conservation of momentum for the control volume: The second approach is to patch the force calculations with a
simpler ML method other than the CNN. Since the deviation at this
𝜕  Mach regime is not significant, patching the forces with a simpler
 𝜌𝐕𝑑 + 𝜌𝐕 (𝐕 ⋅ 𝐧) 𝑑
∭𝐶𝑉
𝜕 𝑡
 ∮𝑐𝑠 method is acceptable. Note that skin friction becomes more dominant
∑ for smaller Mach numbers, and model accuracy becomes important.
= 𝐅𝐜𝐯 = 𝐅𝐞𝐱𝐭 + −𝐩𝐜𝐬 𝐧𝐝 + 𝜏𝐢𝐢 𝑑
∮𝐜𝐬 ∮𝐜𝐬
3.6. Discussion of the data representation
For steady flow field the first term is canceled. Also for 2D flow
and for 𝑥-direction the shear stress can be estimated by numerical The first step in our training process is to represent the CFD data
differentiation of: efficiently and feed this data to neural networks. In this study, each
flow field parameter in CFD data is projected onto a 256 × 256
𝑑𝑢 Cartesian grid. The simplification and standardization of CFD data
𝜏𝑥𝑥 = 𝜇
𝑑𝑦 representation are necessary for an uncomplicated CNN implementa-
The shear stress given above is calculated numerically at the top and tion.uncomplicated CNN implementation. However, the Cartesian grid
bottom surfaces for calculation of drag force. Note that the shear stress projection costs a certain loss of accuracy due to the interpolation of
contribution is significant only at a very small distance from the airfoil. grid data. This is particularly important at the vicinity of airfoil walls.
The shear stress contribution is negligible for the CV shown in Fig. 11, Mesh refinement is necessary to capture sudden changes in the flow
and the shear stress calculations can safely be eliminated. Therefore, features generated by no-slip boundary condition. Fig. 12 shows the
CV approach eliminates the skin friction estimations. comparison of raw CFD data with interpolated data with respect to
Only downside of this approach is to requirement of training for the surface pressure coefficient distribution on Eppler 547 surface at
velocity and the density. The inclusion of other parameters may effect 𝛼 = 4◦ . Interpolated data agrees well with the raw CFD data. It is
the accuracy of the CNN model. For this study, we have trained our important to notice that there are considerable deviations between the
model for Mach number to assess the prediction quality around shocks, CFD data and the interpolated data at the suction side and around the
therefore all force predictions are provided with the method described stagnation location at the leading edge. Also, there is a slight shift on

12
C. Duru et al. Computers and Fluids 236 (2022) 105312

Fig. 15. Histograms of the test set accuracy results. (a) 𝐶𝑝 field, (b) 𝑀 field.

the shock location. On the other hand, this deviation does not discredit respect to the 100 − 𝐌𝐀𝐏𝐄. Reminding most of the errors are stemmed
the interpolated data. A comparison of the lift coefficient and the drag from the pixel location shift of the shocks, we can argue that the neural
polar is made between the CFD results, interpolated data and the model network does an excellent job in learning the flow around airfoils
output in Figs. 13 and 14 respectively. While the interpolated data for the given conditions. When we consider the nonlinearity of the
shows good agreement with the CFD data for the lift coefficients as problem, including discontinuities of shock waves and flow separation,
shown in Fig. 13, there are some offsets in the drag polar (see Fig. 14) our model performs remarkably well.
in the drag direction. However, these offsets cannot be associated only
to poor resolution in the leading edge. Main contributing factor is the 3.7. Assessment of prediction accuracy and utilization potential.
lack of the friction drag as mentioned before. The model outputs agree
well with the interpolated data (ground truth) as shown in Figs. 13 and The current training database was extracted with a carefully crafted
14. computational grid and accurate CFD-based flow predictions. Each CFD
Overall, considering the 𝐌𝐀𝐏𝐄 metric, we achieve overall 94% and simulation takes 85 min on 32 Xeon cores. On the contrary, ML-based
99% accuracy for 𝐶𝑝 and 𝑀 field predictions, respectively. Fig. 15 predictions require less than a second for each case. Of course, this
shows the histograms of the accuracy results of the test cases with comparison can be misleading since the airfoil performance predictions

13
C. Duru et al. Computers and Fluids 236 (2022) 105312

Fig. 16. Comparison of CFD and machine-learning based 𝐶𝑝 distribution around the Eppler 547 airfoil.

are conducted with simpler, faster and less accurate methods during predictive ability of the model is demonstrated by the flow field con-
the initial design phase. For a better assessment of the performance tours and validated various accuracy metrics quantitatively. We have
of our method, we have defined a use case scenario. Assume that tested the performance of the model on the impacts of the angle of
we are designing a transonic aircraft that operates at transonic Mach attack and airfoil shape variations on the flow field. When we compared
numbers and operates at a wide AOA. Therefore, the flow field will with the CFD simulations, we observe that the model predictions are
contain shocks and flow separations, and a viscous flow solver should very accurate for both capturing effects of shape and angle of attack
be employed. We have employed a fast and robust pressure-based variations on the flow field. The errors are accumulated around the
implicit flow solver for faster CFD simulations with the same Spalart– high gradient of 𝐶𝑝 and 𝑀 along the flow direction. The proposed
Allmaras turbulence model for similar performance with our training model made it possible to avoid time-consuming CFD simulations with
data. The wing surface is modeled with 500 nodes around the wing. a slight loss of accuracy. After 360 h training process using 2 𝑥 TESLA
Twenty high-quality boundary layer meshes are employed for viscous V100 GPU, the prediction time for a single case is approximately 1
predictions. The final mesh consists of 40 K mixed quad-tetrahedral s while the CFD solver produces the result of the corresponding case
cells. The simulations are conducted until not only residuals but also
in approximately 85 min using 32 Xeon E5-2690 cores. We strongly
forces are stabilized. The resulting simulations require four minutes on
believe that this work has a potential to help researchers in design
a similar parallel computer with 28 Xeon cores. Therefore, the results
optimization studies. Further to this study, the presented model can
can be obtained at 1/20 time compared to our solutions for a realistic
be trained with a wide range of the Reynolds numbers and the Mach
use case.
numbers to investigate the effects of the Reynolds number on the flow
The CFD predictions are compared with the ground truth and the
field in different flow regimes.
ML-based predictions in Fig. 16. It is seen that the Machine -Learning
based predictions are significantly better compared to quick CFD sim- The findings of this research are limited to 𝑀 = 0.7 flow conditions.
ulations. Also, the computing time for ML predictions is still a couple We have tested our algorithm in other subsonic to transonic Mach
of hundred times better than CFD simulations. Faster flow simulations numbers and obtained promising results. The flow predictions between
will further deteriorate the prediction accuracy, more accurate simu- the trained Mach numbers is still a challenge, especially at the transonic
lations will further increase the computing time gap. We can argue flow regime.
that our method provides an excellent alternative for fast and accurate The robustness and accuracy of our flow field prediction tool allow
airfoil predictions. The attention while constructing the ground-truth our research group to expand the tool to predict three-dimensional
database pays well. However, as discussed before, our method has one flows, Mach divergence and other challenging problems. One short-
drawback compared to the CFD-based method at low AOA. Although coming of our method is the insensitivity to Reynolds numbers. The
the CFD solution described above does not provide accurate pressure Reynolds effects are less important for the transonic to supersonic flow
distribution, it provides skin friction as the output. This is known and fields. However, the extension of the method to lower Mach numbers
discussed shortcoming of our method and will be addressed in future requires carefully designed Reynolds number variation in the training
studies. set data.

4. Conclusion
CRediT authorship contribution statement
In this work, we have presented a deep learning methodology for
the prediction of pressure and Mach fields around airfoils for the angle Cihat Duru: Conceptualization, Methodology, Software, Validation,
of attack range from −10◦ to 20◦ at free-stream Mach number, 𝑀 = 0.7 Writing. Hande Alemdar: Conceptualization, Methodology, Software,
and free-stream Reynolds number, 𝑅𝑒 = 6 × 106 utilizing and improving Validation, Writing. Ozgur Ugras Baran: Conceptualization, Method-
our previously proposed neural network model CNNFOIL [24]. The ology, Software, Validation, Writing.

14
C. Duru et al. Computers and Fluids 236 (2022) 105312

Declaration of competing interest [17] Yilmaz E, German B. A convolutional neural network approach to training pre-
dictors for airfoil performance. In: 18th AIAA/ISSMO multidisciplinary analysis
and optimization conference. 2017, p. 3660.
The authors declare that they have no known competing finan-
[18] Zhang Y, Sung WJ, Mavris DN. Application of convolutional neural network
cial interests or personal relationships that could have appeared to to predict airfoil lift coefficient. In: 2018 AIAA/ASCE/AHS/ASC structures,
influence the work reported in this paper. structural dynamics, and materials conference. 2018, p. 1903.
[19] Bhatnagar S, Afshar Y, Pan S, Duraisamy K, Kaushik S. Prediction of
Data availability aerodynamic flow fields using convolutional neural networks. Comput Mech
2019;64(2):525–45.
[20] Sekar V, Jiang Q, Shu C, Khoo BC. Fast flow field prediction over airfoils using
The data that support the findings of this study are available from deep learning approach. Phys Fluids 2019;31(5):057103.
the corresponding author upon reasonable request. The neural net- [21] Thuerey N, Weißenow K, Prantl L, Hu X. Deep learning methods
work model is publicly available on https://github.com/Rakuen88/ for Reynolds-averaged Navier–Stokes simulations of airfoil flows. AIAA J
2020;58(1):25–36.
CNNFOIL.
[22] Hui X, Bai J, Wang H, Zhang Y. Fast pressure distribution prediction of airfoils
using deep learning. Aerosp Sci Technol 2020;105:105949.
Acknowledgment [23] Wang J, He C, Li R, Chen H, Zhai C, Zhang M. Flow field prediction of
supercritical airfoils via variational autoencoder based deep learning framework.
The authors are grateful to TÜBİTAK ULAKBİM for the computing Phys Fluids 2021;33(8):086108.
[24] Duru C, Alemdar H, Baran OU. CNNFOIL: convolutional encoder de-
resources. coder modeling for pressure fields around airfoils. Neural Comput Appl
2021;33(12):6835–49.
References [25] Paszke A, Gross S, Massa F, Lerer A, Bradbury J, Chanan G, et al. PyTorch: An
imperative style, high-performance deep learning library. In: Advances in neural
[1] Melin T. Parametric airfoil catalog. Linköping University Electronic Press; 2013. information processing systems. 2019, p. 8024–35.
[2] Fielding JP. Introduction to aircraft design, vol. 11. Cambridge University Press; [26] Selig M. UIUC airfoil coordinates database. Urbana, IL: UIUC Applied
2017. Aerodynamics Group; 2019, Accessed 19 May.
[3] Abbott IH, Von Doenhoff AE. Theory of wing sections: Including a summary of [27] Ioffe S, Szegedy C. Batch normalization: Accelerating deep network training
airfoil data. Courier Corporation; 2012. by reducing internal covariate shift. In: International conference on machine
[4] Faller WE, Schreck SJ. Neural networks: applications and opportunities in learning. PMLR; 2015, p. 448–56.
aeronautics. Prog Aerosp Sci 1996;32(5):433–56. [28] Clevert D, Unterthiner T, Hochreiter S. Fast and accurate deep network learning
[5] Rai MM, Madavan NK. Aerodynamic design using neural networks. AIAA J by exponential linear units (ELUs). In: Bengio Y, LeCun Y, editors. 4th Interna-
2000;38(1):173–82. tional conference on learning representations, ICLR 2016, San Juan, Puerto Rico,
[6] Hacioglu A. Fast evolutionary algorithm for airfoil design via neural network. May 2-4, 2016, conference track proceedings. 2016, URL http://arxiv.org/abs/
AIAA J 2007;45(9):2196–203. 1511.07289.
[7] Thirumalainambi R, Bardina J. Training data requirement for a neural network to [29] Spalart P, Allmaras S. A one-equation turbulence model for aerodynamic flows.
predict aerodynamic coefficients. In: Independent component analyses, wavelets, In: 30th aerospace sciences meeting and exhibit. 1992, p. 439.
and neural networks, vol. 5102. International Society for Optics and Photonics; [30] Toro EF, Spruce M, Speares W. Restoration of the contact surface in the
2003, p. 92–103. HLL-Riemann solver. Shock Waves 1994;4(1):25–34.
[8] Suresh S, Omkar S, Mani V, Prakash TG. Lift coefficient prediction at high angle [31] Cook P, Firmin M, McDonald M, Royal Aircraft Establishment. Aerofoil RAE
of attack using recurrent neural network. Aerosp Sci Technol 2003;7(8):595–602. 2822: pressure distributions, and boundary layer and wake measurements. RAE;
[9] Krizhevsky A, Sutskever I, Hinton GE. Imagenet classification with deep convo- 1977.
lutional neural networks. In: Advances in neural information processing systems. [32] Haase W, Brandsma F, Elsholz E, Leschziner M, Schwamborn D. EUROVAL—
2012, p. 1097–105. A European Initiative on validation of CFD codes—Results of the EC. 1990,
[10] Jin X, Cheng P, Chen W-L, Li H. Prediction model of velocity field around circular BRITE-EURAM Project EUROVAL 1992.
cylinder over various Reynolds numbers by fusion convolutional neural networks [33] Thomas JL, Salas M. Far-field boundary conditions for transonic lifting solutions
based on pressure on the cylinder. Phys Fluids 2018;30(4):047105. to theEuler equations. AIAA J 1986;24(7):1074–80.
[11] Fukami K, Fukagata K, Taira K. Super-resolution reconstruction of turbulent flows [34] Kingma DP, Ba J. Adam: A method for stochastic optimization. 2014, arXiv
with machine learning. J Fluid Mech 2019;870:106–20. preprint arXiv:1412.6980.
[12] Viquerat J, Hachem E. A supervised neural network for drag prediction of [35] Kuga R, Kanezaki A, Samejima M, Sugano Y, Matsushita Y. Multi-task learning
arbitrary 2D shapes in laminar flows at low Reynolds number. Comput & Fluids using multi-modal encoder-decoder networks with shared skip connections. In:
2020;210:104645. Proceedings of the ieee international conference on computer vision workshops.
[13] Lee S, You D. Data-driven prediction of unsteady flow over a circular cylinder 2017, p. 403–11.
using deep learning. J Fluid Mech 2019;879:217–54. [36] Eigen D, Fergus R. Predicting depth, surface normals and semantic labels with
[14] Chen J, Viquerat J, Hachem E. U-net architectures for fast prediction of a common multi-scale convolutional architecture. In: Proceedings of the IEEE
incompressible laminar flows. 2019, arXiv preprint arXiv:1910.13532. international conference on computer vision. 2015, p. 2650–8.
[15] Guo X, Li W, Iorio F. Convolutional neural networks for steady flow approxi- [37] Liu F, Shen C, Lin G. Deep convolutional neural fields for depth estimation from
mation. In: Proceedings of the 22nd ACM SIGKDD international conference on a single image. In: Proceedings of the IEEE Conference on computer vision and
knowledge discovery and data mining. ACM; 2016, p. 481–90. pattern recognition. 2015, p. 5162–70.
[16] Singh AP, Medida S, Duraisamy K. Machine-learning-augmented predictive [38] White FM, Corfield I. Viscous fluid flow, vol. 3. McGraw-Hill New York; 2006.
modeling of turbulent separated flows over airfoils. AIAA J 2017;2215–27. [39] Hue D, Esquieu S. Computational drag prediction of the DPW4 configuration
using the far-field approach. J Aircr 2011;48(5):1658–70.

15