Optimizing the Number of Airfoils in Turbine

Design using Genetic Algorithms

José M. Chaquet?1 , Enrique J. Carmona2 , and Roque Corral1

1
Technology and Methods Dep., Industria de TurboPropulsores S.A.
Avda. de Castilla 2, 28830, San Fernando de Henares, Madrid, Spain
2
Dpto. de Inteligencia Artificial, Escuela Técnica Superior de Ingenerı́a Informática,
Universidad Nacional de Eduación a Distancia, Madrid, Spain

Abstract. A method for optimizing the number of airfoils of a turbine

design is presented. The optimization consists of reducing the total num-
ber of airfoils meanwhile a set of geometric, aerodynamic and acoustic
noise restrictions are fulfilled. It is described how is possible to reduce
the problem degrees of freedom to just one variable per row. Due to the
characteristics of the problem, a standard Genetic Algorithm has been
used. As a case study, a real aeronautical Low Pressure Turbine design
of 6 stages has been optimized.

1 Introduction

A turbine of a gas turbine engine is a device that extracts work from a pressured
gas stream. It is normally made up of three modules, called HPT, IPT and
LPT (High, Intermediate and Low Pressure Turbine). The extraction of work
from the fluid is done by means of several aerodynamic surfaces called airfoils
which are placed in an annular way forming row s. A turbine stage is formed by
two consecutive rows, called stator and rotor. Stator airfoils are called vanes,
meanwhile rotor airfoils are called blades. The stator is attached to the casing
and directs the flow towards the rotor, meanwhile the rotor transmits the power
to the turbine shaft. The number of airfoils of a row is called NumberOff.
The design process of an aeronautical turbine is a very challenging task. A
LPT can contribute with one third to the total weight and with up to 15%
to the total cost [1]. A lot of different constraints must be taken into account
when designing the LPT airfoils and usually the final decision on the optimum
particular configuration requires a trade-off among different requirements.
In this work it is presented a method for optimizing the NumberOffs of a
turbine. The optimization consists of reducing the total number of airfoils mean-
while a set of geometric, acoustic and aerodynamic restrictions are fulfilled. It
will be demonstrated that is possible to reduce the problem Degrees of Freedom
?
Corresponding author: Telf.: +34 91 2079147; Fax: +34 91 2079411.
E-mail: [email protected] (J.M. Chaquet), [email protected] (E.J. Car-
mona), [email protected] (R. Corral).
(DoF) to just the NumberOff for each row. The approach adopted to solve the
optimization problem uses a Genetic Algorithm (GA).
There are several applications in the literature that make use of GAs in the
design process of gas turbine components, like the control system unit [2,9],
blade cooling holes [3,5,6], the combustor [4,10], rotor system [7] or the 2D and
3D design of airfoils [8,11]. Some applications use standard GA [4,8,11], while
others implement specific GA, as Multi-objective GA (MOGA) which evolves
a Pareto-optimal solution [2,7,9,10]. In some methods the initial strategy in-
volves the identification of high performance (HP) regions of conceptual design
spaces and the extraction of relevant information regarding the characteristics of
the solutions within these regions [3,5]. HP regions are rapidly identified using
the COGA approach (Cluster-oriented GA). Another special GA used is called
GAANT, which is based upon ant colony concepts and genetic algorithms [5].
Other possible approach is a Generalized Regression GA (GRGA) which explores
the relationship among the variables of the solutions belonging to any continues
portion of the Pareto front using non-linear multivariable regression analysis [6].
In this paper, first a description of the problem to solve is presented (section
2). It will be shown how to reduce the problem DoF. Then the GA approach will
be described (section 3) and the results obtained in the optimization of a real 6
stage aeronautical gas turbine are presented (section 4). Finally, the conclusions
and future works are given (section 5).

2 Problem Description

The problem consists of the turbine total number of airfoils minimization for a
given flow-path (Fig. 1a) and aerodynamic exit angles. The minimization process
has to fulfil a set of aerodynamic, acoustic and geometric restrictions that may
be reduced to a set of explicit analytical expressions. As a consequence, both the
objective function and the restrictions are extremely fast to evaluate.
In order to parametrize the problem, a simplified geometry will be used ap-
proximating each row by a rectangle (Fig. 1b). For a turbine of M number of
rows, each row is defined with only 5 parameters: NumberOff (Ni ), gap (gi ),
chord (ci ), mean radius (Ri ) and span (Si ) where i goes from 1 to M . The mean
radius is the distance of the middle point of the row to the turbine axis. It is
also needed one global variable, L, which is the total axial length of the turbine.
The turbine inner and outer annuli are supposed to be optimized in an outer
loop and in this exercise are kept constant. Therefore the mean radius and the
span of all the rows are constant. NumberOffs, gaps and chords will be modified
in order to find optimum feasible configurations. Gaps for row i is the distance
between the trailing edge of row i and the leading edge of next row i + 1, or
the exit station for the last row. The initial gap, g0 , is defined as the distance
between the inlet station and the leading edge of first the row (figure 1b).
The geometric constraints are defined with the following parameters for each
row: maximum aspect ratio (M Ai ), minimum pitch to chord ratio (mP Ci ), max-
imum pitch to chord ratio (M P Ci ), minimum gap (mGi ), minimum gap to chord
 a)

b)
ci

Si

g0

gi
Ri

Fig. 1. Real geometry (a) and simplified one (b)

ratio (mGCi ), maximum gap to chord ratio (M GCi ), maximum NumberOff

(M Ni ) and the NumberOff for each package (Pi ). The maximum aspect ratio
should be limited by mechanic and flutter response. The pitch to chord ratio is
limited in order to maintain Zweiffel coefficient bounded. Gaps are bounded in
order to avoid mechanical interferences and by noise restrictions. The package
parameter imposes that the NumberOff must be multiple of Pi . For the inlet gap
g0 , two constraints are given for bounding its value between a minimum and a
maximum value: mG0 and M G0 .
It is well known that one way of reducing the generation of noise associated
to pure tones is to force that the NumberOffs ratio for two consecutive rows
lies within some specific intervals [1]. When the NumberOff ratio fulfills this
conditions, the acoustic wave amplitudes decrease with the axial distance, the
stage is said to be cut-off and the perturbations do not propagate outside the
turbine. The cut-off condition depend as well on the flow variables, but in our
problem these are assumed to remain constant. Noise constraints are given by
four parameters: αi , βi , γi and δi . These parameters define two intervals [αi , βi ]
and [γi , δi ] where the ratio of NumberOff of row i and row i + 1 must be located.
Normally 0 ≤ αi ≤ βi ≤ 1 ≤ γi ≤ δi . When both ranges are used, the configura-
tion is called Mixed cut-off. For Direct cut-off mode, [αi , βi ] interval is chosen for
even rows and [γi , δi ] for odd rows, therefore it will be more vanes than blades.
The opposite is chosen for Reverse cut-off mode.
 Putting all together, the mathematical problem consist of finding the M
positive integer numbers Ni and the M positive real numbers ci and gi which
fulfill the following constraints for i ∈ [1, M ]

Si
≤ M Ai (1)
ci

2 · π · Ri
mP Ci ≤ ≤ M P Ci (2)
Ni · ci

mGi ≤ gi (3)
gi
mGCi ≤ ≤ M GCi (4)
ci
Ni ≤ M N i (5)
Ni %Pi = 0 (6)
M
X
mG0 ≤ g0 ≡ L − (ci + gi ) ≤ M G0 (7)
i=1

Ni
if (i 6= M & M ixed) ∈ [αi , βi ] ∪ [γi , δi ] (8)
Ni+1
Ni
if {i 6= M & (Direct & i%2 = 0) or (Reverse & i%2 = 1)} ∈ [αi , βi ] (9)
Ni+1
Ni
if {i 6= M & (Direct & i%2 = 1) or (Reverse & i%2 = 0)} ∈ [γi , δi ] (10)
Ni+1
In equation (6) the symbol % means the remainder of integer division. Equation
(7) computes the first gap g0 and it imposes that g0 must be in between mG0
and M G0 .
Taking into account the three parameters for each row (Ni , ci and gi ), there
are 3M DoF. We will show that the problem may be reduced to that of finding
the M DoF associated to the number of airfoils for each individual row.
Fig. 2a displays Gap-Chord space. The shaded region represents gi and ci
feasible pairs where constraints (3) and (4) are represented. Points A and A’
have the same chord, but A’ has the minimum possible gap. The same happens
with points B and B’. If point A is feasible regarding to all constraints except
those in equation (7), point A’ will be feasible as well. But point A’ could be
considered better than A because gives more room to other rows to increase
their gaps and chords. For that reason and regardless of other considerations,
gaps will be set to the minimum for a given chord:

gi ≡ gi (ci ) = max (mGi , mGCi · ci ) (11)

 g MGC i a) Ni
i b)
A
MN i
Ni = 2. π.R / (mPC . c i)
i i
B’ B
A’
B mGC i A’
A
mG i B’

ci Ni = 2. π.Ri/ (MPC i. c ) ci
i
cmin,i cmax,i
mG

Fig. 2. Feasible domain spaces (shaded area): Gap-Chord (a) and NumberOff-Chord
(b)

In Fig. 2b a typical NumberOff-Chord diagram for a specific row is plotted.

The shaded region represents ci and Ni feasible pairs without taking into account
the Noise constraints. Using constraints (1), (2), (3) and (4) we obtain:
 
Si mGi 2 · π · Ri
cmin,i = max , , (12)
M Ai M GCi M Ni · M P Ci

Using (7), (11) and (12) expressions it is possible to compute for each row a
maximum chord considering that the rest of rows have their minimum chords.
It is important to notice that this limit is not absolute, but it depends on the
rest the turbine rows. It is an upper limit, but will decrease if at least one row
has his chord larger than his minimum chord.
M
X
cmax,i = L − (cmin,j + gj (cmin,j )) − mg0 (13)
j=1,j6=i

From Fig. 2b we can argue that if point A is feasible regarding all constraints,
point A’ will be feasible as well. A’ is considered better because it has the min-
imum chord for a given NumberOff. Smaller chords give more room to other
rows. The same consideration may be done for points B and B’. Then, given a
NumberOff, the optimum chord can be chosen using the following expression:
 
2 · π · Ri
ci ≡ ci (Ni ) = max cmin,i , (14)
Ni · M P Ci

Finally, the range of Ni to explore may be derived from expressions (2), (5)
and (12):
  
2 · π · Ri
Nmax,i = f loorPi min M Ni , (15)
mP Ci · cmin,i
   
2 · π · Ri
Nmin,i = ceilPi min Nmax,i , (16)
M P Ci · cmin,i
In these expression f loorPi () function is the largest integer value not greater
than the argument and multiple of Pi . Function ceilPi () is the smallest integer
value not less than the argument and multiple of Pi .
Summarizing, it has been demonstrated than we can reduce the problem to
only M DoF, which will be the number of airfoils of every row, Ni . Once Ni is
known, the ci value can be calculated from expression (14) and, in turn, the gi
value can be derived from expression (11).

3 Genetic Algorithm

Once it has been demonstrated that each design configuration is determined

by a set of NumberOffs, it could be possible to perform an exhaustive search
computing all the possible configurations. Using expressions (15) and (16), the
QM
number of configurations to explore will be i=1 (Nmax,i − Nmin,i ) /Pi . As it
will be showed in section 4, huge numbers appear in real problems.
Due to the multiple restrictions, it is difficult to define a continuous and
derivable optimization function. Therefore methods based in gradient of the op-
timization function are not recommended. Several characteristics of the problem
makes appropriate the use of a Genetic Algorithm (GA). First of all, it is easy
to transform a Constrained Optimization Problem (COP) into a Free Optimiza-
tion Problem (FOP). Secondly, the optimization function is not necessary to
be continuous. Thirdly, the solution codification is made easily using a numeric
vector.
The first step for defining a GA is to link the real world to the GA world. Ob-
jects forming possible solutions within the original problem context are referred
to as phenotypes, while their encoding are called genotypes. In our problem, the
phenotypes are vectors of natural numbers with the NumberOff for each row.
Each NumberOff only can change in the range given by (15) and (16). The en-
coding of each genotype is a bit string for each NumberOff. The number of bit
strings will be M , one for each row. Each bit string is called a gene. Each gene
may have a different number of bits
  
Nmax,i − Nmin,i
Bi = ceil1 ln / ln 2 (17)
Pi

The way of decoding the genotype to the phenotype consists of obtaining the
integer number ni from the bit string. A Gray coding is used instead of the usual
binary coding because of its advantages, described extensively in the literature
[12]. Then, the NumberOff will be Ni = Nmin,i + ni · Pi . Gaps and chords are
obtained using the expressions (11) and (14). Maybe it will be necessary to
modify the gaps in order to fulfill restriction (7).
As it was mentioned in section 2, equation (11) does not take into account
the constraint (7). A repairing process may be necessary if, in obtaining the
phenotype, g0 does not meet that constraint. If mG0 ≤ g0 ≤ M G0 the fixing is
not necessary. On the other hand, if g0 ≤ mG0 it is not possible to repair and
the individual receives a high penalty in its fitness. If g0 > M G0 a repairing
process is needed. The repairing process is made in the phenotypic space and
this consists of distributing the amount 4g = g0 − M G0 among the the rest of
the gaps maintaining the constraints gi ≤ M GCi · ci .
The next step in the design process of the GA is to choose the fitness function.
The role of the fitness function is to represent the requirements to be optimized.
The fitness function implemented transforms our initial COP into a FOP. With
the representation of individuals adopted, all the constraints are satisfied but
(7), (8), (9) and (10). Being M the number of rows, a penalty function FR is
defined as 0 for individuals placed in the feasible regions and negative values
increasing exponentially in the following way

   P
mG0 −g0 M
 1 − exp λ L

 + i=1 Fi if g0 < mG0
PM
FR =
 i=1 F
i if mG0 ≤ g0 ≤ M G0 (18)

 1 − exp λ g0 −M G0 + PM F

if M G0 < g0
L i=1 i

where g0 is computed using expression (7). The value of constant parameter λ

is used for modulating the exponential decreasing in the unfeasible regions. Its
value is taken experimentally and does not have a big influence in the perfor-
mance of the algorithm. Fi deals with the noise restrictions depending of the
cut-off mode. For instance, for Mixed cut-off mode and i%2 = 0 or Reverse
cut-off mode and i%2 = 1 :
 h  i

 1 − exp λ αi − NNi+1
i
if NNi+1
i
< αi

Ni
Fi = 0 if αi ≤ Ni+1 ≤ βi (19)
 h  i
 1 − exp λ Ni − β

if β < Ni
Ni+1 i i Ni+1

In order to perform an simple optimization among the feasible individuals, the

fitness functions is defined:
(
FR if FR < 0
FN = PN (20)
1/ i=1 Ni if FR ≥ 0

GA uses a population of possible solutions. The parent selection mechanism

implemented here is the tournament method, i.e. k individuals with replacement
are chosen randomly from the population and the final individual chosen will be
the best of these k in terms of their fitness value. Once the parents have been
selected, there is a recombination probability pr which sets if the offspring of two
parents are just a copy of the parents or a real recombination is produced. Two
recombination methods have been implemented: one point crossover over the
genes (called gene crossover ) and one point crossover over the bits for each gene
(bit crossover ). Another parameter that control the algorithm is the mutation
probability pm . After doing the crossover of the parents, the offspring is mutated.
The mutation is done in each gene swapping a random bit. A generational model
is used, so for each generation all parents are changed by their offspring. It has
been implemented the feature of elitism, swapping the worst individual after the
mutation operator is applied by the best individual of the previous generation.
The initialization is made by taking a random representation of possible
solutions from the design space and carrying out fitness evaluations on all the
individuals.

4 Case study and Experimental Results

The GA described in the previous section has been applied to an aeronautical

LPT made up of 6 stages and an Outlet Guide Vane (OGV), which makes a
total of 13 rows (Fig. 1). The total number of airfoils of this turbine is 1486 and
its axial length is 1.225 meters. The restrictions imposed has been the same used
by the design team, so the study case must be considered a real one. Only the
Reverse cut-off condition is considered feasible.
Using the expressions (15) and (16), it is possible to compute the possibilities
of each NumberOff. The number of possible values varies greatly between each
row. The minimum number is 1 possibility for the first and last row, and the
maximum value is 63 for the fourth row. Multiplying all the possibilities we
obtain a total number of configurations to explore of 2.8 · 1018 . If we use a
exhaustive search and consider that each configuration was evaluated in 10−6
seconds, the computing time would be 73117 years. So an exhaustive search can
not be used in this case.
The parameters of the GA have been chosen by trial and error. The best
results are obtained for a population of 5 · 104 individuals, 50 generations, k = 5,
pr = 0.8, pm = 0.01, gene crossover and elitism. Due to the stochastic nature
of the GA, the algorithm has been run 5 times. The average total NumberOff
achieved has been 1461.4, with a standard deviation of 1.96. The best total
NumberOff has been 1460. The average time needed has been 2 minutes and 22
seconds using a 2.40 GHz Intel Core Duo machine with a 4 GB of RAM memory
and a operative system Linux openSUSE 10.3.
Fig. 3a shows the NumberOff for each row of the original design and the
new ones obtained by the best run of the GA. Several configurations with the
same total NumberOff of 1460 have been found. All these solutions are indistin-
guishable by the algorithm. As is clear from the interpretation of equation (20),
the evolving process involves two phases: an initial stage (FR < 0) dedicated to
satisfying the constraints, and a second stage (FR ≥ 0) devoted to minimizing
the total number of airfoils once the constraints are satisfied. This can be seen in
Fig. 3b, where the evolution of the best fitness of the population in each run is
plotted against the generation number. Fig. 3c is an enlarged view of Fig. 3b and
shows the border between these two stages. More specifically, 15 to 19 genera-
tions are needed for having at least one individual that meets all the restrictions.
In Fig. 3d it is showed the new real geometry in meridional plane obtained by
the GA. Notice that the algorithm has changed the chord and the gap of some
rows in order to fulfill the constraints.

a) b) c)
150

-04
0

6.85e
-04
6.80e
NumberOff
100

-1
Fitness

-04
6.75e
Run 1
Run 2
Run 3

-04
Run 4

6.70e
Run 5
50

-2
Original Design, Total NumberOff = 1486

-04
Total NumberOff in the best run = 1460

6.65e
-04
0

-3

6.60e
5 10 0 10 20 30 40 50 10 20 30 40 50
Row Number Generation Generation
d)

Fig. 3. Summary of experimental results: NumberOffs achieved in the best run (a),
progress plots (b) and (c), where (c) shows a magnified region of (b), and optimized
real geometry (continuous line) compared with the original one (dashed line) (d).

5 Conclusions and Future Work

A GA has been applied to perform the airfoil number optimization of a LPT gas
turbine fulfilling a set of realistic restrictions. The turbine model used as input
to the GA corresponds to the final design of a turbine made from a standard
design methodology. The algorithm has reduced the total number of airfoils in
1.74%. The improvement is not bigger because the input to the GA is an al-
ready optimized final design using the same set of restrictions, so it was close
to the optimum. However, from the standpoint of the real turbine design, the
improvement is not negligible. With the appropriate parameters for the GA,
the optimization process found the same configuration with a very low disper-
sion (standard deviation in the total NumberOff less than 2 blades). The time
consuming of the algorithm is low, despite of dealing with a high population
number.
As a future work it would be possible to consider more ambitious goals. For
example, the total number of airfoils is a rough estimation of the weight and
efficiency of the LPT since other factors such as the chords, spans and thickness
are not taken into consideration. A more refine model should consider the trade-
off between efficiency and weight to include other optimization functions.

References
1. P. de la Calzada: Aerothermodynamic Design of Low Pressure Turbines. ”Aero-
engine design: from state of the art turbofans towards innovative architecture”.
Von Karman Institute Lecture Series LS-2008-04, Brussels, Belgium, 3-7 March
(2008)
2. A. Chipperfield, P. Flemming: Multiobjective Gas Turbine Engine Controller De-
sign Using Genetic Algorithms. IEEE Transactions on Industrial Electronics, Vol.
43(5), pp. 583-587 (1996)
3. I. C. Parmee, C. R. Bonham: Towards the Support of Innovative Conceptual Design
Through Interactive Designer/Evolutinary Computing Strategies. Artificial Intel-
ligence for Engineering Design, Analysis and Manufacturing, Vol. 14(1), pp.3-1614
(2000)
4. J. M. Rogero, A. Tiwari, O. Munaux, P.A. Rubini, R.Roy, G.Jared: Applications of
evolutionary algorithms for solving real-life design optimisation problems. In: 6th
Int. Conf. Parallel Problem Solving from Nature, Paris, France, September (2000)
5. I. C. Parmee: Evolutionary Computing Strategies for Preliminary Design Search
and Exploration. In: Proceedings US United Engineering Foundation’s ‘Optimisa-
tion in Industry’ Conference, Tuscany, Italy, 2001. Springer-Verlag, London (2002)
6. R. Roy , A. Tiwari , J. Corbett: Designing a Turbine Blade Cooling System Using
a Generalised Regression Genetic Algorithm. CIRP Annals, Vol. 52(1), pp. 415-418
(2003)
7. A. Angantyr, J. O. Aidanpaa: A Pareto-Based Genetic Algorithm Search Approach
to Handle Damped Natural Frequency Constraints in Turbo Generator Rotor Sys-
tem Design. J. Eng. Gas Turbines Power 126, 619, DOI:10.1115/1.1760529 (2004)
8. V. Ahuja, A. Hosangadi: Design Optimization of Complex Flowfields Using Evolu-
tionary Algorithms and Hybrid Unstructured CFD. In: 17th AIAA Computational
Fluid Dynamics Conference, Toronto, Ontario, Canada, 6-9 June (2005)
9. M. Parrilla, J. Aranda, and S. Dormido-Canto: Parallel Evolutionary Computation;
Application of an EA to Controller Design. Proceeding of IWINAC 2005. Springer-
Verlag Berlin Heidelberg, LNCS 3562, pp. 153–162 (2005)
10. L. Elliott, D. B. Ingham, A. G. Kyne, N. S. Mera, M. Pourkashanian, C. W. Wilson:
A Novel Approach to Mechanism Reduction Optimization for an Aviation Fuel/Air
Reaction Mechanism Using a Genetic Algorithm. J. Eng. Gas Turbines Power 128,
255, DOI:10.1115/1.2131887 (2006)
11. R. Varvill, G. Paniagua, H. Kato, M. Thatcher: Design and Testing of the Contra-
Rotating Turbine for the Scimitar precooled Mach 5 Cruise Engine. In: IAC-08-
C4.5.3 (2008)
12. A.E. Eiben, J.E. Smith: Introduction to Evolutionary Computing. Springer (2007)