Paper ID: ETC2017-326 Proceedings of 12th European Conference on Turbomachinery Fluid dynamics & Thermodynamics

ETC12, April 3-7, 2017; Stockholm, Sweden

A NOVEL OPTIMIZATION BASED DESIGN METHOD

FOR CENTRIFUGAL FANS

K. Bamberger - T. Carolus
Institute for Fluid- and Thermodynamics, University of Siegen, Germany
konrad.bamberger@uni-siegen.de

ABSTRACT
A novel design method for impellers of centrifugal fans with backward curved blades is
proposed. The overall geometry of the impellers considered is kept close to customary
industrial designs (e.g. circular arc blades). The design method is based on seven simple
correlations which determine the geometrical impeller parameters as a function of the
targeted aerodynamic duty (design) point. Unlike classic design methods, these correlations
are derived from aerodynamic optimization. The objective function of the optimization is the
maximization of total-to-static efficiency while precisely matching the targeted design points.
The range of design points studied covers the full typical realm of centrifugal fans according
to Cordier's diagram. A comparison to the classic slip theory as e.g. by Pfleiderer reveals that
the novel method is superior regarding both fulfillment of the targeted design point and
energy efficient performance.

KEYWORDS
centrifugal fan, design method, optimization, slip theory

NOMENCLATURE
Latin symbols ρ density, kg/m³
D diameter, m σ specific speed (nondimensional)
P power ω angular velocity (nondimensional)
Q flow rate, m³/s ψ pressure coefficient (nondimensional)
S blade thickness, m
c velocity, m/s Subscripts
h (blade) width, m 1 upstream of the fan
ki constants in the design correlations 2 downstream of the fan
l length, m dyn dynamic
n rotational speed, s-1 m meridional
p pressure t total
r radius ts total-to-static
s tip gap tt total-to-total
z number of blades θ circumferential
Greek symbols Abbreviations
β blade or flow angle, degrees ANN artificial neuronal network
δ specific diameter (nondimensional) CFD computational fluid dynamics
η efficiency (nondimensional) DoE design of experiment
ϕ flow coefficient (nondimensional) LMN local model networks
µ slip factor (nondimensional) MLP multi-layer perceptrons
RANS Reynolds averaged Navier Stokes

OPEN ACCESS 1 Copyright © by the Authors

Downloaded from www.euroturbo.eu
Paper ID: ETC2017-326 Proceedings of 12th European Conference on Turbomachinery Fluid dynamics & Thermodynamics
ETC12, April 3-7, 2017; Stockholm, Sweden

INTRODUCTION
This paper deals with the aerodynamic design of impellers for centrifugal fans, Figure 1. A good
aerodynamic design method reliably yields the geometry of a new fan that fulfills the targeted
performance in terms of pressure rise ∆p and volume flow rate Q with high energy efficiency.

Figure 1: Examples of centrifugal fan impellers for different design points

Q and ∆p are the so called design point. ∆p can be either total-to-total (∆ptt) or total-to-static
(∆pts). The total-to-total pressure rise is the difference of total pressure downstream (pt2) and
upstream of the fan (pt1):

∆ ptt = pt ,2 − pt ,1 (1)

All common design methods of fans use this pressure rise definition to specify the design point. In
many applications (e.g. rotor only fans), however, the dynamic pressure at the fan outlet (pdyn,2) is
useless as it dissipates in the surroundings. In that case, the total-to-static pressure rise becomes
relevant:

∆ pts = p2 − pt ,1 = ∆ ptt − pdyn ,2 (2)

Assuming that the inflow is free of swirl, ∆ptt can be computed by Equation (3):

D2
∆ ptt = ηblade ⋅ ρ ⋅ ω ⋅ cθ 2 (3)
2

ηblade is the hydraulic efficiency between inlet and outlet, ρ the fluid density, ω = 2πn the angular
velocity of the impeller and cθ2 the circumferential velocity of the flow at the impeller outlet. The
outer impeller diameter D2 and rotational speed n are the main machine parameters and are assumed
to be given at the beginning of a design process. Hence, the main challenge of the designer is to
estimate ηblade and to find an impeller geometry which yields the cθ2 which is required to produce
∆ptt. Starting in the early 1900's, successful methods for linking cθ2 to the impeller geometry have
been developed by Pfleiderer [1, 2], and later by Eck [3] as well as many others. In essence it is the
underperformance factor

∆ ptt
µ= (4a)
∆ ptt ,th

OPEN ACCESS 2 Copyright © by the Authors

Downloaded from www.euroturbo.eu
which links the actual to the theoretical pressure rise for blade-congruent flow. Pfleiderer' approach
[2] for instance is

−1
  β  
 k ⋅ 1 + b 2  
µ = 1 +  60°  
( )
(4b)
 z 1 − ( D1 D2 ) 
2

 
 

where the factor k is usually between 1.2 and 2. From (3) and (4a, b) it is possible to compute ∆ptt,th
and the corresponding theoretical circumferential velocity cθ2,th which eventually yields the blade
outlet angle required to achieve the desired pressure rise. Obviously µ depends on the blade outlet
angle βb2, which in turn depends on the underperformance factor. Hence Pfleiderer's approach
requires an iterative determination of µ. A user-friendly summary of these methods is given by
Carolus [4]. Moreover, Carolus summarizes recommendations by Bommes [5] and Pohl [6] to
determine adequate values of the other geometrical parameters, e.g. the blade widths h1 and h2 and
the inner diameter D1. Those recommendations are supposed to lead to reasonably good designs
with an acceptable impeller efficiency defined as

Q⋅∆p
η= (5)
Pshaft

where Pshaft is the shaft power driving the impeller. In general, the resulting efficiency of a fan can
be computed with both ∆ptt and ∆pts yielding the total-to-total-efficiency ηtt or the total-to-static-
efficiency ηts, respectively. In all common design methods, however, the focus is on ηtt. The
aforementioned design methods have the advantage of being applicable to a wide range of design
points and that they require nearly no computational resources. The main drawback lies in the
inherently empirical base (indicated for instance in the considerable scatter in the coefficient of
approach (4b) from 1.2 to 2.0) which limits the reliability of the method regarding the fulfillment of
the design point. Moreover, the maximum achievable efficiency may not be achieved.
More recently Computational Fluid Dynamics (CFD) is coupled with optimization algorithms. A
typical example (among many others) is the work by Lee et al. [7] who used CFD-based
optimization to reduce the shaft power of a centrifugal fan by 9 %. While CFD-based optimization
can yield a highly efficient design of a specific machine, it suffers from long computing times.
Hence, the latest technique is to substitute CFD by CFD-trained metamodels (e.g. artificial neural
networks, ANN). The effectiveness of this approach is generally acknowledged and even applied
commercially, see e.g. [8-10]. However, the computational cost is still significant due to the
necessity to generate a CFD-database required to train the ANNs.
The aim of this work is to develop a novel design method which can be applied similarly as the
classic methods, i.e. a design point is given and the corresponding geometrical parameters of the
impeller are computed analytically. By contrast with the classical method the semi-empirical
equations are replaced by correlations that essentially stem from aerodynamic optimization with a
generalized metamodel for the complete class of typical radial fans. The metamodel is obtained one
time only prior to any individual design task by a comprehensive CFD-study (as reported in [11] by
the authors of this paper). This novel design method aims at yielding fan impeller geometries within
the full class of centrifugal fans that fulfill more reliably the targeted design point - with the highest
efficiency possible.

3
 OPTIMIZATION SCHEME
Initially, the geometry of the impeller needs to be parameterized. Those parameters are then
varied by Design of Experiment (DoE) and CFD simulations are performed for each geometry
variation. The resulting CFD dataset is then used for training artificial neural networks which are
eventually coupled with an evolutionary optimization algorithm. In the subsequent section this
optimization scheme is applied for deriving the desired design correlations.

Parameterization
Since the design correlations shall be generally applicable, all correlated quantities are non-
dimensional. Lengths are non-dimensionalized with the outer fan diameter D2 and all velocities are
non-dimensionalized with the circumferential blade velocity ω·r2 yielding a velocity coefficient ϕc.
The design point is non-dimensionalized as usual in turbomachinery literature (see e.g. [12])
yielding the flow and the pressure coefficient

Q ∆p
ϕ= ,ψ = . (6a, b)
π 2
π 2

Dn 3
2 Dnρ
2
2
2

4 2

The main objective of the impeller parameterization is to enable a high level of geometrical
flexibility while keeping the number of free geometry parameters to a minimum. For that reason,
basic parameters which are also used in classic literature [2, 5] are selected. Those parameters are
graphically illustrated in Figure 2 and listed in Table 1. The blades have a circular arc shape
between the inner diameter D1 and the outer diameter D2. The hub is a flat disc perpendicular to the

Figure 2: Graphical illustration of the geometrical parameters in side and top view

Table 1: Independent geometrical parame- Table 2: Dependent and constant geometri-

ters cal parameters
Symbol/ Min. Max. Symbol/
Description Description Value
Definition Value Value Definition
Number of blades1 z 5 16 Outer diameter D2 0.3 m
Diameter ratio D1/D2 0.25 0.8 Rotational speed n 50 s-1
Inlet blade angle βb1 20° 60° Fluid density ρ 1.2 kg/m3
Inlet blade angle βb2 20° 60° Nozzle radius rn/D1 0.25
Inlet blade width2 h1/D2 0.025 0.4 Clearance lc/D1 0.02
Outlet blade width2 h2/D2 0.025 0.4 Overlap lo/D1 0.03
Shroud radius rs/D1 0.14 0.3
1
only integers are possible
2
h2 must be smaller than or equal to h1

4
Paper ID: ETC2017-326 Proceedings of 12th European Conference on Turbomachinery Fluid dynamics & Thermodynamics
ETC12, April 3-7, 2017; Stockholm, Sweden

axis of rotation. The shroud is not parallel to the hub but tapers between D1 and D2. The innermost
part of the shroud has a circular shape with the radius rs. The outer (mostly much larger) part of the
shroud has a linear shape determined by the blade widths h1 and h2. The blade thickness S is
constant. The radius and the center point of the circular blade arc are given via the inlet angle βb1
and the outlet angle βb2 which are measured between the circumferential direction and a tangent at
the blade leading or trailing edge, respectively. There is a small cut-back of the blade in the area
where the leading edge touches the shroud. The cutting edge starts at the radius D1/2 at the distance
h1 from the bottom disc and perpendicularly intersects the circular part of the shroud. Together with
the number of blades z, this set of parameters represents simple yet widely used impeller geometries
which can be manufactured employing cost-efficient techniques. In the context of this study the
number of parameters is reasonably low to allow for a systematic variation by design of
experiments. D2 and S are excluded from the design of experiment. D2 is held constant because we
consider the dimensionless aerodynamic performance which is independent of D2. This also applies
to the rotational speed n and the fluid density ρ which are held constant, too. Hence, strictly spoken
all results in this study are only valid for one value of a characteristic Reynolds number. The reason
why S is no independent parameter is that the blade thickness only has a minor effect on the
aerodynamic performance and is usually selected for structural reasons. We here assume that the
blade thickness increases with increasing ratio D1/ D2 as

S D
= 0.006 ⋅ 1 . (7)
D2 D2

Further geometrical parameters are required to describe the inflow nozzle. Since the inflow nozzle
is less relevant with respect to the aerodynamic performance, those parameters are a function of the
impeller parameters and do not need to be varied by design of experiments. The inflow nozzle has a
simple geometry and is fully described by its outflow diameter and the radius rn. Between the
nozzle and the impeller there is a clearance of the size lc. The overlap of the nozzle with the fan is
lo. All dependent parameters are listed in Table 2.

Design of Experiment (DoE)

As recommended by Santner et al. [13], a space filling DoE should be chosen, if no prior
knowledge about the process is available. Due to its advantages and simplicity [13] a latin
hypercube (LH) design has been used. The non-collapsing property of LH designs is one advantage
compared to commonly used grid designs. In order to achieve good space filling properties the LH
design was optimized with the extended deterministic local search method described by Ebert et al.
[14]. Approximately 2000 different impeller geometries were simulated with that DoE and used to
train preliminary metamodels. In a second DoE, 2000 additional geometries were defined and
simulated. The geometry selection for those 2000 impellers followed two basic ideas: (i) Avoidance
of irrelevant sections of the input space (which are aerodynamically suboptimal), (ii) Avoidance of
the simulation of geometries which are unlikely to enhance to metamodels (areas of the input space
where the aerodynamic performance is almost a linear function of the geometrical parameters and
hence easy to model).

CFD Model
The CFD model used for this work emulates a typical experimental set-up of an impeller in a
chamber test rig, i.e. the inflow is assumed to traverse a chamber and the inflow nozzle whereas the
outflow area is free of obstacles. Figure 3 depicts a sketch of all components considered for the
CFD-model. The boundary conditions are constant mass flow rate at the inlet, ambient pressure at
the opening and no slip at the walls. Hub, shroud and blade are modeled as rotating walls with the

OPEN ACCESS 5 Copyright © by the Authors

Downloaded from www.euroturbo.eu
Figure 3: CFD model

rotational speed n whereas all other walls are stationary. Due to the high number of CFD
simulations required to train the metamodels, only stationary simulations were affordable, i.e. the
Reynolds-Averaged Navier-Stokes (RANS) method was used. The solver selected was ANSYS
CFX 14.5.7 and the turbulence model used was shear stress transport (SST). As usual in RANS
simulations, only one blade channel was simulated and rotational periodicity was assumed at the
lateral boundaries. The interface type between the rotating blade and the stationary environment
was frozen rotor. In order to determine the fan performance, the quantities ∆pts, ∆ptt and Pshaft must
be extracted from the CFD results. ∆pts is the difference between the area-averaged pressure at the
opening and the inlet. The dynamic pressure required to determine ∆ptt is evaluated at the interface
between the impeller outlet and the environment. The overall shaft power Pshaft equals the sum of
the powers associated with all rotating surfaces, i.e. the surfaces at the blade, the hub and the
shroud. The computational grid of the impeller was generated with ANSYS TurboGrid 14.5.7 and
the grid of the environment was generated with ICEM CFD. A grid optimization was performed
to find the best compromise between accuracy and computational time. Details about this grid
optimization and an experimental validation of the CFD model are discussed in [11].

Metamodel Training
Two meta-model types were tested: Local Model Networks (LMN) and Multi-Layer
Perceptrons (MLP). LMNs follow a divide-and-conquer strategy. The input space spanned by all
geometry parameters is divided into small sub-spaces, where simple local models are able to
sufficiently well describe a part of the whole function to be approximated. For the training of the
LMNs the hierarchical local model tree (HILOMOT) algorithm (see Nelles [15]) was used with
local linear models. The structure of the LMN grows incrementally by adding a further division of
one pre-existing sub-space. Therefore, one local model is added in each step until a termination
condition is met. For the automatic prevention of overfitting the available data can be split into two
datasets for training and validation. An increase of the validation error indicates beginning
overfitting. For the termination of the LMN training process 20% of the available data was utilized
as validation data. As soon as the validation error increased two times in succession, the training
was terminated in order to prevent the model from overfitting.
MLPs consist of the input layer, one or more hidden layer(s) and the output layer. The number
of hidden layers and the number of neurons in each of the hidden layers determines the complexity
of the model. In the present work, two hidden layers were used and the number of neurons in each
hidden layer was optimized aiming at a minimal validation error. Training an MLP for a given
6
structure means optimization of the weights in the hidden layers. This was performed with the
Neural Network Toolbox of Matlab which uses the Levenberg-Marquardt [16] method. The target
function of the weight optimization was the mean squared difference between the MLP output and
the CFD results. 80% of the available CFD data was used for that purpose whereas the remaining
20% was used for validation. It was found that the MLPs performs slightly better than the LMNs,
but that the most reliable prediction is obtained when using both models and averaging the
respective outputs (weighted with the inverse validation error).
Optimization Algorithm: An evolutionary optimization algorithm was used. The
implementation mainly follows the suggestions by Thévenin and Janiga [17]. The basic settings are
2.000 individuals per generation, reproduction via cross-over (80% probability) or averaging (20%
probability) and random mutations at the end of each reproduction process. The objective function
is maximization of total-to-static efficiency with a penalty term for the violation of the targeted
pressure coefficient ψtt,target:

η * = ηts − ψ tt −ψ tt ,t arg et (8)

ψtt and ηts are evaluated by the metamodels. Due to the extremely efficient metamodels the
performance of one generation can be computed in less than one second on one CPU of a standard
personal computer.

THE NEW CENTRIFUGAL IMPELLER DESIGN METHOD

Step 1 (Heuristic Determination of the Structure of the Design Correlations)

The optimization tool described above was applied for finding the optimal impeller geometries
for the 94 design points illustrated in Figure 4. The design points are computed by

1
ψ tt ,t arg et = 1 − ϕ+∆ (9)
0.35

where ϕ is varied in ten equidistant steps between 0.04 and 0.25 and ∆ is varied in ten equidistant
steps between -0.3 and +0.3. This initially leads to 100 design points. Due to the constraint ψtt,target ≥
0.25, however, the overall number of different design points reduces to 94. The optimal geometries
were examined heuristically to find general correlations about the optimal choice of the geometrical
parameters. These correlations are described in the following. As mentioned previously, the
correlations shall be as simple as possible. Often, it was easier to define correlations for flow
quantities from which the geometrical parameters are derived instead of directly defining
correlations for the respective parameters. Most of the following correlations contain one or more
[-]
tt,target

Figure 4: Design points used for the derivation of the design correlations
7
so far unknown constants ki. The values of the constants will be identified by subsequent
optimization in the second step.

The first design correlation is rather simple. All 94 optimizations yield blade inlet angles which
are as flat as possible within the limits defined in Table 1. Therefore

!
β b ,1 = 20° . (10)

Most probably, the actual optimum is even lower, but this cannot be examined with the existing
metamodel which is only valid within the limits defined in Table 1. A constant inlet angle
irrespective of the design point is not intuitive. Normally, the angle increases with the flow
coefficient to minimize incidence losses. The design target of maximum total-to-static efficiency,
however, leads to large throughflow areas to decrease exit losses and consequently to flat inlet
angles for all design points. The inlet area is determined by the inner diameter D1 and the inlet
width h1. These geometrical parameters are basically determined by the second and third design
correlation. However, these design correlations do not directly yield the geometrical parameters but
the dimensionless meridional inflow velocity ϕcm1 and the incidence angle ∆β1 from which D1 and
h1 can be derived via Eq. 17 and 18. ϕcm1 is always on a similar level and can be approximated by a
constant:

cm1 !
ϕcm1 = = k1 (11)
ω r1

The incidence angle ∆β1 (i.e. the difference between the blade angle and the flow angle) is
negative at small flow coefficients and asymptotically approaches a shock-free inflow (i.e. ∆β1 = 0)
with increasing flow coefficient. This can be expressed via the third design correlation:

!
∆β1 = β b ,1 − β1 = k2 ⋅ e − k ⋅ϕ
3
(12)

Given these three correlations, the inlet geometry is described unambiguously. The next two design
correlations are required to describe the outlet geometry. As with the inlet geometry, the outlet
design correlations yield flow quantities from which the corresponding geometrical parameters can
be derived by Eq. 17 and 18. The fourth design correlation defines the optimal meridional
deceleration which is found to be a function of the flow coefficient but independent of the pressure
coefficient:

ϕcm 2 !
= k 4 ⋅ ϕ 2 + k5 ⋅ ϕ + k6 (13)
ϕcm1

The fifth design correlation deals with the underperformance factor µ, i.e. with the quotient of
the actually achieved pressure coefficient ψtt,actual and the theoretical pressure coefficient ψtt,th that
would ensue if the relative exit flow direction was congruent with the blade. Precise prediction of
the underperformance factor is of fundamental importance since it is decisive for the accuracy with
which the targeted design point is met. One possibility to achieve an accurate prediction is defining
a very complex design correlation with many free constants ki. This strategy, however, would
counteract the objective to keep the design correlations as simple as possible. As a consequence, it
is suggested to firstly approximate a preliminary underperformance factor by a linear function of ϕ
and ψtt,target. This underperformance factor is then applied to design a preliminary impeller and the

8
actual pressure coefficient is determined for that machine. In case of unsatisfactory results, the
underperformance factor can be corrected:
!
µ prelim = k7 ⋅ ϕ + k8 ⋅ψ tt ,target + k9
ψ (14)
µ = µ prelim ⋅ tt ,t arg et
ψ tt ,actual

During the development of the design correlations, ψtt,actual was determined by the metamodels. In
the application of the design correlations, it is assumed that the user determines ψtt,actual via CFD
simulations or experiments. This extra-effort is considered acceptable since in most practical
applications corrections are required anyway due to different gaps or inflow conditions as compared
to the assumptions used in this study. If the correction is not affordable, the preliminary
underperformance factor still will lead to a very acceptable performance.
The sixth and seventh design correlation directly determine geometrical parameters, namely the
number of blades z and the radius of the cover disc nozzle rn. Both parameters are linear functions
of ϕ and ψtt,target. Note that z always needs to be rounded to an integer.

!
rn !
z = k10 ⋅ ϕ + k11 ⋅ψ tt ,target + k12 , = k13 ⋅ ϕ + k14 ⋅ψ tt ,target + k15 (15, 16)
D2

Since many of the design correlations deal with flow quantities, Equations 17 to 20 are required
to determine the corresponding geometrical parameters.

D1 ϕcm1 b ϕ
= , 1 = (17, 18)
D2 tan ( β b ,1 + ∆β1 ) D2 4 ⋅ ϕ ⋅ D1
cm1
D2

 ϕcm 2 
b2 b D ϕ 
−1
 ϕcm1 ⋅ ϕ 
= 1 ⋅ 1 ⋅  cm 2  , β b ,2 = a tan  cm1
 (19, 20)
D2 D2 D2  ϕcm1  ψ
 1 − tt ,t arg et 
 2⋅µ 
 

Throughout the development of the design correlations, it was assumed that the limits according
to Table 1 must not be exceeded. In case that any of the design correlations leads to parameter
values below or above the ranges defined in Table 1, the lower or upper limit must be used,
respectively.

Step 2 (Identification of the Constants in the Design Correlations)

The values of the constants ki are found with the Simplex method by Nelder and Mead [18]. The
objective function is maximization of the averaged total-to-static efficiency of the 94
aforementioned design points with penalty terms for the violation of the design points. The resulting
optimal values of the constants are summarized in Table 3.

Table 1: Optimal constants for the design correlations

Constant Value Constant Value Constant Value Constant Value Constant Value
k1 0,296 k4 -0,012 k7 -0,225 k10 93,2 k13 0,233
k2 27,3 k5 0,541 k8 0,498 k11 26,9 k14 0,453
k3 12,6 k6 0,599 k9 0,238 k12 -19,1 k15 0,04

9
 RESULTS
The full geometry of impellers for the aforementioned 94 design points was identified by means
of the design correlations. Subsequently, their actual aerodynamic performance was simulated by
means of independent CFD to assess the quality of the design correlations. The first question is how
precisely the actual pressure coefficient ψtt,acutal matches the targeted pressure coefficient ψtt,target.
Figure 5(a) illustrates the deviation ∆ψtt = ψtt,actual - ψtt,target. At most design points, the deviation
∆ψtt is close to zero. Only at very low targeted pressure coefficients (ψtt,target ≈ 0.25) the deviation
reaches values of up to 0.04. That means that the actual pressure coefficient ψtt,acutal of nearly all
designs investigated matches the targeted pressure coefficient ψtt,target very well, or, with other
words, the novel design correlations seem to be quite reliable.
The second question is concerned with the efficiency of the various designs. The total-to-static
efficiency ηts associated with all designs is depicted in Figure 5(b). At almost all design points, it is
around 60 % or moderately higher. The lowest value is obtained at design points with high flow
coefficient but low pressure coefficient and amounts to approximately 45 %. Such design targets are
more typical for axial fans. Figure 6 depicts the difference between the efficiencies obtained with
impellers designed by the novel design correlations and the classic Pfleiderer method. Obviously,
the novel design method yields impellers with better performance regarding both total-to-static and
total-to-total efficiency. The differences can be as high as 10 percentage points at design points with
high targeted pressure coefficients. This observation is in full agreement with an earlier similar
study on axial fans [19] which revealed that the potential for improvement through optimization
increases with increasing blade loading. Remarkably, the new design method enhances the total-to-
total efficiency even more than total-to-static efficiency although the objective function of the
optimization algorithm was to maximize total-to-static efficiency. The reason for that is the self-
imposed constraint βb,2 ≥ 20° (see Table 1) which prohibits large exit areas and consequently low
meridional flow velocities.

CONCLUSIONS
A novel design method for the impellers of centrifugal fans was developed and successfully
tested. The design method consists of seven simple correlations for geometrical parameters or flow
quantities. In terms of simplicity and quickness the method compares well to classic methods such
as the slip theory by Pfleiderer. However, since the novel correlations have been obtained by
application of a comprehensive optimization tool, the accuracy in terms of meeting the design target
is better than the accuracy of classic methods. Moreover, it was found that the efficiency of the
designs obtained with this new method is always superior over designs via classic methods. This
holds true for all design points of typical centrifugal fans but is more pronounced for design points
requiring high blade loading. Corrections may become necessary if the characteristic Reynolds
number and secondary parameters (e.g. inflow conditions, size of gaps) differ substantially from the
assumptions on which the design correlations are based on.

ACKNOWLEDGMENT
Part of this work was funded by the German Ministry for Economic Affairs and Energy
(BMWi), the German Federation of Industrial Research Associations (AiF) and the Research
Association for Air and Drying Technology (FLT) under the project number 18084.

10
(a) (b)

Figure 5: Impellers designed by means of the novel design correlations: (a) Deviation of actual
and targeted pressure coefficient (∆ψtt = ψtt,actual - ψtt,target); (b) Total-to-static efficiency; (one
correction step for µ according to Equation (14))

Figure 6: Difference of efficiency between impellers designed with the novel design correla-
tions and the classic Pfleiderer method (∆ηtt/ts = ηtt/ts,actual - ηtt/ts,Pfleiderer)

REFERENCES
[1] Pfleiderer, C., 1961, Die Kreiselpumpe für Flüssigkeiten und Gase, Springer Verlag,
Berlin-Heidelberg.
[2] Pfleiderer, C., Petermann, H., 1991, Strömungsmaschinen, Springer-Verlag, Berlin,
Germany.
[3] Eck, B., 2003, Ventilatoren, Springer Verlag, Berlin, Germany.
[4] Carolus, T., 2012, Ventilatoren - Aerodynamischer Entwurf, Schallvorhersage,
Konstruktion, Springer Vieweg, Wiesbaden, Germany.
[5] Bommes, L., Fricke, J., Grundmann, R., 2003, Ventilatoren, Vulkan Verlag, Essen.
[6] Pohl, C., 2001, "Schaufelformen bei Industrie-Radialventilatoren", Proc. VDI
Ventilatortagung, Braunschweig, VDI Verlag.
[7] Lee, Y.-T., Ahuja, V., Hosangadi, A., Slipper, M. E., Mulvihill, L. P., Birkbeck, R.,
Coleman, R., 2011, "Impeller Design of a Centrifugal Fan with Blade Optimization",
INternational Journal of Rotating Machinery, Article ID 537824, p. 16.
[8] Demeulenaere, A., Ligout, A., Dijkers, R., Purwanto, A., Hirsch, C., Visser, F., 2005,
"Design and Optimization of an Industrial Pump: Application of Genetic Algorithms and
Neural Networks", Proc. ASME Fluids Engineering Division Summer Meeting &
Exhibition 2005, Houston, TX.

11
[9] Pierret, S., Hirsch, C., 2003, An Integrated Optimization System for Turbomachinery
Blade Shape Design, Defense Technical Information Center, Brussels, Belgium.
[10] Hildebrandt, T., Starke, A., Ruck, G., 2010, "Leistung- und Effizienzsteigerung einer
Diagonallüfterstufe: Von der CFD basierten Optimierung mit FINE™/Turbo bis zur
Serieneinführung", Proc. VDI Ventilatoren 2010, Braunschweig, Germany.
[11] Bamberger, K., Belz, J., Carolus, T., Nelles, O., 2016, "Aerodynamic Optimization of
Centrifugal Fans Using CFD-Trained Meta-Models", Proc. International Symposium on
Transport Phenomena and Dynamics of Rotating Machinery (ISROMAC) 2016, Honolulu,
Hawaii, USA.
[12] Dixon, S. L., Hall, C. A., 2010, Fluid Mechanics and Thermodynamics of Turbomachinery
- Sixth Edition, Butterworth Heinemann, Oxford.
[13] Santner, T., Williams, B., Notz., W., 2003, The design and analysis of computer
experiments, Springer Science & Business Media.
[14] Ebert, T., Fischer, T., Belz, J., Heinz, T. O., Kampmann, G., Nelles, O., 2015, "Extended
Deterministic Local Search Algorithm for Maximin Latin Hypercube Designs", Proc.
IEEE Symposium on Computational Intelligence in Control and Automation.
[15] Nelles, O., 2006, "Axes-Oblique Partitioning Strategies for Local Model Networks", Proc.
International Symposium on Intelligent Control.
[16] Marquardt, 1963, "An Algorithm for Least-Squares Estimation of Nonlinear Parameters",
SIAM Journal of Applied Mathematics, 11, pp. 431-441.
[17] Thévenin, D., Janiga, G., 2008, Optimization and Computational Fluid Dynamics,
Springer Verlag GmbH, Heidelberg, Germany.
[18] Nelder, J. A., Mead, R., 1965, "A Simplex Method for Function Minimization", Computer
Journal, 7, pp. 308-313.
[19] Bamberger, K., 2015, "Aerodynamic Optimization of Low-Pressure Axial Fans", PhD
Thesis, University of Siegen.

12