Performance Enhancement of Axial Fan Blade Through Multi-Objective Optimization Techniques
Performance Enhancement of Axial Fan Blade Through Multi-Objective Optimization Techniques
Performance Enhancement of Axial Fan Blade Through Multi-Objective Optimization Techniques
www.springerlink.com/content/1738-494x
DOI 10.1007/s12206-010-0619-6
(Manuscript Received March 1, 2010; Revised June 7, 2010; Accepted June 11, 2010)
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Abstract
This paper presents an axial fan blade design optimization method incorporating a hybrid multi-objective evolutionary algorithm (hy-
brid MOEA). In flow analyses, Reynolds-averaged Navier-Stokes (RANS) equations were solved using the shear stress transport turbu-
lence model. The numerical results for the axial and tangential velocities were validated by comparing them with experimental data. Six
design variables relating to the blade lean angle and the blade profile were selected through Latin hypercube sampling of design of ex-
periments (DOE) to generate design points within the selected design space. Two objective functions, namely, total efficiency and torque,
were employed, and multi-objective optimization was carried out, to enhance the performance. A surrogate model, Response Surface
Approximation (RSA), was constructed for each objective function based on the numerical solutions obtained at the specified design
points. The Non-dominated Sorting of Genetic Algorithm (NSGA-II) with local search was used for multi-objective optimization. The
Pareto-optimal solutions were obtained, and a trade-off analysis was performed between the two conflicting objectives in view of the
design and flow constraints. It was observed that, by the process of multi-objective optimization, the total efficiency was enhanced and
the torque reduced. The mechanisms of these performance improvements were elucidated by analysis of the Pareto-optimal solutions.
Keywords: Axial fan blade; Evolutionary algorithm; Surrogate model; Pareto-optimal solutions; Total efficiency; Torque
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
zation of a low-speed fan cascade using Bezier-curve-selected Table 1. Design specifications of axial fan.
design variables including thickness distribution and camber
Flow Coefficient 0.41
line. Benini [16] used multi-objective optimization and a
Total Pressure Coefficient 0.30
Bezier curve to define blade-section profiles in consideration
of the total pressure ratio and adiabatic efficiency, and used Rotor Rotation Frequency, rpm 1000
these as design objectives for an axial compressor blade. He Tip Radius, mm 287.5
employed the camber line and a thickness profile as the design Number of blades 9
parameters. Lastly, Pierret et al. [17] reported multi- Hub/Tip Ratio 0.52
disciplinary and multiple-operating-point optimizations for Inlet Angle at Rotor Tip, degrees 68.8
fixed rotor speeds. Outlet Angle at Rotor Tip, degrees 63.8
In the present work, a hybrid multi-objective evolutionary
algorithm (hybrid MOEA) [18] coupled with the RSA model
was applied to obtain a global Pareto-optimal front for design
of an axial fan blade. Three-dimensional RANS analyses were
performed to obtain numerical solutions for the selected de-
sign points. The blade shape was optimized by six design
variables relating to the blade lean angle and blade profile.
Two conflicting objectives, that is, the total efficiency and the
torque, were selected for optimization.
2. Numerical analysis
Reynolds-averaged Navier-Stokes (RANS) equations were Fig. 1. Computational domain and structure of grid system.
solved using the Shear Stress Transport model as turbulence
closure. The commercial software ANSYS CFX 11.0 [19]
consisting of Blade-Gen, Turbo-Grid, CFX-Pre and CFX- computational domain. A grid-independency test had been
Solver applications, was employed in flow analyses. The carried out to determine the optimal grid number, and a grid of
blade profile was generated in Blade-Gen and exported to 630,000 points was selected. Fig. 1 shows the grid system
Turbo-Grid for mesh generation. The meshed geometry was structure.
then imported into CFX-Pre for boundary and initial condition The solutions are made to converge; among the conver-
definitions, and finally, the model was run in CFX-Solver, gence criteria, the root mean square (RMS) residual values of
which solves three-dimensional steady incompressible RANS the momentum and mass were set below 1.0E-06, and the
equations. The code uses a finite volume solver that utilizes imbalances of mass and energy were maintained below 1.0E-
the coupled algebraic multigrid (AMG) method [19]. 03. The converged solutions were obtained after approxi-
An important issue in turbulence modeling is the formula- mately 500 iterations. The computations were carried out us-
tion of near-wall turbulence, which determines the accuracy of ing a PC with an Intel Pentium IV CPU with a processing
the wall shear stress formulation. The shear stress turbulence speed of 3.0 GHz. The computation time according to the
(SST) model is used as a turbulence closure. The SST model, geometry considered and the rate of convergence was 8-9
having the advantages of both k-ω and k-ε, employs the k-ω hours.
model at the near-wall and the k-ε model in the bulk flow
regions, a blending function ensuring a smooth transition be-
3. Design variables and objective functions
tween the two models. Recently, the SST model has been
regarded among two-equation turbulence closures as the most Six design variables were selected from among the various
accurate model for aerodynamic applications. blade profile parameters affecting axial fan performance.
The computational domain in which the present simulations These were airfoil maximum camber (α), maximum camber
were performed for a single passage of an axial fan is shown location (β), leading edge (LE) radius (γ), TE radius (δ), lean
in Fig. 1. The simulations performed were steady-state, and angle at mid-span (κ) and lean angle at tip span (ξ). The other
the working substance was 25°C air. The total pressure at the relevant parameters were kept constant. The blade profile was
inlet was set to 1.0atm, and at the outlet, a designed mass flow varied by third-order Bezier curve, as shown in Fig. 2(a). The
rate for a single passage was established. The solid surfaces main advantage of blade curve parameterization by Bezier
were considered to be hydraulically smooth under no-slip and curves is that a certain limited number of points (the “control
adiabatic conditions, the periodic conditions were set at the points”) can be used to control the curves, rendering them
blade passage interfaces, and the tip clearance modeled was smooth and free of discontinuities. By moving these control
2.0 mm. The major design specifications are listed in Table 1. points, which are considered as design variables, the blade
A hexahedral grid system was employed as the mesh in the shape can be varied. In the present problem, some of the con-
J.-H. Kim et al. / Journal of Mechanical Science and Technology 24 (10) (2010) 2059~2066 2061
approximate Pareto-optimal solutions are obtained using the Table 2. Ranges of design variables.
real-coded NSGA-II developed by Deb [18] for the two objec-
Variables Lower Upper
tive functions, total efficiency and torque. Here, real-coded
α (mm) -4.0 4.0
indicates that the crossover and mutations are conducted in
β (mm) -12.0 12.0
real space to obtain a response from NSGA-II. The Pareto-
optimal solutions are then refined by searching for a local γ (mm) -0.5 1.0
optimal solution for each objective function over all NSGA-II- δ (mm) -0.75 0.5
derived optimal solutions; the search uses Sequential Quad- κ (deg) -15.06 16.94
ratic Programming (SQP) [22] with NSGA-II solutions as ξ (deg) -6.85 10.15
initial guesses. SQP is a generalization of Newton’s method,
which is a gradient-based optimization technique. To perform Table 3. Results of ANOVA and regression analysis.
a local search, usually two approaches are applied [18]. In one
Cross-
approach, all objectives are combined into a single composite Objective
R2 R2adj RMSE validation
objective, and the optimum is searched. In the other approach, functions
errors
one objective is optimized by treating the others as equality η 0.969 0.920 1.03E-3 1.97E-3
constraints, and the process is repeated for all objectives. τ 0.999 0.998 3.92E-4 7.15E-4
In the present study, the first objective was optimized, and
the second objective was treated as an equality constraint. The 0.7
local search was repeated for the second objective function by Axial-TEST
0.6 Axial-RANS
treating the first as an equality constraint. This process yielded Tangential-TEST
two new sets of optimal solutions, which were then merged 0.5 Tangential-RANS
with the NSGA-II solutions. From these solutions, the domi- Vt / Ut, Va / Ut 0.4
nated solutions were discarded and then the duplicate solu-
0.3
tions were removed so as to obtain the global Pareto-optimal
solutions. Subsequently, the process of local search improved 0.2
MOEA CFD
Design variables Increment
Design prediction calculation
α (mm) β (mm) γ (mm) δ (mm) κ (deg) ξ (deg) τ η τ η τ (%) η (%P)
Reference 0 0 0 0 0 0 - - 0.474 0.8557 - -
Cluster A 3.51 0.04 0.32 -0.75 16.94 8.46 0.297 0.8193 0.283 0.8051 67.71 ↓ 5.06 ↓
Cluster B 1.67 0.04 0.29 -0.75 16.40 10.14 0.342 0.8396 0.332 0.8353 42.93 ↓ 2.04 ↓
Cluster C -0.69 0.04 0.36 -0.75 10.94 10.15 0.398 0.8566 0.397 0.8541 19.48 ↓ 0.16 ↓
Cluster D -2.30 0.04 0.73 -0.53 9.19 10.15 0.436 0.8629 0.437 0.8613 8.47 ↓ 0.56 ↑
Cluster A B C D
1.1
0.5 1
0.4 C 0.6
C 0.5
α
B 0.4
0.35 β
0.3 γ
0.2 δ
A B κ
0.3 0.1 ξ
A 0
0.25 -0.1
0.8 0.81 0.82 0.83 0.84 0.85 0.86 0.87 0.81 0.82 0.83 0.84 0.85 0.86 0.87
Total efficiency Total efficiency
Fig. 4. Pareto-optimal solutions by hybrid MOEA. Fig. 5. Distributions of design variables over cluster points with total
efficiency.
the improvement of one objective leads to the impairment of the least sensitivity to objective functions along the Pareto-
the other. Here, it can be seen that no solution out of the optimal front. This design variable, then, can suitably be util-
Pareto-optimal solutions is superior to any other, in either ized to economize the procedure for multi-variable, multi-
objective, since each solution is a global Pareto-optimal solu- objective and multi-disciplinary design optimizations requir-
tion. A trade-off analysis showed that a higher efficiency can ing simultaneous consideration of many design variables and
be obtained at the cost of a higher torque, whereas the lower many performance objectives.
torque values are associated with lower values of efficiency. A The cluster solutions were reproduced by means of a RANS
designer can choose any optimal solution that accords with the analysis, as shown in Table 4. All of the representative opti-
available power that delivered the torque required for driving mal designs indicated that there was a significant decrease in
the fluid. the torque, though only a relatively small decrease of effi-
A greater insight into the relationship between the objective ciency was associated with them. In order to determine the
function and the design variables was attained by analyzing main factors responsible for the increase in the objective func-
the Pareto-sensitivity of the design variables, which is to say, tions, the internal flow fields were compared with the refer-
the change in the values of the design variables with the ence model and the optimal designs A and D.
change in the objective-function (e.g. total efficiency) values Fig. 6 shows the distributions of the torque value at the blade.
along the Pareto-optimal front. Out of the 620 global Pareto- The maximum torque value for solution A was observed near
optimal solutions, 4 clusters, A, B, C and D, were formed to the 70% blade span, whereas for the other designs it was
find the representative Pareto-optimal solutions, as shown in observed near the 80% span. Designs A and D had lower
Fig. 4. Fig. 5 illustrates the RSA-predicted Pareto-sensitivity torque values along the blade spans of 20~90% and 20~70%,
of the design variables for those representative Pareto-optimal respectively, than did the reference model. However, for the
solutions. The Pareto-sensitivity analysis not only revealed the blade spans beyond 90%, their torque values were higher.
effectiveness of the design variables but also determined the
extremes of the active design space in the Pareto-optimal de- Fig. 7 shows the pressure distributions along the blade spans
sign paradigm. The two design variables α and κ showed a of 10%, 50% and 90% for the suction and pressure surfaces in
decreasing trend, whereas the three design variables γ, δ, and ξ the reference model and optimal models A and D. On the
revealed an increasing trend. However, β was almost constant whole, the pressure distribution in model A, with the lowest
with efficiency. Thus, higher efficiency, in the context of the torque, has the lowest values among the three models. How-
Pareto-optimal designs, was obtained for low values of α and ever, it also had the highest pressure values for blade LE. The
κ and high values of γ, δ and ξ. The design variable β showed pressure distribution of the pressure surface in model D was
2064 J.-H. Kim et al. / Journal of Mechanical Science and Technology 24 (10) (2010) 2059~2066
100
80
60
Span (%)
40
Cluster A
20 Cluster D
Reference
0
0 0.1 0.2 0.3 0.4 0.5 0.6
Torque [Nm]
(a) Reference
Fig. 6. Comparison of blade torques.
200
Pressure at 10% span [Pa]
-200
-400 Cluster A
Cluster D
Reference
-600
0 0.25 0.5 0.75 1 (b) Cluster A
Streamwise locaton
400
Pressure at 50% span [Pa]
200
-200
Cluster A
-400 Cluster D
Reference
-600
0 0.25 0.5 0.75 1 (c) Cluster D
Streamwise location
(b) At 50% span Fig. 8. Streamline distributions on near-hub and blade suction surface.
600
Cluster A plane and the blade suction surface for the reference model
400 Cluster D
Pressure at 90% span [Pa]
Reference and the optimal models A and D. The reference model had the
200 widest recirculation zone at the near-hub and blade suction
0
surface, and diagonal streamlines. But the optimal model A
had almost no separation zone at the suction surface; the opti-
-200
mal model D, moreover, had a smaller separation zone than
-400 the reference model. Therefore, it is clear that optimized
shapes have more stable flow fields.
-600
0 0.25 0.5
Streamwise location
0.75 1
As discussed above, the optimal designs showed improve-
ment in torque, whereas the total efficiency was increased in
(c) At 90% span
the optimal model D and decreased in the optimal model A,
Fig. 7. Pressure distributions on blade surfaces. compared with the reference model. It is thought that the de-
sign variables were changed with the variation in the inci-
entirely more uniform than in the case of the reference model. dence angle (see Fig. 5), resulting in a significant reduction in
This trend was caused by the variation of the incidence angle the torque and an equally impressive enhancement of the effi-
with the changes of the design variables. ciency.
Fig. 8 shows the streamline distributions on the near-hub
J.-H. Kim et al. / Journal of Mechanical Science and Technology 24 (10) (2010) 2059~2066 2065
mization of a transonic compressor rotor, AIAA Journal of Jae-Ho Choi received his B.S. degree
Propulsion and Power, (20) (3) (2004) 559-565. from Inha University, Korea, in 1993,
[17] S. Pierret, R. F. Coelho and H. Kato, Multidisciplinary and and his M.S. and Ph.D. degrees in
multiple operating points shape optimization of three- Thermodynamics and Fluid Mechanics
dimensional compressor blades, Structural and Multidisci- at the same University in 1995 and
plinary Optimization, 33 (1) (2007) 61-70. 2000, respectively. He is currently a
[18] K. Deb, Multi-objective optimization using evolutionary principal research engineer and the
algorithms, 1st ed. John Wiley & Sons Inc. Chichester, Eng- leader of aerodynamic design group at
land (2001). Samsung Techwin R&D Center, Seongnam, Korea. He is an
[19] ANSYS CFX-11.0, Solver Theory, Ansys Inc. (2006). editor in the compressor division of the Korean Fluid Machin-
[20] N. V. Queipo, R. T. Haftka, W. Shyy, T. Goel, R. Vaid- ery Association (KFMA). His research and development in-
yanathan and P. K. Tucker, Surrogate-based analysis and op- terests are aerodynamic designs with conventional approaches
timization, Progress in Aerospace Science, 41 (2005) 1-28. and numerical optimizations, flow analyses, and performance
[21] R. H. Myers and D. C. Montgomery, Response surface tests on axial-, centrifugal-, and mixed-flow compressors used
methodology: process and product optimization using de- for energy equipment systems and gas turbines.
signed experiments, John Wiley & Sons, Inc., New York,
USA (2005). Afzal Husain received B.E. and
[22] MATLAB®, The language of technical computing, Re- M.Tech. degrees in Mechanical Engi-
lease 14, The Math Works Inc. (2004). neering with specialization in Thermal
[23] JMP 6.0.0, The statistical discovery software, version 6.0.0, Sciences from Aligarh Muslim Univer-
SAS Institute Inc., Cary, NC, USA (2005). sity, India in 2003 and 2005, respec-
[24] C. M. Jang, D. Sato and T. Fukano, Experimental analysis tively. He successfully completed Ph.D.
on tip leakage and wake flow in an axial flow fan according degree in Thermodynamics and Fluid
to flow rates, ASME J. of Fluids Eng., 127 (2005) 322-329. Mechanics at Inha University, Republic
[25] A. A. Giunta, Aircraft multidisciplinary design optimiza- of Korea. His research interests are computational fluid dy-
tion using design of experimental theory and response sur- namics, numerical analysis and optimization of fluid flow and
face modeling methods; Ph.D. dissertation, Virginia Poly- heat transfer systems using surrogate models, development of
technic Institute and State University, Virginia (1997). heat transfer augmentation and optimization techniques for
conventional- and micro-systems, thermal analysis of micro-
electromechanical systems (MEMS), and electronic cooling.
Jin-Hyuk Kim received his bachelor’s
degree to be honored with the early Kwang-Yong Kim received his B.S.
graduation of excellent from Sunmoon degree from Seoul National University
University, Korea, in 2007, and his in 1978, and his M.S. and Ph.D. degrees
master’s degree from Inha University, from the Korea Advanced Institute of
Korea, in 2009. He also received the Science and Technology (KAIST), Ko-
excellent master’s thesis award in fluid rea, in 1981 and 1987, respectively. He
engineering division from Korean Soci- is currently an Inha Fellow Professor
ety of Mechanical Engineers (KSME), Korea. Currently he is and the head of the School of Mechani-
pursuing his research towards Ph.D. degree in Thermodynam- cal Engineering of Inha University, Incheon, Korea. Professor
ics and Fluid Mechanics at Inha University, Korea. His re- Kim is also the current editor-in-chief of the International
search interests are designs of turbomachinery, numerical Journal of Fluid Machinery and Systems (IJFMS), and the
analyses and optimization techniques. president of the Korean Fluid Machinery Association
(KFMA). He is also a fellow of the American Society of Me-
chanical Engineers (ASME) and an associate fellow of the
American Institute of Aeronautics and Astronautics (AIAA).