Professional Documents
Culture Documents
Airfoil Shape Optimization Using Genetic Algorithm Coupled Deep Neural Networks
Airfoil Shape Optimization Using Genetic Algorithm Coupled Deep Neural Networks
Airfoil Shape Optimization Using Genetic Algorithm Coupled Deep Neural Networks
net/publication/373468165
CITATION READS
1 183
6 authors, including:
All content following this page was uploaded by Ming-Yu Wu on 01 September 2023.
CrossMark
View Export
Online Citation
An upper diameter bound for compact Ricci solitons with application to the Hitchin–Thorpe inequality. II
J. Math. Phys. (April 2018)
Ming-Yu Wu (吴明雨),1 Xin-Yi Yuan (袁心怡),2 Zhi-Hua Chen (陈志华),1 Wei-Tao Wu (吴威涛),2 Yue Hua (华越),3,a)
and Nadine Aubry4
AFFILIATIONS
1
Key Laboratory of Transient Physics, Nanjing University of Science and Technology, Nanjing 210094, China
2
School of Mechanical Engineering, Nanjing University of Science and Technology, Nanjing 210094, China
3
Sino-French Engineer School, Nanjing University of Science and Technology, Nanjing 210094, China
4
Department of Mechanical Engineering, Tufts University, Medford, Massachusetts 02155, USA
a)
Author to whom correspondence should be addressed: yhua@njust.edu.cn
ABSTRACT
To alleviate the computational burden associated with the computational fluid dynamics (CFD) simulation stage and improve aerodynamic
parameterization method based on the study of Sobieczky13 and then trained network can generate more realistic airfoil samples. In nature,
applied GA, supplemented by mesh movement strategy, for optimiz- this method was proposed for geometric filtering to improve the opti-
ing the airfoil shape under turbulent flow conditions. He et al.14 pro- mization efficiency. Achour et al.38 adopted a conditional generative
posed the B-splines free-form deformation (FFD) method along with adversarial network (CGAN) to optimize the airfoil shape, where
adaptive parameterization to define the airfoil shape, overcoming CGAN can guide the shape transformation toward particular classes
robustness issues related to shape parameterization, mesh deforma- of labeled samples by labeling each airfoil shape with precalculated
tion, and flow solver convergence. Then, they utilized mesh deforma- aerodynamic characteristics. Li et al.39 employed deep learning techni-
tion techniques to change the CFD surface mesh and finally employed ques to develop a tailored airfoil modal parameterization method for
sparse nonlinear optimizer (SNOPT) to find the optimal airfoil shape. low-Reynolds-number airfoils. By applying this method to solve the
Since no unique correlations exist between functional requirements optimization problems using the sequential least squares programing
and airfoil geometric shape,15 the above processes need to be iterated (SLSQP) algorithm, the novel parameterization method showcased its
continuously to find the optimal solution, which will take a substantial capability to effectively define the design space and enhance the design
amount of time and computing resources. efficiency. To sum up, Li et al.40 provided a comprehensive review of
To alleviate the computational burden associated with the CFD aerodynamic shape optimization (ASO) through the lens of machine
simulation stage and improve optimization efficiency, surrogate mod- learning. They introduced the state-of-the-art techniques and the
els have been widely applied in aerodynamic optimization and remaining challenges in aerodynamic shape optimization (ASO) from
design.16–20 However, when sampling the design space, this method three fundamental perspectives: compact geometric design space, fast
can still exponentially increase the computational cost as the dimen- aerodynamic analysis, and efficient optimization architecture.
sion of the design space grows. To combat this, researchers have According to this, it is known that there are a few studies coupling
devoted to dimensionality reduction (DR) techniques for the original deep learning-based fast aerodynamic analysis and optimization algo-
parametric design space. Raul and Leifsson21 used a Kriging regression rithms. Therefore, to improve the efficiency of the existing optimiza-
surrogate model coupling PARSEC parameterization method to tion methods, this work focuses on implementing an efficient
reduce the computational burden of ASO, mitigating the airfoil’s optimization framework using deep learning and evolutionary
dynamic stall characteristics. Berguin and Mavris22 used principal algorithm.
component analysis (PCA) to re-formulate the optimization problem In addition to the studies on the computational cost reduction of
in a lower dimensional coordinate system, where the adjoint formula- the classic optimization process, scientists also pay increasing attention
which offered the reference metrics for designer to select the suitable aerodynamic coefficient prediction network with the GA optimization
parameterization based on the specific problem. method (GA-ACPN). Then, the primary methods are followed,
In this work, to develop an alternative to addressing the time- including airfoil geometry parameterization, the establishment of the
consuming CFD simulations during ASO, aerodynamic coefficient ACPN, and the coupling mechanism of bonding ACPN with GA opti-
modeling is conducted using a fully connected neural network (FCN), mization. In Sec. III, we demonstrate and discuss the prediction and
a nonlinear mapping function between the airfoil geometry and the optimization results obtained using the GA-ACPN framework.
aerodynamic coefficient. In particular, the aerodynamic coefficient Section IV concludes the current work.
prediction network (ACPN) takes as input a two-dimensional matrix
represented airfoil profile, which is constructed by varying the design II. METHOD
variables in the improved Hicks–Henne bump functions. To verify the In general, this paper focuses on proposing a novel approach for
ACPN model’s predictive performance for aerodynamic coefficient, aerodynamic shape optimization combining a deep learning-based
several test cases different from the previous airfoil design space are aerodynamic coefficient predictor and an evolutionary algorithm.
carried out. Subsequently, the optimization procedure that couples Specifically, aerodynamic coefficient modeling is first accomplished
ACPN with the classic GA optimization method is implemented in a via a fully connected network, which fast calculates the aerodynamic
numerical code to seek out the optimal airfoil satisfying both geomet- coefficient from the airfoil shape, replacing the computationally expen-
ric constraints and functional requirements. In general, the main sive CFD simulations. Then, GA is coupled with the ACPN model to
objective and contributions of the present work are summarized as find the optimal airfoil under multiple constraints.
follows:
(1) This paper developed a high-precision aerodynamic surrogate A. Overall architecture of GA-ACPN framework
model called the ACPN model using the deep learning tech-
The overall GA-ACPN framework is mainly composed of three
nique, which implements a fast and accurate surrogate model-
ing from airfoil geometry to aerodynamic performance. parts: airfoil geometry parameterization, aerodynamic coefficient
(2) The improved Hicks–Henne bump functions were adopted to modeling, and airfoil shape optimization, as shown in Fig. 1. First, the
parameterize airfoil shapes and generate abundant airfoil coor- improved Hicks–Henne bump functions are adopted to parameterize
dinates library. the airfoil geometry. Particularly, to ensure the uniformity of the airfoil
space, the Latin hypercube sampling (LHS) technique is utilized to
FIG. 2. Comparison of the original and improved Hicks–Henne bump functions [the upper surface of NACA0012 with 0 angle of attack (AOA) in red]: (a) original and (b)
improved.
the improved Hicks–Henne bump functions are adopted to describe can be significantly altered without airfoil distortion, where a is the
various airfoil profiles55 slope coefficient and b is the decay coefficient. In this study, a and b
are set as 5 and 10, respectively.
X
k¼n
Theoretically, the more design variables are employed, the more
yup ¼ y0up þ ck fk ðxÞ; (1)
accurate airfoil curves will be obtained by improved Hicks–Henne
k¼1
bump functions, and the more satisfied optimization results are. To
X
k¼n simplify the airfoil parameterization process and reduce complexity
ylow ¼ y0low þ ckþn fk ðxÞ; (2) and time consumption, this study opts to select only five control posi-
k¼1 tions on both the upper and lower surfaces of the airfoil, i.e.,
8
FIG. 3. Comparison of the airfoil shape generated by the original and improved Hicks–Henne bump functions: (a) original and (b) improved.
FIG. 4. Effect of design variables of Hicks–Henne bump functions on the airfoil’s baseline geometry: (a) effects of the individual design variable on the airfoil shape and (b)
design space.
0.007 individually while keeping other design variables zero. As a two-dimensional coordinates depicting the airfoil profile, which are
result, the baseline geometry of NACA0012 airfoil expands outward, generated by the improved Hicks–Henne bump functions. The output
as shown in Fig. 4(a). Combined with Fig. 2, it can be seen that if the is the lift and drag coefficients. In general, the functional mathematical
design variable changes, the shape at the corresponding control point expression of the ACPN model is as follows:
is most altered obviously, along with the slightly weaker change from
the leading to the trailing edge of the airfoil, which is caused by the lin- ðCl ; Cd Þ ¼ f ðx; y; W; bÞ; (5)
ear combination of a set of basis Hicks–Henne bump functions. In where Cl and Cd indicate the lift and drag coefficient, respectively. (x,
addition, when we set c3 as 0.007 individually, it can be observed y) represents the two-dimensional coordinates of the given airfoil
that the baseline shrinks inward. Furthermore, Fig. 4(b) depicts the air- shape, and W and b are the weight and bias parameter matrix of the
foil space used for the dataset preparation, where all the design varia- ACPN model. It should be emphasized that W and b are the parame-
bles are limited in the range of [0.007, 0.007]. ters of the ACPN model to be trained and learned, with which the
ACPN model can be implemented to regress a nonlinear function f
TABLE I. ACPN model parameters of each fully connected layer. the f in Eq. (5) can be derived as follows if regarding L as the index of
the last layer:59
Layer name Executing operation/neuron Shape
f ¼ W ½L r z ½L1 þ b½L ; (9)
Input 61 2
Flatten 132 1 where W ½L and b½L represent the weight and bias matrices of the last
FC1 Fully connected/1024 1024 1 layer, z ½L1 represents the output of the ðL 1Þ layer, and rðz ½L1 Þ
FC2 Fully connected/512 512 1 refers to the activated output. It can be deduced from Eq. (8) that the
FC3 Fully connected/64 64 1 ReLU function can avoid exponential operations during forward prop-
FC4 Fully connected/2 21 agation, and since its output is unsaturated, the problem of gradient
vanishing can be prevented during the backpropagation. Both advan-
Output Reshape 21
tages speed up the update of network parameters. In addition, the
backpropagation involves the update of network parameters W and b,
which will be introduced in Sec. II C 3.
a½l ¼ r z ½l ; (7)
½l ½l
where W and b denote the weight and the bias, and z is the accu- ½l 2. Data preparation
mulated output value of the l th layer. Furthermore, to make the net- To ensure the reliability of the data fed into the ACPN model,
work deeper with sufficient nonlinear expression ability, each hidden CFD validation is conducted before preparing the data in batches. The
layer is activated by an active function, performed by rðÞ. a½l and a½l1 typical NACA0012 (National Advisory Committee for Aeronautics)
are the active value of z ½l and z ½l1 , respectively. Specifically, the recti- 2D airfoil structure is adopted. To generate the hybrid mesh with
fied linear unit (ReLU) function is employed as the activation function boundary layers aligned on the airfoil surface, Gmsh is utilized. Gmsh
in this work, which can be mathematically described as follows:58 is an application programing interface (API) in Python that facilitates
the creation of high-quality meshes for CFD simulations. To allevi-
x; x 0;
rðxÞ ¼ (8) ate the effect of the far-field boundary on the evolution of flow field
0; x < 0:
near the airfoil, the lengths in the perpendicular and chord direction
We input the airfoil coordinates at the first layer, and then, the forward are set to 40c and 100c, respectively, where c is the airfoil chord length.
FIG. 6. Schematic diagram of the calculation domain and mesh in the validation case.
the local region mesh with boundary layers in the sub-window. The
total number of grids in the computational domain is approximately
4 104.
The simulations are performed in OpenFOAM with the gener-
ated mesh, and the Spalart–Allmaras (SA) turbulence model is chosen
to solve the incompressible (Reynolds-averaged Navier–Stokes) RANS
equations.60 The operating conditions for the simulation are defined
as: the Mach number (Ma) is 0.15, the Reynolds number (Re) is
6 106, and the angle of attack (AOA) is 10 . The boundary condi-
tions on the SA turbulence field variable read61
~ wall ¼ 0; ~ farfield ¼ 4 fluid ; (10)
utilizes the first and second moments to update the network 1. Objective functions
parameters64
The objective function is a crucial aspect of the optimization pro-
Vdh :¼ b1 Vdh þ ð1 b1 Þdh; (13) cess. In our study, the optimization goal is to maximize the airfoil’s lift
coefficient and simultaneously minimize the drag coefficient. To con-
Sdh :¼ b2 Sdh þ ð1 b2 Þdh2 ; (14)
vert the problem into a single objective, we introduce the ratio of lift
where h represents the network parameters W and b collectively as and drag coefficient, and two objective functions are investigated,
their updating rules are the same, dh and dh2 refer to the first and sec-
Cl
ond gradients of the loss function J with respect to h, b1 and b2 are f1 ðck Þ ¼ ; k ¼ 1; 2; …; 14; (19)
the exponential decay factors of the moment estimation, and Vdh and Cd
Sdh are the first and second moment estimations. Considering that Vdh Cl
f2 ðck Þ ¼ þ d ðCl Cl r Þ; k ¼ 1; 2; …; 14: (20)
and Sdh will be biased toward zero during the updating process when Cd
they are initialized using zero vectors, to rectify these biases within The first objective function f1 directly equals the lift and drag coeffi-
each training iteration, the following corrections are applied: cient ratio. While the second objective function f2 adds one more pen-
Vdh alty term d ðCl Cl r Þ to drive the lift coefficient Cl greater than
corrected
Vdh ¼ ; (15) Cl r , the lift coefficient of the original airfoil, whose value is 0.5. d is
1 bt1
the adjustment factor and is set to 500 in our study.
Sdh
Scorrected
dh ¼ ; (16)
1 bt2 2. Optimization procedure of GA-ACPN framework
where the superscript “t” of b denotes the exponential operator, and The optimization procedure of GA-ACPN for airfoil shape opti-
corrected
Vdh and Scorrected
dh denote the corrected first and second moments. mization in the study is summarized in Fig. 8; the details are in the fol-
Finally, the network parameters h are updated iteratively by lowing steps:
corrected
Vdh (a) First, the control points xk are selected, and the correspond-
h :¼ h a qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi; (17)
Scorrected þ e ing ranges of design variable ck are determined. Then, accord-
dh
ing to binary encoding rules, the population of the first
adopt the single-point crossover operator, randomly selecting a cross- (SIM), which randomly specifies a real number to represent the posi-
over point with a certain probability. Then, the genetic information tion to be mutated and then reverses the string at the position in an
of the two parents beyond the point is swapped with each other.67 individual solution.67
Figure 10 illustrates the scheme of the single-point crossover. Selection is a crucial step in GA, which determines whether a
Mutation operators modify the values of some genes in a given string will participate in reproduction or not, and the con-
genome. This work employs the simple inverse mutation operator vergence rate of GA is greatly influenced by the selection pressure.
The rank selection technique utilizes the ranks given according to their
TABLE III. Hyperparameters setting of GA.
fitness values to endow each individual with a chance to be selected,
reducing the premature convergence probability of the solution to a
Description Value
local minimum.67
Binary encoding length 10 The used hyperparameters of the GA are summarized in Table III.
Population size 60
III. RESULTS AND DISCUSSIONS
Crossover probability 0.9
Mutation probability 0.01 A. Hyperparameters analysis of the ACPN model
Evolutionary generations 400 This subsection focuses on investigating the impact of two crucial
hyperparameters, namely, the learning rate and the mini-batch size,
FIG. 11. Convergence history under different sets of hyperparameters (smoothened by the exponential weighted moving average and presented in the dark colors): (a) various
learning rates with a batch size of 64, (b) various mini-batch sizes with a learning rate of 0.001, and (c) various training data size.
TABLE IV. Comparison of testing accuracy of the ACPN model with different training validation accuracy. The generalization ability of the ACPN model
data sizes. under the three cases will be given in Table IV (Sec. III B).
Number of training data CL Cd Average
B. Prediction performance of the ACPN model
1000 99.02% 98.53% 98.78% Following the methodology given in Sec. II C 2, the network was
3000 98.99% 98.75% 98.87% trained, and the hyperparameters were identified in Sec. III A. This
6000 99.16% 98.88% 99.02% subsection compares the aerodynamic coefficients predicted by the
trained ACPN model and the CFD results (the ground truth), where
the prediction performance of the ACPN model is evaluated by calcu-
on the convergence performance of the proposed ACPN model. lating the relative errors using Eq. (18). Unlike the training process,
We conducted a study considering various combinations of these the network parameters remain fixed and are reloaded during the eval-
hyperparameters values. Figures 11(a) and 11(b) illustrate the results, uation period. At this stage, the ACPN model can be treated as a
where the red lines stand for the reference combination. In the refer- purely nonlinear mapping function. In order to employ the trained
ence combination, the learning rate is set to 0.001, and the mini-batch ACPN model to predict aerodynamic coefficients, referencing the
size is 64. Additionally, two more values of the learning rate, one input and output of the ACPN model, the design variables need to be
before and the other after 0.001 with a batch size of 64 are studied in first sampled, followed by generating airfoil coordinates according to
Fig. 11(a), and similarly, two more results with the learning rate of the Hick–Henne parameterization method. Finally, fed it into the
0.001 are investigated in Fig. 11(b). Moreover, the influence of the size ACPN model, the forward propagation figures out the corresponding
of training data is shown in Fig. 11(c), where “N” denotes the training aerodynamic coefficients. The prediction example of aerodynamic
data size. coefficients is shown in Fig. 12. Following the same approach, the
Figure 11(a) shows both the training and validation accuracy batch predicted aerodynamic coefficients can be obtained, and the pre-
curves with the three learning rates (abbreviated as “lr” in the figure). diction performance of the ACPN model will be evaluated below.
Clearly, all curves of the three learning rates rapidly increase in the first Figures 13(a) and 13(b), respectively, depict the predicted lift and
2000 epochs, and then, the growth becomes slow until the end of train- drag coefficients of the proposed model. In the left column of the fig-
ing. The learning rate of 0.001 results in the highest accuracy. In addi- ure, it can be observed that the majority of samples align closely with
the line of y ¼ x, signifying a strong agreement between the ACPN
FIG. 12. The prediction example of aerodynamic coefficients using the trained and tested ACPN model.
needs to explore the huge airfoil space, although the mapping from 50% of the relative errors in lift coefficient prediction fall within the
61 2 airfoil coordinates to 2 1 aerodynamic coefficient is simple. range of 1.48%–0.4%, while approximately 50% of the relative errors
Figure 14 intuitively demonstrates the relative error distribution in drag coefficient prediction are concentrated between 1.15% and
for the lift and drag coefficients. It can be observed that approximately 0.48%. Moreover, the absolute values of relative errors do not exceed
TABLE V. Comparison of the time cost for the lift and drag coefficient predictions by TABLE VI. Ranges of design variables.
the ACPN model and the CFD simulation.
Design variables Range Design variables Range
Method Data preparation Training cost Prediction cost
c1 [0.007, 0.007] c8 [0.007, 0.007]
CFD 3 min c2 [0.006, 0.006] c9 [0.006, 0.006]
ACPN (GPU) 50 h 20 s 5 ms c3 [0.004, 0.006] c10 [0.006, 0.004]
c4 [0.004, 0.007] c11 [0.007, 0.004]
c5 [0.005, 0.005] c12 [0.005, 0.005]
3.8% for both coefficient predictions. These findings suggest that the c6 [0.005, 0.005] c13 [0.005, 0.005]
predicted aerodynamic coefficients by the ACPN model can replace
c7 [0.006, 0.006] c14 [0.006, 0.006]
the CFD results for subsequent airfoil shape optimization and
design.
In addition to the high accuracy prediction level, the proposed
ACPN model also shows its extreme prediction speed, as seen in C. Optimization results via GA-ACPN framework
Table V. Although 50 h are required for the data preparation, it takes In this subsection, our focus lies on conducting an effectiveness
only 20 s to train the ACPN model, and then, only 5 ms are required test of the proposed GA-ACPN framework. The ranges of design vari-
for each prediction using the graphics processing unit (GPU) RTX ables ck are listed in Table VI. Moreover, to validate the reliability of
3080. The prediction time is four orders of magnitude faster compared the GA-APCN framework, the same optimization framework but with
to the CFD solver, so once the ACPN model replaces the CFD simula- the CFD simulator is performed to provide the referential optimized
tor, the computational time for aerodynamic design and optimization aerodynamic coefficients.
will be significantly saved. Additionally, our previous work35 have The GA-ACPN and GA-CFD optimization procedures are exe-
pointed out that compared with the Kriging model that needs to spend cuted for airfoil shape optimization. Figures 15 and 16 compare their
5179 s for training, deep learning based surrogate models only take a convergence histories throughout the optimization process using
very short time of 54 s, three time orders of magnitude faster than the objective functions 1 and 2 in Sec. II D 1, respectively. As seen from
traditional surrogate model. Furthermore, in comparison to our previ- the left column of the two figures, the fitness scores of different popu-
ous work,35 the FCN-based surrogate model developed in this study lations among each generation vary greatly at first, then climb rapidly,
FIG. 15. Convergence history and Cl, Cd evolution using the objective function 1 (the average fitness score is in deep red): (a) GA-ACPN and (b) GA-CFD.
coefficients, along with optimization target the lift–drag coefficient first objective function, the final lift coefficients are not, indicating that
ratio, increase with generations. These results signify that the optimal the optimization process may have been stuck in a local optimum.
airfoil shape satisfying both the geometric constraints and functional Figure 17 compares the optimized airfoil shapes with the GA-
requirements can be found by iterating the optimization process. ACPN and GA-CFD methods. It is noted that the optimized airfoils
However, it cannot be ignored that the final lift and drag coeffi- using the two methods for objective function 1 are slightly different
cients of the airfoils optimized by GA-ACPN and GA-CFD methods while very close for objective function 2; the reason may be some
are not exactly the same, as shown in Table VII; the reason may come uncertain randomness when initializing the population and slight
from the unknown randomness brought by the random initialization deviation between ACPN predicted and CFD simulated aerodynamic
of GA process. First of all, we can see that both the GA-ACPN and coefficients. Therefore, the optimized airfoil shapes using GA-ACPN
GA-CFD methods drive the lift–drag ratio to a similar level, 51.4% and GA-CFD methods are inconsistent. Even so, these results still con-
and 50.3% increment, respectively, using the first objective function, firm the effectiveness of the method coupling the proposed ACPN
and 55.4% and 60.0% with the second. It is apparent that the second model with optimization algorithms.
objective function gives a higher lift–drag ratio, where compared with Figure 18 illustrates the pressure distribution on the airfoil sur-
the first objective function, the only difference is that we add a penalty face before and after optimization by GA-ACPN and GA-CFD meth-
term to lead the lift coefficient of the optimized airfoil shape to greater ods. For objective function 1, the two results slightly deviate from each
than that of the original one. Therefore, the optimized lift coefficients other, whereas for objective function 2, they are in good agreement,
using the second function are both higher than 0.5076, while with the which is consistent with the previous results of optimized airfoil
FIG. 17. Comparison of optimized airfoils: (a) objective function 1 and (b) objective function 2.
shapes. Furthermore, it can be clearly seen that either for objective pressure near these areas is weakened and strengthened, as shown in
function 1 or 2, the upper surface pressure variation between the refer- Fig. 19, resulting in an enlarged pressure difference (see Fig. 18). These
ence airfoil and the optimized airfoil nearly equals the inverse value of changes contribute to an increased lift force and an improved overall
that of the corresponding lower surface, as the airfoil thickness of the aerodynamic performance of the airfoil. Furthermore, the pressure varia-
design points is constrained to be unchanged and thereby the displace- tion at the airfoil’s trailing edge also gains favorable aerodynamic
ments of the upper surface and lower surface are equal. In addition, performance.
the optimized pressure difference has a sudden increase near the air- Turning now to the computational time analysis, it is worth men-
foil’s trailing edge compared to the original one. The underlying rea- tioning that although it takes about 50 h to construct the ACPN
son for these phenomena is that the modification in geometric airfoil model, this timeframe is acceptable for building the high effective
shape at 90% of the chord length, causing changes in the flow: the ACPN predictor as the GA-ACPN method can obtain the optimiza-
high-pressure region on the pressure surface of the optimized airfoil tion results within 25 s while the GA-CFD method has to spend about
FIG. 18. Comparison of pressure distributions: (a) objective function 1 and (b) objective function 2.
• The trained and tested ACPN model demonstrates a high accu- In summary, the proposed GA-ACPN optimization technique
racy in predicting the lift and drag coefficients, achieving approx- largely improves the efficiency of classic aerodynamic shape optimiza-
imately 99.02%. Furthermore, each coefficient prediction takes tion and at the same time, assures the optimization effect, showing its
only 5 ms, four orders of magnitude faster than using the CFD great advantages and huge potential for fast aerodynamic design and
solver. optimization. Furthermore, these results suggest that it has great
• The proposed GA-APCN optimization framework can improve potential to address the optimization problems with varying con-
the lift–drag ratio by 51.4% in just 25 s, while 50.3% improve- straints for the same flow condition.
ment is achieved by the GA-CFD optimization approach with 5.5 In addition, the optimization method in this paper is applicable
days, 19 000 times slower than the GA-APCN method. Despite it in other similar research since the essence of current work is to con-
takes additional 50 h and 20 s for data preparing and model struct a data-driven surrogate model for fast and accurate perfor-
training, the GA-ACPN framework still leads to an impressive mance estimation of the complex systems. Then, by replacing the
62.1% reduction in the overall calculation cost. CFD solver in a traditional optimization framework with the surro-
• Moreover, when constraints were imposed on the lift coefficient, gate model, we can fast optimize the performance of the system.
the GA-ACPN optimization framework can still fast optimize the Therefore, as long as the optimization framework uses a CFD solver
airfoil shape, achieving a comparable increment in the lift–drag to do the evaluation, the proposed method can improve the optimi-
ratio with the GA-CFD optimization. zation efficiency.
However, the surrogate model constructed in this work is based specific flow condition. Suppose we want extent this approach to other
on the data from one specified flow condition. Therefore, using this flow conditions. In that case, the focus should be on developing an
model as the simulation tool, the current optimization framework can ACPN model that can accurately predict the aerodynamic coefficient
only achieve efficient aerodynamic shape optimization under the for diverse flow conditions, which is an aspect of our future work. The
related research includes the dataset design for different flow condi-
TABLE VIII. Comparison of time components consumed by different optimization tions, training the model efficiently with big datasets, etc.
framework.
Data Training Optimization Time This research was funded by the Natural Science Foundation
Method preparation cost time reduction of Jiangsu Province (No. BK20201302), the Key Laboratory of
Thermal Management and Energy Utilization of Aircraft, Ministry
CFD 5.5 days of Industry and Information Technology (Grant No.
(6 threads) CEPE2022016), and the State Key Laboratory of Mechanics and
ACPN (GPU) 50 h 20 s 25 s 62.1% Control for Aerospace Structures (Nanjing University of
Aeronautics and Astronautics) (Grant No. MCMS-E-0323Y01).
AUTHOR DECLARATIONS 16
Y. Yu, Z. Lyu, Z. Xu, and J. R. R. A. Martins, “On the influence of optimization
algorithm and initial design on wing aerodynamic shape optimization,”
Conflict of Interest Aerosp. Sci. Technol. 75, 183–199 (2018).
The authors have no conflicts to disclose. 17
Z. Hao, Z. Wang, and J. Liu, “Tail rudder optimization design and dynamics
modeling for the small flexible membrane wing aircraft,” in Proceedings of the
10th International Conference on Modelling, Identification and Control (IEEE,
Author Contributions 2018), pp. 1–6.
Mingyu Wu: Investigation (equal); Software (equal). Xinyi Yuan: 18
A. Sanchez-Carmona and C. Cuerno-Rejado, “Design process and environmen-
Visualization (equal). Zhihua Chen: Supervision (equal). Weitao Wu: tal impact of unconventional tail airliners,” Aerospace 8, 175 (2021).
19
J. S. Gray, C. A. Mader, G. K. W. Kenway, and J. R. R. A. Martins, “Coupled
Resources (equal); Validation (equal). Yue Hua: Writing – review &
aeropropulsive optimization of a three-dimensional boundary-layer ingestion
editing (equal). Nadine Aubry: Writing – review & editing (equal). propulsor considering inlet distortion,” J. Aircr. 57, 1014–1025 (2020).
20
A. S. Batrakov, A. N. Kusyumov, S. A. Mikhailov, and G. N. Barakos,
“Aerodynamic optimization of helicopter rear fuselage,” Aerosp. Sci. Technol.
DATA AVAILABILITY 77, 704–712 (2018).
Data sharing is not applicable to this article as no new data were
21
V. Raul and L. Leifsson, “Surrogate-based aerodynamic shape optimization for
delaying airfoil dynamic stall using Kriging regression and infill criteria,”
created or analyzed in this study.
Aerosp. Sci. Technol. 111, 106555 (2021).
22
S. H. Berguin and D. N. Mavris, “Dimensionality reduction using principal
REFERENCES component analysis with gradient information,” AIAA Paper No. 2014-0112,
1
H. Jiang, M. Xu, and W. Yao, “Aerodynamic shape optimization of co-flow jet 2014, pp. 1–18.
airfoil using a multi-island genetic algorithm,” Phys. Fluids 34, 125120 (2022).
23
S. H. Berguin, D. Rancourt, and D. N. Mavris, “Method to facilitate high-
2
Q. Du, T. Liu, L. Yang, L. Li, D. Zhang, and Y. Xie, “Airfoil design and surro- dimensional design space exploration using computationally expensive analy-
gate modeling for performance prediction based on deep learning method,” ses,” AIAA J. 53, 3752–3765 (2015).
24
Z. J. Grey and P. G. Constantine, “Active subspaces of airfoil shape parameter-
Phys. Fluids 34, 15111 (2022).
izations,” AIAA J. 56, 2003–2017 (2018).
3
A. R. Al-Obaidi, “Numerical investigation on effect of various pump rotational 25
D. Cinquegrana and E. Iuliano, “Investigation of adaptive design variables
speeds on performance of centrifugal pump based on CFD analysis technique,”
bounds in dimensionality reduction for aerodynamic shape optimization,”
Int. J. Model., Simul., Sci. Comput. 12, 2150045 (2021).
Comput. Fluids 174, 89–109 (2018).
4
M. M. Bhatti, M. Marin, A. Zeeshan, and S. I. Abdelsalam, “Recent trends in 26
Q. I. U. Yasong, B. A. I. Junqiang, L. I. U. Nan, and W. Chen, “Global aerody-
computational fluid dynamics,” Front. Phys. 8, 593111 (2020).
namic design optimization based on data dimensionality reduction,” Chin. J.
39
J. Li, M. Zhang, C. M. J. Tay, N. Liu, Y. Cui, S. C. Chew, and B. C. Khoo, “Low- 53
J. Achleitner, K. Rohde-Brandenburger, and M. Hornung, “Airfoil optimization
Reynolds-number airfoil design optimization using deep-learning-based tai- with CST-parameterization for (un-)conventional demands,” in 34th OSTIV
lored airfoil modes,” Aerosp. Sci. Technol. 121, 107309 (2022). Congress Conference Proceedings (2018), pp. 117–120.
40
J. Li, X. Du, and J. R. R. A. Martins, “Machine learning in aerodynamic shape 54
J. Kou, L. Botero-Bolıvar, R. Ballano, O. Marino, L. de Santana, E. Valero, and
optimization,” Prog. Aerosp. Sci. 134, 100849 (2022). E. Ferrer, “Aeroacoustic airfoil shape optimization enhanced by autoencoders,”
41
T. T. Zhang, Z. G. Wang, W. Huang, and L. Yan, “A review of parametric Expert Syst. Appl. 217, 119513 (2023).
approaches specific to aerodynamic design process,” Acta Astronaut. 145, 55
J. Wang and Z. Gao, “Analysis and improvement of HicksHenne airfoil param-
319–331 (2018). eterization method,” Aeronaut. Comput. Technol. 40, 47–49 (2010).
42
W. Zhu, M. H. McCrink, J. P. Bons, and J. W. Gregory, “The unsteady Kutta 56
D. A. Masters, N. J. Taylor, T. C. S. Rendall, C. B. Allen, and D. J. Poole,
condition on an airfoil in a surging flow,” J. Fluid Mech. 893, R2 (2020). “Geometric comparison of aerofoil shape parameterization methods,” AIAA J.
43
F. Mohebbi, B. Evans, and M. Sellier, “On the Kutta condition in compressible 55, 1575–1589 (2017).
flow over isolated airfoils,” Fluids 4, 102 (2019). 57
H. Taud and J. F. Mas, “Multilayer perceptron (MLP),” Geomatic Approaches
44
M. Gomez, R. Haimes, and M. C. Galbraith, “On analysis driven shape design for Modeling Land Change Scenarios (Springer, 2018), pp. 451–455.
using B-splines,” AIAA Paper No. 2022-1736, 2022, p. 1736. 58
K. Eckle and J. Schmidt-Hieber, “A comparison of deep networks with ReLU
45
D. Rajnarayan, A. Ning, and J. Mehr, “Universal airfoil parametrization using activation function and linear spline-type methods,” Neural Networks 110,
B-splines,” AIAA Paper No. 2018-3949, 2018. 232–242 (2019).
46
J. Wang, C. Wang, B. Zhou, L. Zeng, and K. Yang, “Airfoil shape and angle of 59
J. Wang, C. He, R. Li, H. Chen, C. Zhai, and M. Zhang, “Flow field prediction
attack optimization based on Bezier curve and multi-island genetic algorithm,” of supercritical airfoils via variational autoencoder based deep learning frame-
J. Fluids Eng. 144, 51203 (2022). work,” Phys. Fluids 33, 86108 (2021).
47
M. G. Lauer and P. J. Ansell, “A parametrization framework for multi- 60
M. Tiberga, A. Hennink, J. L. Kloosterman, and D. Lathouwers, “A high-order
element airfoil systems using Bezier curves,” AIAA Paper No. 2022-3525, discontinuous Galerkin solver for the incompressible RANS equations coupled
2022, p. 3525. to the k turbulence model,” Comput. Fluids 212, 104710 (2020).
48
X. Wei, X. Wang, and S. Chen, “Research on parameterization and optimiza- 61
P. Spalart and S. Allmaras, “A one-equation turbulence model for aerodynamic
tion procedure of low-Reynolds-number airfoils based on genetic algorithm flows,” AIAA Paper No. 1992-439, 1992, p. 439.
and Bezier curve,” Adv. Eng. Software 149, 102864 (2020). 62
S. L. Krist, R. T. Biedron, and C. L. Rumsey, “CFL3D user’s manual (version
49
J. C. Yu and R. Wulandari, “Airfoil aerodynamics optimization under uncer- 5.0),” NASA Technical Reports Server (NTRS) No. NASA-TM-208444, 1998.
tain operating conditions,” J. Phys. Conf. Ser. 1446, 012014 (2020). 63
C. L. Ladson, “Effects of independent variation of Mach and Reynolds numbers
50
S. Zhou, H. Zhou, K. Yang, H. Dong, and Z. Gao, “Research on blade design on the low-speed aerodynamic characteristics of the NACA 0012 airfoil sec-
method of multi-blade centrifugal fan for building efficient ventilation based tion,” Report No. NASA-TM-4074, 1988.
on Hicks-Henne function,” Sustainable Energy Technol. Assess. 43, 100971 64
I. K. M. Jais, A. R. Ismail, and S. Q. Nisa, “Adam optimization algorithm for
(2021). wide and deep neural network,” Knowl. Eng. Data Sci. 2, 41–46 (2019).