Hindawi

International Journal of Aerospace Engineering

Volume 2021, Article ID 5583681, 25 pages
https://doi.org/10.1155/2021/5583681

Research Article
Design Optimization of a Multistage Axial Flow Compressor Based
on Full-Blade Surface Parameterization and Phased Strategy

Jinxin Cheng ,1,2 Shengfeng Zhao ,1,2 Zhaohui Dong ,3 and Chengwu Yang1,2
1
Key Laboratory of Light-Duty Gas-Turbine, Institute of Engineering Thermophysics, Chinese Academy of Sciences,
Beijing 100190, China
2
College of Engineering Science, University of Chinese Academy of Sciences, Chinese Academy of Sciences, Beijing 100049, China
3
College of Engineering, Peking University, Beijing 100190, China

Correspondence should be addressed to Shengfeng Zhao; zhaoshengfeng@iet.cn and Zhaohui Dong; zhdong@pku.edu.cn

Received 19 January 2021; Revised 12 April 2021; Accepted 10 June 2021; Published 1 July 2021

Academic Editor: Andre Cavalieri

Copyright © 2021 Jinxin Cheng et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

A new design optimization method is proposed for the problem of high-precision aerodynamic design of multistage axial
compressors. The method mainly contains three aspects: full-blade surface parametrization can significantly reduce the number
of control variables per blade row and increase the degrees of freedom of the leading edge blade angle compared with the
traditional semiblade parametric method; secondly, the artificial bee colony algorithm improved initialization and food source
exploration and exploitation mechanism to enhance the global optimization ability and convergence speed, and a distributed
optimization system is built on the supercomputing platform based on this method; finally, a phased optimization strategy
based on the “synchronous change in multirow blades” is proposed, and expert experience is introduced to achieve a better
balance between exploration and exploitation. The optimization method is applied to the AL-31F four-stage low-pressure
compressor. As a result, the adiabatic efficiency is improved by 0.67% and the surge margin is improved by 3.1% under the
premise that the total pressure ratio and mass flow rate satisfy the constraints, which verifies the effectiveness and engineering
practicality of the proposed optimization method in the field of multistage axial flow compressor aerodynamic optimization.

1. Introduction dynamic optimization of multistage axial flow compressors,

which has a very important engineering application value.
Aeroengines and gas turbines are needed for strategic national Before the year 2000, limited by the computing power at
development, and one of the core components of aeroengines that time, aerodynamic optimization of the multistage axial
and gas turbines is the multistage axial flow compressor. For flow compressor was mainly limited to one dimension and
the design of multistage axial flow compressors, an aerody- two dimensions [2–5]. With the great improvement in com-
namic optimization design method can effectively reduce the puting power, three-dimensional steady aerodynamic opti-
dependence on manual experience and greatly improve the mization of compressors has been developed rapidly, and a
design capability. It is well known that the aerodynamic opti- few three-dimensional unsteady optimization methods have
mization design of a multistage axial flow compressor has the even been proposed [6]. However, the cost of unsteady calcu-
characteristics of typical high-dimensionality, expensive cost, lation is too large, much simplification is required, the actual
and black box (HEB) [1] problem. With an increase in optimi- calculation accuracy is limited, and there is no engineering
zation control variables, the design space will grow exponen- practicability. At present, research on multistage axial flow
tially and fall into the “curse-of-dimensionality.” Therefore, compressors still mainly focuses on three-dimensional steady
obtaining the optimized solution within the acceptable time optimization. With the development of optimization regard-
range of engineering is a frontier problem in the field of aero- ing multidiscipline and high-fidelity problems, the HEB
2 International Journal of Aerospace Engineering

problem mentioned above has increasingly become the key Another important part of the optimization process is the
to restrict the design of multistage axial flow compressors. optimization algorithm. To avoid falling into a local extre-
Traditional methods for solving the HEB problem of mum, this paper adopts an improved artificial bee colony
aerodynamic optimization of multistage axial flow compres- (IABC) algorithm with good global optimization ability
sors can be divided into six categories: (1) screening variables [31], which does not depend on the initial solution. In addi-
[7, 8], (2) stage-by-stage optimization [9], (3) spatial decou- tion, due to the improvements in initialization, food source
pling [10], (4) surrogate model method [11–15], (5) adjoint exploration, and exploitation mechanisms from the standard
algorithm [16–18], and (6) parametric dimensionality reduc- artificial bee colony [32] (ABC) algorithm, its convergence
tion [19, 20]. Among them, the screening method is only speed is improved. Because the evolutionary algorithm has
suitable for examples with obvious bad flow fields but is not the advantage that it does not depend on the gradient infor-
universal. The stage-by-stage optimization process is com- mation of the function, it can perform a simultaneous search
plex, and the coupling effect of the flow field between the of multiple individuals in the design space with the potential
front and back stages is ignored. The spatial decoupling for multitask concurrency. Combined with a commercial
method proposed in literature [10] is only suitable for the supercomputing platform with powerful parallel capabilities,
case of a weak radial flow but not for a blade with a low aspect the IABC algorithm can greatly reduce the optimization time.
ratio. The surrogate model method is difficult to train in However, because evolutionary algorithms essentially
high-dimensional problems, and the number of samples require many high-fidelity performance evaluations, the
needed increases exponentially with the dimensionality, so above strategy is still not enough to reduce the computational
the sample training cost is difficult to handle. Although the cost to an acceptable range. This paper proposes a phased
adjoint algorithm has the advantage that optimization is optimization strategy based on the “synchronous change in
independent of the dimensionality of variables, it easily falls multirow blades,” which can greatly reduce the time cost of
into local extrema [21] and has defects in handling multicon- optimization. The core of this strategy is to draw on the
straint multiobjective problems. idea of “exploration first, exploitation later” in the optimi-
The literature [22, 23] notes that one of the most promis- zation process and divide the entire optimization process
ing strategies for solving the HEB problem is parametric into two phases to effectively reduce the number of itera-
mapping dimensionality reduction. For aerodynamic optimi- tions required for the spatial search, thereby reducing the
zation of multistage axial flow compressors, the aim is to find optimization time.
a parametric method to reduce the design variables and The structure of this paper is as follows: Section 2
achieve dimensionality reduction while retaining as much shows the optimization method; Section 3 performs the
of the optimal solution from the original design space as pos- optimization case verification and analysis; Section 4 draws
sible. The traditional parametric method has remarkably a few conclusions.
high-dimensional characteristics because each section of the
blade is modeled separately and independently of each other 2. Optimization Method Based on Full-Blade
[24]. An effective method to constrain the control parameters Surface Parameterization
of each section is the surface parametric method, which was
first applied in outflow aerodynamic optimization and is an 2.1. Parametric Method. The traditional parametric method
effective dimensionality reduction method [25–28]. Lee shapes each radial section of the blade separately; on the
et al. compared B-spline and CST surface methods and con- one hand, it makes the blade profile difficult to smooth; on
cluded that the two parametric methods have their own the other hand, it makes the blade control parameters
advantages and disadvantages and are suitable for different numerous (a single row of blades can have hundreds of
situations. In 2003, the semiblade surface parametric method parameters). To better promote the solution of the HEB
(blade suction surface modification by a Bezier surface) was problem of multistage axial compressor aerodynamic optimi-
applied in the internal flow field of compressors by Burgu- zation, the surface parametric method with significant low-
buru and le Pape [29], which was further developed by Wang dimensional characteristics was selected. This method
and Zhou [30] and Cheng et al. [1]. However, the defect of understands the blade as a combination of the suction surface
this surface parametric method is that only half of the blade and pressure surface rather than the accumulation of two-
surface is modified, and the influence of the other half surface dimensional sections. This paper proposes a full-blade
on the flow field is ignored. In addition, this method cannot surface parametric method that is an improvement of the
change the blade angle at the leading edge; this defect plays semiblade surface parametric method [29]. Figure 1 shows
an important role in the development of the flow field as it the construction process of the method. The blade is opened
results in limiting the improvement space of the aerodynamic radially from the leading edge to form an unfolded surface,
performance for blade geometry. In this paper, a full-blade and the Bezier perturbed surface is superimposed to form a
surface parametric method is presented that regards the new unfolded surface. Then, the new unfolded surface is
suction surface and the pressure surface as a whole surface stitched at the leading edge in the radial direction to generate
and superimposes the perturbed surface on it. This method a new blade. The parametric method can be divided into the
reduces the control variables and realizes an effective dimen- following four steps:
sionality reduction under the condition of ensuring the
degree of freedom of the leading edge, trailing edge, and (1) The blade opens radially at the leading edge, forming
blade body. an unfolded surface
International Journal of Aerospace Engineering 3

Control
points

Original blade
Trailing edge
Casing
Bezier
Leading surface
Pressure dge
Suction surface Unfold the blade alone the
𝜂 surface leading edge
Leading 𝜉
edge Hub

1 Calculation of the moving distance

Chord-length parameterization of of blade surface points based on
0.8
the original blade Bezier surface formula
0.6
V2

0.4

0.2
Tip
0
0 0.2 0.4 0.6 0.8 1
Circumferential movement
V1
in each pint of the original
blade
L

Leading Trailing edge

dge
New blade

Hub

Original
Optimal

Figure 1: The construction process of parametric method.

(2) Chord length parameterization is applied. Because (3) Bezier perturbed surface generation is applied. The
the three-dimensional surface of the physical space perturbed surface is defined by the following formula:
and the unit Bezier surface are to be superimposed,
the coordinates of each point of the blade in the phys- ( )
n m
ical space need to be converted to the unit plane in !
the calculation space, and ξ and η are the axes of R= 〠 〠 P k ,l Bm
l ðv Þ Bnk ðuÞ, ð3Þ
k=0 l=0
the computational space. The formula for chord
length parameterization is as follows:
Bnk ðuÞ = C nk uk ð1 − uÞn−k , ð4Þ
8
>
< n!
∑im=1 lm , if 0 ≤ k ≤ n,
ξi , j = , ð1Þ n
Ck = ð n − kÞ!k! ð5Þ
T lj >
:
0, if not,
!
j where R refers to the change value of each point on the per-
∑ L turbed surface; Pk,l is the ðm + 1Þ × ðn + 1Þ control points of
ηi, j = n=1 n , ð2Þ
T Li the perturbed surface; Bm l ðvÞ and Bk ðuÞ refer to the Bernstein
n

basis functions calculated by equation (4); v and u refer to the

where i ∈ ð1, np Þ and j ∈ ð1, ns Þ; np and ns refer to the number independent coordinate variables of the perturbed surface,
of data points of a section and the total number of sections, and its variation range is ½0, 1; and Cnk is the combination
respectively; T l j refers to the sum of the chord lengths of number calculated by equation (5)
each segment of the jth section in the η direction; T Li refers !
to the sum of the chord lengths of the ith section in the ξ (4) The R value of each point of the obtained perturba-
direction; lm is the mth chord length of the jth section in tion surface is superimposed on the circumferential
the η direction; and Ln is the nth chord length of the ith direction of the corresponding point of the unfolded
section in the ξ direction surface, thereby generating a new blade
4 International Journal of Aerospace Engineering

The full-blade surface parametric method is a new tion is ranked according to the food source. The
surface parametric method that retains the high-order first half of the food source’s fitness is taken as the
smoothness characteristic while improving the existing semi- target of the employed bees, and the rest is left to
blade surface parametric method, which is mainly manifested the onlooker bees
in the following two points:
(3) According to formula (1), all bees explore the vicinity
(1) Since the full-blade surface parametric method com- of the initial source and calculate the concentration of
bines the suction surface and the pressure surface as a the food source. When the concentration of the new
whole, it can take into account the influence of the food source is greater than that of the original one,
suction surface and pressure on the aerodynamic per- the original food source will be replaced by the better
formance while not significantly increasing the num- one. If the opposite is true, the bees will continue
ber of control variables, and the low-dimensional to explore
characteristics are maintained well. The control
parameters of a single blade row can be reduced to 16    
j j j j j j j j
V i = X i + ξi X i − X k + 1:5λi X best − X i , ð6Þ
(2) The full-blade surface parametric method increases
the degree of freedom of the geometric variation at j
the leading edge. The semiblade surface parametric where V i refers to the jth component of the food source
method lacks the freedom of the blade angle at the location of the ith bee, j ∈ f1, 2,⋯,Dg ; k ∈ f1, 2,⋯,N/2g,
leading edge, so it has certain restrictions on the j j
i ∈ f1, 2,⋯,N/2g and k ≠ i ; ξi and λi are random numbers
geometric exploration function. Figure 2 shows a j
between ½−1, 1 ; and X best is the best food source location
schematic diagram of the point distribution of the
in the general information of the food source
full-blade parametric method. The yellow point is a
free active point, and the blue point is a limited active (4) Onlooker bees explore the food source. The onlooker
point; that is, all the blue points need to keep chang-
bees use the Russian roulette rule to select food
ing synchronously so that the leading edge of the sources, as shown in formula (2); that is, the food
suction pressure can be smoothly connected source is selected with a probability proportional to
the concentration of the food sources determined
2.2. Optimization Algorithm. For the purpose of searching for by the employed bees and is explored in the vicinity
the global optimal solution of the multistage axial flow com- according to formula (3). If the concentration of the
pressor, an evolutionary algorithm with global search ability obtained food source is higher than the concentration
is adopted. Since evolutionary algorithms do not require of the original food source, the original food source is
the gradient information of function values, multiple individ- replaced, and if not, the exploration is continued
uals can be concurrent, which can greatly reduce the optimi-
zation time. Among various evolutionary algorithms, the
ABC [32] algorithm has a better optimization accuracy and f ðX i Þ
convergence speed. In this paper, the ABC algorithm is devel- P= Ne
,
∑m=1 f ðX m Þ ð7Þ
oped through an initialization mechanism, a food source
      
exploration, and an exploitation mechanism, and its sche- j
V Neib =
j
X Neib
j
+ ξi
j
X Neib
j
− Xk ,
matics are shown in Figure 3. best best best
The IABC [31] algorithm involves employed bees,
onlooker bees, and detector bees. At the beginning, the where the value of N e is N/2, referring to the number of
employed and onlooker bees constitute the entire bee colony, employed bees. P refers to the probability of selecting a cer-
and the number of bees is divided equally between these tain food source; f ðX i Þ refers to the concentration of the ith
j
types. The food source in the schematic refers to a set of fea- and f ðX m Þ refers to the mth food sources; ðX Neib Þbest refers
sible solutions, and the food source concentration refers to to the food source position that has the largest concentration
the fitness of the feasible solutions. The detailed steps of the j
in the adjacent area; and ðV Neib Þbest refers to the location of
IABC algorithm are as follows:
the new food source of onlooker bees. The Chebyshev dis-
(1) Food source initialization. To obtain as much design tance formula is used to calculate the best point in the neigh-
space information as possible from as few sample borhood, which is shown in formula (4):
points as possible, instead of the random initializa- !1/q
D 
tion mechanism in the standard ABC algorithm, the  j j q
optimal Latin square experiment method [33] is d ði, t Þ = lim 〠 X i − X t  , ð8Þ
q⟶+∞
j=1
utilized to initialize the distribution of the food
source, and the feasible solution X i ði = 1, 2, 3,⋯,NÞ
where j ∈ f1, 2,⋯,Dg and dði, tÞ is the Chebyshev distance
of ND-dimensional vectors is generated
between X i and X t .
(2) The concentration of the initial food source is calcu- The neighborhood of X i is determined by formula (5),
lated, the best value is recorded, and the concentra- which means that if the Chebyshev distance between X i
International Journal of Aerospace Engineering 5

Casing Trailing edge

𝜂4

Leading edge 𝜂3 Leading edge

𝜂2

𝜂1
𝜉1 𝜉2 𝜉3 𝜉4 𝜉5 𝜉6 𝜉7
Hub
Free active points
Limited active points

Figure 2: Schematic diagram of the point distribution of the full-blade parametric method.

Initialize the food source with

Output
Output thethe best
best value
value the optimal Latin method
Become an employed bee
YES

NO Employed bees explore

Exit or not
the food source
Initial the food source
NO Explore for detect bees
Is the new food source better times+1
Record the best
value
YES >Explore times
Replace the original food source limit Become a
detect bee
Calculate the selection
probability of the onlooker bees
Onlooker bees explore the food source

NO Explore
Is the new food source better
times+1

YES
Replace the original food source

Figure 3: Schematic diagram of the IABC algorithm.

and X t is less than or equal to the product of r (the neighbor- ðX Neib Þbest is calculated by
hood radius) and mdi (the average Chebyshev distance), then
X t locates in the neighborhood of X i ; otherwise, it does not:         
fit ðX Neib Þbest = max fit ðX Neib Þ1 , fit ðX Neib Þ2 ,⋯,fit ðX Neib ÞS ,

( ð10Þ
d ði, t Þ ≤ r ∗ mdi , t ∈ S,
ð9Þ
d ði, t Þ ≥ r ∗ mdi , t ∉ S, where fitðÞ refers to the individual fitness of each bee.
For the purpose of testing the performance of the above
algorithm, the ABC algorithm and the genetic algorithm
where md i is the average Chebyshev distance between X i and [34] (GA) are used for comparative analysis.
the entire onlooker bee population, S is the neighborhood of The algorithm parameters of IABC and ABC are set the
X i , and r is the radius of the neighborhood. When r is 1, the same, and the total number of bee colonies is 200, among
algorithm has the best convergence. which the number of employed bees is 100, the number of
6 International Journal of Aerospace Engineering

Table 1: Performance comparison of the IABC, GA, and ABC algorithms based on 5 benchmark functions [35].

Minimum Griewank Sphere Rastrigin Rosenbrock Zakharov

GA 1:05E00 2:42E − 02 3:08E + 01 1:94E + 02 8:20E + 01
ABC 3:91E − 01 4:30E − 03 1:95E + 01 2:53E + 01 2:35E + 02
IABC 2:98E − 10 3:38E − 12 4:12E − 10 2:71E + 01 1:55E + 02

onlooker bees is 100, the maximum number of explorations used to check the calculation to obtain the optimized solution
is 100, and the maximum number of iterations is 200. The under the high-fidelity flow field calculation.
GA algorithm has 200 chromosomes and 200 iterations and
a crossover probability of 0.8 and mutation probability of
0.005. Since these algorithms are probabilistic algorithms, 2.3.2. Multitask Concurrent Parallel Computing Based on the
the result of each calculation is uncertain, so the average min- IABC Algorithm and Supercomputing Platform. Since
imum value of 30 independent cycles is used for comparison. evolutionary algorithm optimization does not depend on
The comparison results are shown in Table 1. the gradient information of the function, it naturally pos-
Table 1 shows that the IABC algorithm has obvious sesses multitask concurrent computing capabilities. Using
advantages for the first three benchmark functions. The min- China’s advanced “Taihu Lake Supernatural Light” super-
imum value obtained is approximately 10 orders of magni- computing platform, a multitask concurrent optimization
tude smaller than that for the other two algorithms. For the system based on the IABC algorithm is built, which can
Rosenbrock function, the ABC algorithm is optimal, but the reduce the optimization time considerably. Figure 5 shows
IABC algorithm is at the same level, and the disadvantage is a schematic diagram of the multitask concurrent system
not obvious. For the Zakharov function, the GA is optimal, based on the IABC algorithm and the supercomputing plat-
but the IABC algorithm is suboptimal by less than an order form. On the user’s computer, the initial blade is used as
of magnitude; hence, the disadvantage is acceptable. the input of the program to start the optimization algorithm;
Figure 4 shows the convergence history of these bench- initialize the food source distribution; iteratively conduct
mark functions. It can be seen that IABC has a better and fas- employed bee, onlooker bee, and detector bee optimization;
ter convergence performance in the first four benchmark and record the location of the best food source for each bee
function tests than the other algorithms. colony generation (i.e., the optimal solution for each genera-
tion). If the loop exit condition has not been met, then the
2.3. Optimization Strategy. Engineering optimization often loop is continued. In any loop iteration, the initial food
does not require an optimal solution, but only an optimized source, employed bees, onlooker bees, and detector bees all
solution that is better than the original one. In this case, a adopt a concurrent mechanism, so the login node needs to
certain optimization strategy is needed to shorten the optimi- be started as an intermediary. Since there is an upper limit
zation cost. The advanced parametric methods and optimiza- for supercomputing concurrence, the number of concurrent
tion algorithms introduced in Sections 2 and 3 reduce the bee colonies must be limited to Max_Run; that is, Max_
dimensionality of the optimization task for the multistage Run bees are concurrently processed. After counting one
axial compressor, thus reducing the optimization calculation bee, the next new bee is started to join the parallel. The pro-
cost to a certain extent. However, considering the high cost of cess of concurrent bees is shown on the user’s computer as
the high-fidelity numerical calculation of multistage axial Max_Run. New geometry files are generated through the
compressors, it is necessary to further adopt an appropriate parametric model and then placed in the corresponding
optimization strategy to complete the optimization task within Max_Run new engineering task folder. At the login node,
the acceptable time range of the project. This article uses a multitask files are generated through scripts, and then, mul-
combination of the following three optimization strategies. titask files are assigned. The function of the multitask file is
to start a remote computing node to perform concurrent
2.3.1. Coarse Grid Optimization and Fine Grid Verification. computing. On the computing node, each task runs indepen-
The literature [36, 37] notes that the coarse mesh has been dently and must complete three steps, namely, grid genera-
able to capture the near wall viscosity, so coarse and fine tion, flow field calculation, and postprocessing analysis. On
meshes can predict the same trend of performance variation the login node, the script responsible for regularly detecting
during optimization. Although the computing power of com- the status of the computing node runs at all times. If an engi-
puter is improving, it is still insufficient for the high-precision neering task in the computing node ends, performance
design optimization of multistage axial compressor. There- results will be written in donelist.txt, and a line of donelist.txt
fore, it is still necessary to adopt certain optimization strate- will be added. At this time, by regularly synchronizing done-
gies to reduce the time consumption of single flow field list.txt on the user’s computer and the login node, the opera-
calculation. “Coarse mesh optimization, fine mesh verifica- tion status on the local user’s computer is known, and a
tion” is a common and effective strategy to reduce the time decision on whether to dispatch a new bee can be made.
consumption of single flow field calculation. The flow field Eventually, the entire optimization cycle reaches the set
calculation in the optimization process uses a coarser grid, number of iteration steps and exits the cycle to obtain the
and after obtaining the optimization solution, a fine grid is optimized blade for this case.
International Journal of Aerospace Engineering 7

250
800

700 200
600

500 150
Fitness

Fitness
400
100
300

200
50
100

0 0
50 100 150 200 50 100 150 200
Iteration count Iteration count
GA GA
ABC ABC
IABC IABC
(a) Griewank (b) Sphere

1E+06
400

800000
300

600000
Fitness

Fitness

200
400000

100
200000

0 0
50 100 150 200 0 50 100 150 200
Iteration count Iteration count
GA GA
ABC ABC
IABC IABC
(c) Rastrigin (d) Rosenbrock

Figure 4: Continued.
8 International Journal of Aerospace Engineering

600

500

400

Fitness
300

200

100
0 50 100 150 200
Iteration count
GA
ABC
IABC
(e) Zakharov

Figure 4: Comparison of the convergence history of the three algorithms based on the 5 benchmark functions.

User's computer Logged node Calculating node

Original blade
14 U 14 U 14 U
Create Max_Run new
geometry files
Initial food sources (new food sources) Create multi-task files

14 U 14 U 14 U
Employed bees Create Max_Run new
engineering task Distribute multitask files
folders
Onlooker bees Mesh generation
donelist.txt Synchronize Check the running status:
Detector bees 1) Model preparation
2) Grid state Flow field calculation
No = 0
Record the best food Number of change 3) Calculation status
source 4) Postprocessing status
Yes = N Postprocessing
No Satisfy the exit
requirement?
Create Max_Run new
geometry files (new 14 U 14 U 14 U
Yes donelist.txt
food sources)
Optimized blade

Create Max_Run new

14 U 14 U 14 U
engineering task
folders

Figure 5: Schematic diagram of the multitask concurrent system based on the IABC algorithm and supercomputing platform.

2.3.3. Phased Parameterization Strategy. To find the closest tion, which can improve the efficiency of space exploration in
possible global optimal solution of the design space within the initialization phase, but it has not yet been able to affect
the limited number of iterations, it is necessary to further the efficiency of the exploration and exploitation of the
study the exploration and exploitation strategy of the design design space in the optimization process. To solve this prob-
space. The IABC algorithm mentioned in the third section lem, drawing on the sample training in the surrogate model
uses a Latin square distribution instead of a random distribu- and the idea of “exploration first, exploitation later” of the
International Journal of Aerospace Engineering 9

Row1

Row2 Rowi

Row3 Flow field analysis

Synchronous Rowj
Independent
change
change
Row4 Blade rows and …
regions selection
Rowk
…

Row10
The first phase The second phase

Figure 6: Optimization strategy diagram.

0.83
4

0.82
Total pressure ratio

3.8
Adiabatic efficiency

0.81

3.6
0.8

3.4
0.79

3.2
0.78
106 108 110 112 114 106 108 110 112 114
Mass flow (kg/s) Mass flow (kg/s)
Exp Exp
Sim Sim
(a) Mass flow-pressure ratio (b) Mass flow-efficiency

Figure 7: Comparison of the experimental and calculated values of the compressor performance at the design speed.

evolutionary algorithm, this paper proposes a phased optimi- lyzed to identify the most promising blade rows and corre-
zation strategy based on the “synchronous change in multi- sponding blade regions (largest loss in this area). These
row blades”; that is, the aerodynamic optimization process blade rows and blade regions are selected for the second
of the multistage axial compressor is divided into two phases, phase of automatic optimization. In this analysis and selec-
as Figure 6 shows. tion process, the experience of the designer is incorporated.
The first phase: the “synchronous change in multirow The second phase: it is the independent optimization of
blades” strategy is adopted; that is, for the purpose of explor- the above selected potential blade rows. The aim is to build
ing the design space at a large step, each blade row of the on the results of the first phase of exploration to enable fur-
multistage axial compressor is modified in the same way ther exploitation of the corresponding design space to obtain
and by the same amount. After the optimization reaches a better performing blade shapes.
certain number of steps, which is determined according to The phased optimization strategy based on the “synchro-
the designer’s experience, the decision of when to end the nous change in multirow blades” can play a practical role in
first phase of the optimization exploration is made, and the the aerodynamic optimization of multistage axial compres-
optimized blade of the first phase is obtained. The first phase sors for the following two reasons:
focuses on the exploration of the design space.
After the first phase, the optimization process comes to a (1) The optimized solution only exists in a few small local
temporary end and the existing optimization results are ana- areas in the huge design space. In the evolutionary
10 International Journal of Aerospace Engineering

Table 2: Performance comparison between the AL-31F low-pressure compressor and advanced international low-pressure compressors.

Low-pressure compressor Total pressure ratio Number of stage Average pressure ratio Adiabatic efficiency Surge margin (%)
AL-31F 3.6 4 1.377 0.81 14.9
F100-PW-100 3.1 3 1.458 0.81 22
F100-PW-229 3.2 3 1.473 0.81 22
F110-GE 3.1 3 1.458 0.825
F404-GE-400 ≥3.7 3 ≥1.547 0.83

Tip:
Inlet: non-slip adiabatic wall
total pressure 101325 Pa
total temperature 293.15 K
air flow axial direction

S5
S4
S3 R4
S2 R3
R2
S1
R1
S0 Outlet:
average static
Outlet tandem blade pressure

Rotor/stator interface:
non-reflecting 2D

Rotation speed: 10,200 rpm

Inlet guide vane

Figure 8: Calculation domain and boundary condition setting for four-stage low-pressure compressor.

algorithm, the change in each blade row in each opti- compressor, with a total of ten rows of blades, as shown in
mization iteration is very small, and it is difficult to Figure 7. The inlet is a guide vane, and the outlet is tandem.
quickly explore the design space. If only the number The reason for using a tandem blade in the last stage of the
of variables of each optimization in the algorithm is stator blade is to ensure that the outlet airflow is along the
increased, then it is easy to increase the possibility of axial direction and at the same time to avoid large separation.
the mutual cancellation of the effects of different vari- If only one stator blade is used, the blade-camber angle of
ables, but it is also difficult to quickly explore the that stator blade will be too large, and thus, the airflow will
space. The “synchronous change in multirow blades” be prone to large separation at the back of the blade, causing
strategy can increase the pace of exploration and great losses and even stalling and surging. Compared with
greatly increase the probability of finding the opti- the performance of international advanced low-pressure
mized solution neighborhood within a limited num- compressors, as shown in Table 2, it can be seen that the aver-
ber of iterations age stage pressure ratio, adiabatic efficiency, and surge margin
of the AL-31F four-stage low-pressure compressors have a
(2) After the first phase, the designer’s experience is certain gap compared with other advanced low-pressure
utilized to analyze the flow field so that fewer optimi- compressors. Therefore, there is some room for improvement
zation variables need to be used for targeted optimi- in the AL-31F’s aerodynamic optimization design.
zation exploitation in the second phase, avoiding
invalid exploration of the design space and improv-
ing the optimization efficiency 3.2. Numerical Method Verification. A CFD numerical simu-
lation uses the Fine module of NUMECA to solve the three-
dimensional Reynolds average N-S equation in finite volume
3. Optimization Case Verification and Analysis form. The Autogrid5 module of NUMECA is used to gener-
ate the grid, and the grid topology of the flow passage and
3.1. Optimization Object. The optimization example used in blade is established automatically. In this test case, the model
this paper to verify the effectiveness of the above optimiza- without tip clearance and the Spalart-Allmaras one-equation
tion method is the AL-31F four-stage axial low-pressure turbulence model are adopted. The time discretization is
International Journal of Aerospace Engineering 11

Midspan

Trailing edge

Leading edge

(a) 3D blade grids (b) Grids of B2B

3.5 0.825

3.45
0.82
Total pressure ration

3.4

0.815 Efficiency
3.35

3.3 0.81

3.25
0.805

3.2
0 0.2 0.4 0.6 0.8 1 1.2 1.4
Number of grids (millions)
Efficiency
Pressure
(c) Verification of grid independence

Figure 9: Grid independence verification.

solved by the fourth-order explicit Runge-Kutta method, near stall point and the blocking point. The near stall point
while the false oscillation near the shock discontinuity and is the condition before calculating divergence that arises from
other small oscillations are controlled by the central differ- increasing the back pressure of the design point, and the
ence scheme and artificial viscosity. To accelerate calculation blocking point is the condition before calculating divergence
convergence, the local time step, the implicit residual fairing, that arises from reducing the back pressure of the design
and multigrid technique are utilized. point.
The calculation domain and boundary condition setting Figure 7 shows the comparison between the characteristic
for four-stage low-pressure compressor are shown in curve of the low-pressure compressor at the design speed
Figure 8. The total temperature at the inlet boundary is obtained by numerical simulation and the experimental data.
293.15 K, the total pressure is 101,325 Pa, and the inlet air At the design speed, the maximum pressure ratio measured
flow direction is axial. The boundary of the solid wall is adi- in the experiment is 4.05, the maximum pressure ratio
abatic and slip-free. Given the back pressure at the outlet obtained by simulation is 4.18, and the simulated value is
boundary, by gradually adjusting the outlet back pressure 3.2% larger than the experimental value. Figure 7(b) illus-
during the calculation process, it will advance towards the trates that the simulated value of the maximum efficiency is
12 International Journal of Aerospace Engineering

Original blade

IABC algorithm to Strategy 1: synchronous change in

Whether
optimize to meet the exit multirow blades
conditions
First phase

Full-blade surface Strategy 2: coarse grid optimization

parameterization and fine grid verification

Flow field Strategy 3: multitask concurrent

Mesh generation calculation parallel computing

Selection of optimized blades IABC algorithm to Whether

in the first phase optimize to meet the exit Optimized blade
Expert experience

conditions

Second phase
Flow field analysis of the Full-blade surface
selected optimized blade parameterization

Selection of optimization Flow field

Mesh generation
variables in the second phase calculation

Figure 10: Schematic diagram of the optimization process.

1.5% larger than the experimental value. At the design speed, food source position represents a new set of control parame-
the calculated pressure ratio range and efficiency range are ters. A new blade is generated under the full-blade surface
larger than the experimental values, and the flow margin is parametric control method, and then, mesh generation, flow
small. The simulated design speed margin is 10.2%, and the field calculation, and performance analysis are carried out for
experimentally measured margin is 14.9%. However, the cal- the new blade. If the requirements are met, the optimization
culated characteristic curve of the low-pressure compressor is will exit the cycle, and the optimized blade will be obtained;
consistent with the trend of the experimental curve, which otherwise, the iterative optimization of the cycle will be con-
proves the validity of the numerical simulation method. tinued. Usually, the decision regarding when to exit the cycle
To make the verification results more credible, grid inde- should depend on the designer’s experience.
pendence verification is performed on the optimized objects.
Figure 9(a) shows the grid of the three-dimensional blade, 3.4. Optimization Settings and Result Analysis. The three
Figure 9(b) shows the B2B grids of hub section, and optimization strategies described in Section 4 are combined
Figure 9(c) shows the calculation performance of four sets to conduct the optimization task of the four-stage low-
of grids under a backpressure of 300,000 Pa. The mesh qual- pressure compressor in two phases.
ity satisfies the near wall condition y+ < 2. As seen from
Figure 9(c), when the number of single channel grids in each 3.4.1. The First Phase: Optimization Based on the “Synchronous
row of blades reaches approximately 1 million, the changes in Change in Multirow Blades.”
efficiency and pressure ratio are very small, so it is reasonable
to believe that the third set of grids (on average, 930,000 per (1) Optimization Goals and Constraints in the First Phase.
single passage of a blade) has met the grid independence From the comparison of the above numerical simulation
requirements. However, by considering further shortening and experimental results, it can be seen that the efficiency
the optimization time, the “coarse grid optimization, fine grid and surge margin of the design speed of the AL-31F four-
verification” strategy is adopted, in which the first set of grids stage low-pressure compressor still have room for improve-
(average 250,000 per single passage of a blade) is used in the ment. Therefore, the optimization objective of the first phase
optimization process and the third set of grids is used for the is mainly to maximize the efficiency of the design point, tak-
flow field calculation of the optimization results. ing into account the surge margin improvement. According
to the author’s optimization experience, when 340,000 Pa of
3.3. Optimization Framework Construction. Figure 10 shows back pressure near the stall point is selected as the optimiza-
the optimization process. First, the flow field of the initial tion condition and the adiabatic efficiency under the back
blade is calculated. If the performance does not meet the pressure is taken as the objective function of the optimization
requirements, it enters into the optimization cycle. The IABC process, an optimization solution that can improve the
algorithm is adopted, which includes four steps: initializa- design point efficiency and surge margin can be obtained
tion, exploration of employed bees, exploration of onlooker under the condition that only one flow field calculation is
bees, and establishment of detector bees. Each step will pro- carried out in each optimization iteration. The flow rate con-
duce a corresponding number of new food sources, and each straint changes within 1%, and the total pressure ratio does
International Journal of Aerospace Engineering 13

not decrease, and the setting of the objective function and Table 3: Performance at the design point before and after
constraint conditions in the actual optimization process is optimization in the first phase.
shown in
Total
Design Mass flow Adiabatic Surge
pressure
maximize f ðx Þ point (kg/s)
ratio
efficiency margin

with respect to x Original

112.1 3.54 0.8207 10.2%
blade
subject to TPR − TPRori ≥ 0
  ð11Þ Optimized
112.7 3.55 0.8255 15%
mass‐massori  blade
  ≤ 1:0%
 mass  Relative
ori +0.54% +0.28% +0.48% +4.8%
change
xLi ≤ xi ≤ xU
i ,

where eff refers to the adiabatic efficiency under opti- in a row, a new food source needs to be generated randomly
mized conditions; TPR and TPRori refer to the total pressure again. On the supercomputing platform, approximately 850
ratio in the optimization process and the total pressure ratio calculations were carried out, and among them, the effective
of the original compressor, respectively; mass and massori point was approximately 800 (the invalid points that did
refer to the flow rate in the optimization process and the flow not meet the constraints were removed). Each optimization
rate of the original compressor, respectively; minus is the task needs 27 cores in parallel, and up to 24 tasks can be per-
minimum value given by experts; and xi refers to the optimi- formed concurrently (when the task is not queued after
zation variable, whose lower and upper limits are xLi and xU i , supercomputing). The optimization time of each optimiza-
respectively. The solutions that do not satisfy the constraints tion task is 30 min; this process takes approximately 30 hours
are directly excluded from the IABC program, and the next in total.
generation of bees will not converge towards this food source,
or the probability of convergence towards this source is (3) Optimization Results and Comparative Analysis in the
almost zero. In this way, the function of constraint can be First Phase. Table 3 shows the performance of the four-
realized in an unconstrained algorithm. stage low-pressure compressor at the design point with the
design speed. It can be seen that under the premise of the
The formulas for TPR, eff , and SM are as follows: “synchronous change in multirow blades” strategy in the first
phase, the optimized blade meets the constraints of the flow
Pout and pressure ratio, and the efficiency increases by 0.48%
TPR = , and the surge margin 4.8%.
Pin
TPRðk−1Þ/k − 1 As shown in Figure 11, the static pressure and Mach
eff = , ð12Þ
T out /T in − 1 number distribution in the hub, middle, and tip sections of
the optimized solution in the first phase are presented. The
TPRs /masss
SM = − 1 × 100%, losses in the R3 and R4 flow fields are not shown because they
TPR0 /mass0 are not significant. It is illustrated that the positive incidence
angles of airflow of R1 and R2 at the hub and middle sections
where Pout and Pin are the total pressure ratios at the outlet are too large; the airflow at the leading edge is accelerated,
and inlet, respectively; k is the specific heat ratio; T out and thus increasing the shock wave intensity; and the maximum
T in are the total temperature at the outlet and inlet, respec- Mach number at the middle reaches 1.4, leading to a corre-
tively; SM refers to the comprehensive surge margin; and sponding large loss at the tip section. At the same time, it
TPRs and TPR0 refer to the total pressure ratios at the surge can also be noted that there are two other relatively large
point and design point, respectively. low-speed areas: one is the tip trailing edge of S1 and the
other is the outlet of S5 at the tip, thus causing corresponding
(2) Optimization Settings in the First Phase. The full-blade loss. Therefore, these areas should be the target of the exploi-
surface parametric method is utilized, and all the blade rows tation optimization in the second phase.
are changed synchronously, with 16 design variables per row.
The four parameters controlling the leading edge are within 3.4.2. The Second Phase: Optimization Based on
[-5.0, 5.0] mm, and the 12 parameters controlling the blade Exploitation Strategy
body geometry are within [-6.0, 6.0] mm. In the IABC opti-
mization algorithm, the population size is 100, and the num- (1) Optimization Objectives and Constraints in the Second
ber of iterations is 8. The maximum number of mines is 3, Phase. According to the flow field analysis of the optimized
which means that the maximum number of times the next blade in the first phase, the control variables of the leading
generation of bees can explore at a particular previous gener- edge at the hub and the middle of R1 and R2, the leading edge
ation of food source cannot exceed three times. That is, if the at the tip of R3, the leading edge at the tip of S1, and the blade
calculated fitness is smaller than the original one three times body at the middle and tip of S5 are selected as the design
14 International Journal of Aerospace Engineering

Relative Mach number: 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3 1.4 1.5

1.4
1.4

R1
R2
Static pressure
Static pressure

(a) Hub

Relative Mach number: 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3 1.4 1.5

1.4
1.4

R1 R2
Static pressure Static pressure

(b) Middle

Relative Mach number: 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 1.1 1.2

0.2
0.1 0.3

S1 S5
(c) Tip

Figure 11: Static pressure and Mach number distribution at the hub, middle, and tip sections in the first phase.

variables of the second phase. The rest of the blade row is (2) Optimization Parameter Settings in the Second Phase.
fixed to the optimized blade obtained in the first phase. After the flow field calculation and analysis of the optimized
geometry obtained in the first phase, a total of five rows of
The optimization objectives, conditions, and constraints blade geometry of R1, R2, R3, S1, and S5 are selected as opti-
of the second phase are the same as those in the first phase. mized in the second phase. To reduce the optimization
International Journal of Aerospace Engineering 15

Casing Trailing edge

𝜂4

Leading edge Leading edge

𝜂3

𝜂2

𝜂1
𝜉1 𝜉2 𝜉3 𝜉4 𝜉5 𝜉6 𝜉7

Hub
Inactive points
Limited active points
(a) R1, R2

Casing Trailing edge

𝜂4

Leading edge Leading edge

𝜂3

𝜂2

𝜂1
𝜉1 𝜉2 𝜉3 𝜉4 𝜉5 𝜉6 𝜉7

Hub
Inactive points
Limited active points
(b) R3

Figure 12: Continued.

16 International Journal of Aerospace Engineering

Casing Trailing edge

𝜂4

Leading edge Leading edge

𝜂3

𝜂2

𝜂1
𝜉1 𝜉2 𝜉3 𝜉4 𝜉5 𝜉6 𝜉7
Hub
Inactive points
Limited active points
Active points
(c) S1

Casing Trailing edge

𝜂4

Leading edge Leading edge

𝜂3

𝜂2

𝜂1
𝜉1 𝜉2 𝜉3 𝜉4 𝜉5 𝜉6 𝜉7
Hub
Inactive points
Active points
(d) S5

Figure 12: Distribution diagram of the control points of each selected blade in the second phase.

Table 4: Time consumption comparison of different optimization methods.

Number of control Number of flow field Single flow field Whether

Total time cost
parameters calculations calculation time concurrent
Direct evolutionary No >5000 h
>100 >10000 30 min
optimization methods Yes >421 h
No >1500 h
Surrogate model method >100 3000 30 min
Yes >126
No >1000 h
DOE method >100 2000 30 min
Yes >84 h
Method proposed in Phase 1: 16 Phase 1: 850 30 h + 48 h = 78 h
30 min Yes
this paper Phase 2: 25 Phase 2: 1000
International Journal of Aerospace Engineering 17

Table 5: Performance comparison before and after optimization at the design point in the two phases.

Design point Mass flow (kg/s) Total pressure ratio Adiabatic efficiency Surge margin
Original blade 112.1 3.54 0.8207 10.2%
Optimized blade of the first phase 112.7 3.55 0.8255 15%
Optimized blade of the second phase 112.2 3.55 0.8274 13.3%
Relative change +0.08% +0.28% +0.67% +3.1%

4
0.83

3.8 0.82

Adiabatic efficiency
Total pressure ratio

0.81
3.6

0.8
3.4
0.79

3.2 0.78
111 111.5 112 112.5 113
111 111.5 112 112.5 113
Mass flow (kg/s) Mass flow (kg/s)
Original Original
Optimal Optimal

(a) Mass flow-total pressure ratio (b) Mass flow-efficiency

Figure 13: Comparison of the performance between the optimized blade and original blade in the second phase.

variables for R1, R2, R3, and S1, the leading edge control task is 30 min, and the optimization takes approximately 48
parameters are optimized to reduce the positive incidence hours in the second phase.
angle. To reduce the optimization variables for S1 and S5, The time consumption of the optimization method pro-
the blade body is controlled in addition to the leading edge posed in this paper is compared with the conventional
variables. Refer to Figure 12 for the specified control variables method in Table 4. From Table 4, it can be seen that the
of each row: all the green points are leading edge control proposed optimization method can greatly reduce the
points, which are variable in the radial direction, but the four number of control variables in the optimization process, thus
points of the same row must keep pace to avoid distortion at greatly reducing the times of flow field calculations required.
the leading edge; all the red points are blade-independent At the same time, the optimization time can be reduced by
control points, and all the black points are fixed points. an order of magnitude due to the construction of a
supercomputing-based concurrent system. Although the
In the second phase, the optimization parameters are set optimization time of the DOE (design of experiment) method
as follows: the full-blade surface parametric method is still is not too long, it only has the function of global exploration
adopted, with a total of 25 optimization variables. The 3 con- and does not have the ability of local exploitation. The
trol parameters of the leading edge of R1 and R2 and the 2 proposed method in this paper has the shortest optimization
control parameters of the leading edge at the hub of R3 all consumption time among several methods.
change within the range [-6.0, 0.0] mm. The variation range
of the 8 control parameters of S1 and 9 control parameters (3) Comparative Analysis of the Optimization Results in the
of S5 is within [-6.0, 6.0] mm. The number of iterations is Second Phase. Since a single-objective optimization (adiabatic
10, the size of the bee colony in the IABC optimization algo- efficiency) is adopted in the second phase, it cannot find a
rithm is 100, and the maximum number of exploitations is 3. solution that is superior to the optimized solution found in
On the supercomputing platform, a total of approximately the first phase in terms of adiabatic efficiency and the surge
1,000 calculations are performed, and the number of effective margin in the results. According to the designer’s experience
points is approximately 980 (without the invalid points vio- and needs, the point where the adiabatic efficiency is most
lating the constraints). Each optimization task requires 27 improved and the surge margin is slightly decreased is selected
cores in parallel, and a maximum of 24 tasks can be per- as our final optimized solution. Table 5 shows a performance
formed concurrently (under the condition that the tasks are comparison of the optimization design point before and after
not queued up). The optimization time of each optimization optimization in the second phase. It can be seen from the table
18 International Journal of Aerospace Engineering

Ori_hub Ori_mid Ori_tip

Opt_hub Opt_mid Opt_tip
(a) R1

Ori_hub Ori_mid Ori_tip

Opt_hub Opt_mid Opt_tip
(b) R2

Ori_hub Ori_mid Ori_tip

Opt_hub Opt_mid Opt_tip
(c) R3

Ori_hub Ori_mid Ori_tip

Opt_hub Opt_mid Opt_tip
(d) R4

Figure 14: Continued.

International Journal of Aerospace Engineering 19

Ori_hub Ori_mid Ori_tip

Opt_hub Opt_mid Opt_tip
(e) S1

Ori_hub Ori_mid Ori_tip

Opt_hub Opt_mid Opt_tip
(f) S2

Ori_hub Ori_mid Ori_tip

Opt_hub Opt_mid Opt_tip
(g) S5

Figure 14: Geometric changes before and after optimization.

that after two phases of optimization, the mass flow rate and them, R4 and S2 underwent only one optimization, and the
pressure ratio at the design point of the optimized blade are rest of the blade deformations that underwent only one
almost unchanged, while the adiabatic efficiency increases by optimization were similar to them. R1, R2, R3, S1, and S5
0.67% and the surge margin 3.1%. Comparing the optimiza- underwent two optimizations.
tion results of the first phase and the second phase, it can be From Figures 14(d) and 14(f), it can be seen that the hub
seen that the adiabatic efficiency of the design point in the and tip shape of the optimized R4 becomes thinner and the
second phase increases by 0.19%, while the surge margin blade-camber angle increases slightly; the leading edge part of
decreases by 1.7%. the middle airfoil is basically unchanged, while the blade angle
at the trailing edge increases slightly; thus, the blade-camber
Figure 13 shows a performance comparison between the angle increases slightly; while the optimized S2 is the opposite
optimized blade of the second phase and the original one. of R4, after optimization, the hub and tip airfoils of S2 become
The change in the flow rate of the four-stage low-pressure thicker and the blade-camber angle becomes smaller, while the
compressor after optimization is almost 0. This is a common blade-camber angle of the middle airfoil decreases slightly.
situation in multistage axial flow compressors. This situation From Figure 14(a), it can be seen that the hub blade-
indicates that although the back pressure is changing, there is camber angle of R1 slightly increases, and the middle and
a flow rate restriction in one or more stages in the multistage tip airfoils do not change much with only a slight normal
compressor, probably due to the fact that it has stalled there. shift. From Figures 14(b) and 14(c), it can be seen that the
We tend to expand the flow surge margin by adjusting the hub blade-camber angle of the optimized R2 and R3
angle of multiple rows of stator blades in the normal operat- increases significantly and the leading edge blade angle
ing condition of a multistage axial compressor. Our optimi- decreases. The middle airfoil is similar to the hub, except that
zation process does not consider the adjustment of the the degree is reduced.
angle of the stator blades, so it appears as a straight line on For S1, the optimized blade-camber angle at the hub,
the pressurizer characteristic graph. The optimized blade middle, and tip regions was reduced, most significantly at
can maintain a high flow rate even under higher back pres- the tip regions, as shown in Figure 14(e). For S5, the
sure and thus improve its surge margin. optimized blade-camber angles at the hub and the tip were
Each blade row changed before and after optimization. significantly reduced, most significantly at the tip, and the
Figure 14 shows the geometric comparison of R1, R2, R3, blade-camber angle at the middle increased, as shown in
R4, S1, S2, and S5 before and after optimization. Among Figure 14(g).
20 International Journal of Aerospace Engineering

Ma 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3 1.4

0.8 0 0.8
0.2
1.4 1.4

(a) Hub of the original R2 (b) Hub of the optimized R2

0 0.2
1.4 1.2

(c) Hub of the original R3 (d) Hub of the optimized R3

1.4 1.4

(e) Midspan of the original R3 (f) Midspan of the optimized R3

Ma
0 0.04 0.08 0.12 0.16 0.2 0.24 0.28 0.32 0.36 0.4

0.24 0.28
0.4 0.28 0.4 0.36 0.32
0.4

(g) Tip of the original S5 (h) Tip of the optimized S5

Figure 15: Comparison of the relative Mach number at the hub of R2, R3 and at the tip of S5 before and after optimization.
International Journal of Aerospace Engineering 21

3.5

Static pressure
2.5
R4 S5
2 S4

1.5
S2 R3 S3
1 S1
R2
R1
0.5 S0

1.5 2 2.5 3 3.5

Axial direction
Ori
Opt
(a) Hub

3.5

3
Static pressure

2.5
S5
2 S4

1.5 R4
S3
1
S2
R3
R1 S1 R2
0.5 S0
1 1.5 2
Axial direction
Ori
Opt
(b) Middle

3.5

3
Static pressure

2.5
S5
S4
2 R4
S3
1.5
R3
1 S1 S2
R1 R2
S0
0.5
0.5 1 1.5
Axial direction
Ori
Opt
(c) Tip

Figure 16: Comparison of the static pressure distribution before and after optimization.
22 International Journal of Aerospace Engineering

Wxyz (m/s)
100

80

60

40

20

(a) Original R1 (b) Optimized R1 (c) Original R2 (d) Optimized R2

Reattachment
line

Separation
line

(e) Original R3 (f) Optimized R3 (f) Original R4 (g) Optimized R4

Figure 17: Comparison of limit streamlines on the suction surfaces before and after optimization.

(4) Comparative Analysis in the Second Phase. The flow field As a result, the flow in the channel is smoother, thereby reduc-
before and after optimization should be compared and ana- ing the airflow loss in the S5 outlet area.
lyzed. However, it is difficult to highlight the details of the
local flow field if all of them are displayed indiscriminately. The limit streamlines on the suction side of R1, R2, R3,
Therefore, in order to better reveal the reason for the perfor- and R4 are shown in Figure 17. From Figures 17(a)–17(d),
mance improvement of the optimized blade, typical blade it can be seen that for R1 and R2, the changes before and after
rows are chosen for flow field analysis. The flow field changes optimization are not significant; for R3, there are large
in the four rotor blades are similar, and the flow field details changes before and after optimization. The closed separation
of the typical blade row regions are selected to show and ana- area at the tip of the optimized R3 is significantly reduced
lyze the changes before and after optimization. Figure 15 after optimization, thus reducing the corresponding loss. In
shows a comparison of the relative Mach number distribu- addition, the separation line of the optimized R3 is shortened
tion at the hub, midspan, and tip of R2, R3, and S5. in the radial direction and thus reduce the radial separation
Figure 16 illustrates the static pressure distribution of each area. At the same time, it can be observed that the optimized
row of blades at the hub, midspan, and tip. Combined with R3 will be attached again after separation, and then, a second
Figures 14–16, it can be seen that although the blade- separation line will be formed, reducing the separation area
camber angle of the hub and mid airfoil shapes of the along the flow direction and thus reducing the corresponding
optimized R2 and R3 becomes larger, the inverse pressure separation loss. For R4, the optimized R4 has an additional
gradient near the trailing edge is reduced due to the change reflux region at the tip, which increases the loss at the tip.
in load and slightly thinner thickness, resulting in a signifi- For R4, original R4 has an additional reflux region at the
cant reduction in the airflow separation area at this location tip, which increases the loss at the tip.
and a reduction in the airflow loss there. The situation of R3 Figure 18 shows the entropy distribution at midspan and
is basically the same as that of R2. As seen from Figures 15(g) tip section before and after optimization, from which it can
and 15(h), as the blade-camber angle of the optimized tandem be seen that the entropy value of the airflow is significantly
S5 tip is reduced, the back pressure gradient near the trailing lower after optimization (such as regions A, B, C, D, E, and
edge is reduced, and hence, the low-speed region caused by air- F), which indicates a smoother airflow and a corresponding
flow mixing at the trailing edge of S5 is significantly reduced. reduction in losses.
International Journal of Aerospace Engineering 23

Original
A B C

Optimized
A B C

Entropy (J/(K•mol))
0 10 20 30 40 50 60 70 80 90 100
(a) Midspan
Original

D
E F

Optimized
D
E F

Entropy (J/(K•mol))
0 30 60 90 120 150 180 210 240 270 300

(b) Tip

Figure 18: Comparison of entropy distribution at the midspan and tip section before and after optimization.

Based on the above analysis, it can be seen that after two stage low-pressure compressor aerodynamics. The following
phases of optimization, the R2 and R3 blade has a reduced conclusions are drawn:
reverse pressure gradient near the trailing edge at the hub
and middle, resulting in a significant reduction in the separa- (1) Compared with the traditional parametric method, the
tion area after the shock wave, and a separation bubble for full-blade surface parametric method can effectively
reattachment is formed under these conditions (shown as reduce the dimensionality and promote the solution
optimized R3), thus reducing the corresponding loss and of the HEB problem for a multistage axial flow
improving the efficiency. The reverse pressure gradient near compressor. At the same time, compared with the
the trailing edge of the S5 blade is also reduced due to the semiblade surface parametric method, the full-blade
reduction of the blade-camber angle at the tip, narrowing surface parametric method increases the degree of free-
the low-speed area near the trailing edge and downstream dom of the blade angle change, therefore enhancing the
and reducing the corresponding airflow loss. probability of the existence of an optimal solution

4. Conclusion (2) After verifying on the benchmark function, the IABC

algorithm is shown to have better optimization per-
This paper proposes a new optimization method for high- formance than the GA and ABC algorithms. The
fidelity global optimization of multistage axial compressors initialization method and exploration mechanism
based on the full-blade surface parameterization. The new are applicable to the multipeak problem of multistage
method is applied to the optimization of the AL-31F four- axial compressor optimization
24 International Journal of Aerospace Engineering

j
(3) The phased optimization strategy based on the ξi : Random numbers
“synchronous change in multirow blades” achieves j
λi : Random numbers
a good balance between global exploration and local j The best food source location in the general
development. By introducing expert experience, the X best :
information of the food source
exploration of an invalid design space by optimiza- N e: Number of employed bees
tion algorithms is avoided, and the optimization
P: Probability of selecting a certain food source
efficiency is greatly improved
f ðX i Þ: Concentration of the ith food sources
(4) The new optimization method based on the full- f ðX m Þ: Concentration of the mth food sources
j
blade parameterization can be successfully applied ðX Neib Þbest : Food source position that has the largest
to the AL-31F four-stage low-pressure compressor. concentration in the adjacent area
When the flow and pressure ratio meet the con- j
ðV Neib Þbest : The location of the new food source of onlooker
straints, the adiabatic efficiency and surge margin bees
are increased by 0.67% and 3.1%, respectively dði, tÞ: Chebyshev distance between X i and X t
md i : Average Chebyshev distance between X i and the
Abbreviations entire onlooker bee population
r: Radius of the neighborhood
HEB: High dimensionality, expensive cost, and black box fitðÞ: Individual fitness of each bee
PDE: Partial differential equations Pout : The total pressure ratios at the outlet
CST: Class-shape function transformation Pin : The total pressure ratios at the inlet
IABC: Improved artificial bee colony T out : The total temperature at the outlet
ABC: Artificial bee colony T in : The total temperature at the inlet.
GA: Genetic algorithm
3D: Three dimensional Data Availability
Ori: Original
Opt: Optimized The geometry and performance data used to support the
mm: Millimeter findings of this study are included within the article.
rad: Radian
TPR: Total pressure ratio Conflicts of Interest
eff: Efficiency
SM: Surge margin The authors declare that they have no conflicts of interest.
DOE: Design of experiment
Max_Run: The maximum number of tasks that can be Acknowledgments
concurrent at one time.
This study was supported by the National Major Science and
Nomenclature Technology Projects of China (CN) (2017-II-0010-0024).
ξ: Chordwise direction
η: Spanwise direction
np : Number of data points
References
ns : Total number of sections [1] J. Cheng, C. Jiang, and H. Xiang, “A surface parametric control
T l j: Sum of the chord lengths of each segment of the and global optimization method for axial flow compressor
jth section in the η direction blades,” Chinese Journal of Aeronautics, vol. 32, no. 7,
T Li : Sum of the chord lengths of the ith section in the pp. 1618–1634, 2019.
ξ direction [2] J. Chen and L. Ji, “Aerodynamic characteristic optimization of
lm : The mth chord length of the jth section in the η a multistage axial flow compressor [J],” Journal of Engineering
direction Thermophysics, vol. 28, no. 4, p. 583, 2007.
Ln : The nth chord length of the ith section in the ξ [3] Z. Q. An, Z. G. Zhou, and Y. H. Zhang, “Axial compressor
direction blade automatic optimization design based on a quasi three-
! The change value of each point on the perturbed dimensional approach [J],” Journal of Propulsion Technology,
R: vol. 35, no. 4, pp. 485–491, 2014.
surface
P k ,l : ðm + 1Þ × ðn + 1Þ control points of the perturbed [4] W. Ding, B. Liu, Z. P. Cao, and Y. Y. CHEN, “Optimization
design of multistage axial flow compressor using multiobjec-
surface
tive genetic algorithm [J],” Journal of Propulsion Technology,
Bml ðvÞ: Bernstein basis functions vol. 27, no. 3, p. 230, 2006.
Bnk ðuÞ: Bernstein basis functions [5] W. Tuo, J. S. Xiong, A. P. Hou, and S. Zhou, “Application of
v, u: Independent coordinate variables of the per- genetic algorithm to multistage compressor aerodynamic opti-
turbed surface mization design [J],” Journal of Aerospace Power, vol. 22, no. 2,
C nk : Combination number pp. 305–309, 2007.
Xi: Feasible solution [6] C. Ma, X. Su, and X. Yuan, Discrete Adjoint Solution of
j The jth component of the food source location
Vi: Unsteady Turbulent Flow in Compressor [C]//ASME Turbo
of the ith bee Expo 2015: Turbine Technical Conference and Exposition,
International Journal of Aerospace Engineering 25

American Society of Mechanical Engineers Digital Collection, [23] S. Shan and G. G. Wang, Survey of modeling and optimization
2015. strategies for high-dimensional design problems [C]//12th
[7] X. J. Liu and F. F. Ning, “Three-dimensional aerodynamic opti- AIAA/ISSMO multidisciplinary analysis and optimization con-
mization of intermediate stator of 7-stage axial compressor [J],” ference, p. 5842, 2008.
Journal of Aerospace Power, vol. 27, no. 8, pp. 1826–1831, 2012. [24] J. Y. Trepanier, A. Lupien, and C. Tribes, “A 3D parameteriza-
[8] Z. Gao, X. Gao, and X. Yuan, “Aerodynamic optimal design of tion for transonic fan blade multidisciplinary design,” Aeronau-
non-axisymmetric endwall for a turbine cascade [J],” Journal tics and Aerospace Open Access Journal, vol. 1, no. 1, 2017.
of Engineering Thermophysics, vol. 28, no. 4, p. 589, 2007. [25] C. Lee, A Comparison of b-Spline Surface and Free-Form Defor-
[9] H. Yu and X. Yuan, “Optimized aerodynamic design of multi- mation Geometry Control Methods for Aerodynamic Shape
blade rows of an axial compressor [J],” Reneng Dongli Gong- Optimization [M], University of Toronto, 2015.
cheng (Journal of Engineering for Thermal Energy and Power), [26] C. Lee, D. Koo, and D. W. Zingg, “Comparison of B-spline
vol. 20, no. 6, pp. 603–606, 2005. surface and free-form deformation geometry control for
[10] W. S. Ma, Investigation of multistage axial-compressor aerody- aerodynamic optimization,” AIAA Journal, vol. 55, no. 1,
namic optimization design [dissertation], Tsinghua University, pp. 228–240, 2017.
Beijing, 2009. [27] T. A. Reist, D. Koo, D. W. Zingg, P. Bochud, P. Castonguay,
[11] P. Champasak, N. Panagant, N. Pholdee, S. Bureerat, and A. R. and D. Leblond, “Cross validation of aerodynamic shape opti-
Yildiz, “Self-adaptive many-objective meta-heuristic based on mization methodologies for aircraft wing-body optimization,”
decomposition for many- objective conceptual design of a AIAA Journal, vol. 58, no. 6, pp. 2581–2595, 2020.
fixed wing unmanned aerial vehicle,” Aerospace Science and [28] C. Liu, Y. Duan, J. Cai, and J. Wang, “Application of the 3D
Technology, vol. 100, p. 105783, 2020. multi-block CST method to hypersonic aircraft optimization,”
[12] J. Li and J. Cai, “Massively multipoint aerodynamic shape Aerospace Science and Technology, vol. 50, pp. 295–303, 2016.
design via surrogate-assisted gradient-based optimization,” [29] S. Burguburu and A. le Pape, “Improved aerodynamic design
AIAA Journal, vol. 58, no. 5, pp. 1949–1963, 2020. of turbomachinery bladings by numerical optimization,” Aero-
[13] L. Baert, E. Chérière, C. Sainvitu, I. Lepot, A. Nouvellon, and space Science and Technology, vol. 7, no. 4, pp. 277–287, 2003.
V. Leonardon, “Aerodynamic optimization of the low- [30] G. Wang and Z. G. Zhou, “Aerodynamic optimization for fan
pressure turbine module: exploiting surrogate models in a rotor design using genetic algorithm,” Civil Aircraft Design &
high-dimensional design space,” Journal of Turbomachinery, Research, vol. 1, pp. 13–19, 2011.
vol. 142, no. 3, 2020.
[31] C. Jinxin, J. Chen, and H. Xiang, “Aerodynamic optimization
[14] Y. He, J. Sun, P. Song, X. Wang, and A. S. Usmani, “Preference- design of compressor blades based on improved artificial bee
driven Kriging-based multiobjective optimization method with colony algorithm,” in Proceedings of the 2018 Joint Propulsion
a novel multipoint infill criterion and application to airfoil shape Conference, p. 4825, Cincinnati, OH, USA, 2018.
design,” Aerospace Science and Technology, vol. 96, p. 105555,
2020. [32] D. Karaboga and B. Akay, “A comparative study of Artificial
Bee Colony algorithm,” Applied Mathematics and Computa-
[15] J. Müller, Surrogate model algorithms for computationally
tion, vol. 214, no. 1, pp. 108–132, 2009.
expensive black-box global optimization problems, Tampere
University of Technology, 2012. [33] J. S. Park, “Optimal Latin-hypercube designs for computer
experiments,” Journal of Statistical Planning and Inference,
[16] G. Yang, A. da Ronch, J. Drofelnik, and Z. T. Xie, “Sensitivity
vol. 39, no. 1, pp. 95–111, 1994.
assessment of optimal solution in aerodynamic design optimi-
sation using SU2,” Aerospace Science and Technology, vol. 81, [34] D. Whitley, “A genetic algorithm tutorial [J],” Statistics and
pp. 362–374, 2018. Computing, vol. 4, no. 2, pp. 65–85, 1994.
[17] R. Lei, J. Bai, and D. Xu, “Aerodynamic optimization of civil air- [35] M. Molga and C. Smutnicki, “Test functions for optimization
craft with wing-mounted engine jet based on adjoint method,” needs [J],” Test functions for optimization needs, vol. 101,
Aerospace Science and Technology, vol. 93, p. 105285, 2019. p. 48, 2005.
[18] A. L. Kaminsky and K. Ekici, “Reduced-order model-based con- [36] J. D. Denton, “Designing in Three Dimensions,” in Turboma-
vergence acceleration of reverse mode discrete adjoint solvers,” chinery design using CFD, AGARD, Neuilly sur Seine, France,
Aerospace Science and Technology, vol. 93, p. 105334, 2019. 1994.
[19] J. Luo, “Design optimization of the last stage of a 4.5-stage [37] S. M. Miner, “Evaluation of blade passage analysis using coarse
compressor using a POD- based hybrid model,” Aerospace grids,” Journal of Fluids Engineering, vol. 122, no. 2, pp. 345–
Science and Technology, vol. 76, pp. 303–314, 2018. 348, 2000.
[20] M. Schnoes, A. Schmitz, and G. Goinis, “Strategies for multi-
fidelity optimization of multistage compressors with through-
flow and 3D CFD,” in Proceedings of the International Society
for Air Breathing Engines, Canberra, Australia, 2019.
[21] Z. Li and X. Zheng, “Review of design optimization methods
for turbomachinery aerodynamics,” Progress in Aerospace Sci-
ences, vol. 93, pp. 1–23, 2017.
[22] S. Shan and G. G. Wang, “Survey of modeling and optimiza-
tion strategies to solve high-dimensional design problems with
computationally-expensive black-box functions,” Structural
and Multidisciplinary Optimization, vol. 41, no. 2, pp. 219–
241, 2010.