Global Optimization Methods:

Theoretical Aspects & Definitions

R.A. Van den Braembussche

von Karman Institute for Fluid Dynamics
Waterloose steenweg, 72
1640, Sint-Genesius-Rode
vdb@vki.ac.be

ABSTRACT
The first part of this lecture describes the basics of global optimizations systems with emphasis
on those based on evolutionary strategies and the use of metafunctions to accelerate the
convergence. The basic method is illustrated by the optimization of a 2D turbine blade.
It is followed by a description on how to extend the method to a multidisciplinary optimization.
The latter is illustrated by the optimization of a radial compressor impeller. It is shown how such
a procedure can lead to innovative designs with high performance.
Further discussed and illustrated by an example are the procedures that may help to guarantee
the performance over a wide operating range (Multipoint Optimization).
The next chapter explains the notion of Robustness and how to verify that small changes in
operating conditions or design and manufacturing errors do not compromise the results.
The last part discusses the techniques that are available to handle designs that have more than
one objective (Multiobjective Optimization).

1. INTRODUCTION
The main goal when designing gasturbine components is to achieve light, compact and highly efficient
systems while reducing the cost and the duration of the design cycle. The traditional trial-and-error
process is now replaced by computerized design systems defining the optimum geometry for a required
performance. They make use of a search technique to find the geometry that is optimum according to a
performance prediction method, while respecting the design constraints.
Advanced design systems must allow full use of all the 3D geometrical features that may improve
performance i.e. lean and sweep. Any limitation of the geometry is acceptable only if it is imposed by
mechanical (stress), manufacturing or cost limitations. The optimal performance can only be guaranteed if
all the real flow phenomena are taken into account i.e. if the performance predictions are made by 3D
Navier-Stokes (NS) solvers. Any use of approximate (incorrect performance measuring systems) may lead
to a false optimum. However those analysis tools require a large amount of computer effort, leading to
excessive design cost. Hence methods must be developed that allow limiting the cost and time without
compromising on the result.
The system must also provide realistic designs i.e. that satisfy the mechanical and geometrical
constraints and guarantee the requested life time of the device. Satisfying all these objectives not only
requires the use of acoustics, stress and heat transfer analysis methods (multidisciplinary optimization)
but also requires a compromise to satisfy the conflicting objectives (multiobjective optimization). High
performance must also be guaranteed over the whole operating range (multipoint optimization). A fast

RTO-EN-AVT-167 3-1
Global Optimization Methods: Theoretical Aspects & Definitions

and fully automated design system is required to achieve all this within a limited time and cost
Following describes optimization systems that reach these objectives in an efficient way i.e. with
improved convergence, while taking into account the design and off-design operating conditions and
constraints imposed by other disciplines. These methods are illustrated with different examples related to
turbomachinery.
An optimization system consists of following components:
• A parameterized definition of the geometry
• An Objective Function (OF) expressing the design goals in a mathematical way.
• A performance prediction system, inclusive automatic grid generators, to providing the input for
the OF for each newly proposed geometries
• A search mechanism that defines the design parameters that correspond to the best performance
while satisfying the constraints (geometrical, mechanical, etc.)

2. SEARCH MECHANISMS
There are two main groups of search mechanisms:
The analytical ones, who calculate the required geometry changes in a deterministic way from the output
of performance evaluations. A common one is the steepest descend method approaching the area of
minimum OF by following the path with the largest negative gradient on the OF surface (Fig 1). This
approach requires the calculation of the direction of the largest gradient of the OF and the step length. A
comprehensive overview of this type of optimization techniques is given in the lecture of J. Peter [1].
Zero-order or stochastic search mechanisms require only function evaluations. They make a random or
systematic sweep of the design space or use evolutionary theories such as Genetic Algorithms (GA) or
Simulated Annealing (SA) to find the optimum parameter combination. Zero order methods may require
more evaluations than gradient methods but the latter have more chances to get stuck in a local minimum.
The present chapter concerns methods using zero order search mechanisms in combination with systems
that allow reducing the computational effort by reducing the number of evaluations.
A systematic sweep of the design space, defining v values between the maximum and minimum
limits of each of the n design variable requires vn function evaluations. Fig.1 illustrates how such a
systematic sweep, calculating the OF for 3 different values of X1 and X2, provides a very good estimation
of where the optimum is located with only 9 function evaluations. This method is a valid alternative for
analytical search methods for small values of n. However it requires more than 14. 106 evaluations for n
=15.

Fig. 1 Zero order sweep of the 2D design space

3-2 RTO-EN-AVT-167
 Global Optimization Methods: Theoretical Aspects & Definitions

Evolutionary strategies such as GA and SA can accelerate the procedure by replacing the
systematic sweep by a more intelligent selection of new geometries using in a stochastic way the
information obtained during previous calculations. Simulated Annealing (SA) is derived from
the annealing of solids [2]. At a given temperature, the state of the system varies randomly. The
new state is immediately accepted if it has a lower energy level. If however the variation results
in a higher level state, it is accepted only with a probability Pr that is function of the temperature.
Eopt − Eact

Pr = e T
As the temperature decreases, the probability of accepting a higher state becomes lower. In a
simulated annealing algorithm, the design parameters characterize the state of the system whereas
the objective function characterizes the energy level.
The method presented in present paper uses a Genetic Algorithm to find the optimum.
This is a numerical technique, which simulates Darwin's evolutionary theory stating that the fitter
survives [3]. According to this theory, an individual (geometry) with favorable genetic
characteristics (design variables) will most likely produce better offsprings. Selecting them as
parents increases the probability that individuals of the next generation will perform better than
the previous one. The method has been developed by Prof. Ingo Rechenberg (Berlin, 1964) [4]
who optimized an articulated plate with 5 degrees of freedom (design parameters) for minimum
drag. Each articulation can take 53 values which results in a total number of
551 = 345 025 251 possible geometries. The solution is quite obvious but difficult to find in a
mathematical way.

Fig. 2 Articulated plate of Ingo Rechenberg

In a standard binary coded GA, the real valued design parameters Xi, defining the geometry are
jointly represented by a binary string.

The substring length l defines the number of digits of a design parameters can take. n is the number of
design parameters. Low values of the substring length decrease the optimization effort by limiting the
possible number of solutions, but the GA may not be able to accurately locate the minimum because of a
too low resolution.

RTO-EN-AVT-167 3-3
Global Optimization Methods: Theoretical Aspects & Definitions

Fig. 3 Schematic of Genetic Algorithm

The operational principle of a standard GA is shown in Fig. 3. Pairs of individuals (parents) are
selected from an initially random population of N geometries; each one is represented by a binary coded
string of length n.l. Genetic material is subsequently exchanged between them (crossover), altered within
the offspring (mutation), followed by an evaluation of each new individual. This process is repeated to
create the N individuals of the next generation. The whole procedure is repeated for tmax generations and it
is assumed that the best individual of the last generation is the optimum.
The GA software can be found on the Webb. The quality of the GA optimizer is measured by:
• the required computational effort i.e. the number of performance evaluations that are needed to
find that optimum (GA efficiency).
• the value of the optimum (GA effectiveness).
The tuning of the GA parameters (N, l, t) to accelerated the convergence will be the subject of a second
lecture [5]
The main issue of the GA is the selection scheme. One of the many selection schemes that have
been proposed is the roulette (Fig. 4left): a system in which the chance that an individual is selected
increases proportional with 1/OF. This scheme favors the best individuals as parent. It is elitist and has
larger chances to get stuck in a local optimum.

Fig. 4 Roulette (left) and tournament (right) selection

3-4 RTO-EN-AVT-167
 Global Optimization Methods: Theoretical Aspects & Definitions

In the tournament selection (Fig. 4right), S individuals are chosen randomly from the population and
the individual with the lowest OF is selected as parent. The same process is repeated to find the second
parent. The parameter S is called the tournament size and can take values between 1 and N. Larger values
of S give more chances to the best samples to be selected and to create off-springs. It favors a rapid, but
maybe premature, convergence to a local optimum. Too small values of S result in a more random
selection of parents. Tests have shown that a standard value of S=2 gives the best
results.
Zero order search methods, even supported by evolutionary theory, also require an excessive
number of performance evaluations. This becomes prohibitive in cases with expensive performance
evaluators. One way to speed up the convergence is by working on different levels of sophistication and
by making better use of the knowledge, gained during previous designs, for subsequent ones. This is
achieved by using fast but approximate prediction methods to find a near optimum geometry, which is
then further verified and refined by the more sophisticated but also more expensive analyzer.

Fig. 5 Flowchart of optimization system

Such a system is illustrated by the flow chart on Fig. 5. [6]. The fast but less accurate optimization
loop is indicated in red, the expensive but accurate one in bleu. The OF minimized by the GA is predicted
by means of a Metafunction or surrogate model i.e. an interpolator using the information contained in the
Database to correlate the performance to the geometry similar to what is done by a Navier-Stokes solver
(NS). Surrogate models have the same input and output as the analysis method they replace. Once they
have been trained on the data contained in the Database, they are very fast predictors and allow the
evaluation of the OF of the many geometries, generated by the GA, with much less effort than a NS
solver. Unfortunately the prediction is not always very accurate and the optimized geometry must be
verified by means of a more accurate but time consuming NS solver. The results of this verification are
added to the Database and a new optimization cycle is started. It is expected that a new learning on the
extended Database will result in more accurate metafunction and that the result of the next GA
optimization will be closer to the real optimum. The optimization cycle is stopped once the ANN
performance is in agreement with the NS calculations i.e. once the GA optimization has been made with
an accurate performance predictor.
The main advantage of this iterative procedure is the fact that, once the system is converged, there will
be no discrepancy between the results of a Metafunction prediction and the one obtained by a Navier-
Stokes calculation. Such an agreement is not guaranteed if a correlation or simplified solver (Euler or NS
on course grid) are used because the inaccuracy is not released during the design process This might drive

RTO-EN-AVT-167 3-5
Global Optimization Methods: Theoretical Aspects & Definitions

the GA to a false optimum.

The accuracy of the Metafunction is a major factor defining the convergence of the system to the
optimum. A GA optimization with an accurate Metafunction would result in a one step optimization,
requiring only one extra performance analysis. The parameters influencing this convergence are discussed
in a separate lecture [5].
The main purpose of the Database is to provide information about the relation between the geometry
and the performance. The more general and complete this information, the more accurate may be the ANN
and the closer the first optimum geometry, defined by the GA, will be to the real optimum. Hence a good
Database may considerably speed up the convergence to the optimum.
Any approximating function can be used as metafunction. Popular ones are: Response surface,
Artificial Neural Network (ANN), Radial Basis Functions (RBS), Kriging, etc. They will be explained in
more detail in a later lecture [5] together with the way to define a more representative Database.

3. 2D TURBINE BLADE OPTIMIZATION

The convergence speed is also strongly influenced by the number of unknown that are needed to
define the optimum geometry. Selecting parameters that have a direct relation to the performance, such as
blade angles, pitch to chord ratio, etc. provide a more straightforward relation between geometry and
performance. The corresponding ANN is simpler and more easily found. Hence less iterations may be
needed to reach agreement with the NS predictions. Another important characteristic is the continuity of
curvature of the blade contour because any discontinuous change in curvature may result in a local
velocity peak.

3.1 Parameterization
A good geometry definition avoids the generation of unrealistic blades while having enough geometrical
flexibility to represent a large number of blade types. The latter is very important because the best
geometry can only be found if it can be generated by the system.

Table 1 2D turbine design parameters

Parameterization is illustrated by the single point optimization of a turbine blade with outlet Mach
number .9. The design requirements are summarized in Table 1. They include the operating conditions (β1
and M2), flow characteristics (Re, γ, Tu (%) and geometrical constraints (Cax, pitch and trailing edge
thickness, maximum cross section area, minimum and maximum moment of inertia and direction),
required performance (β2 and losses).
The blade geometry (Fig. 8) is specified by four key points (LE, 2, 3, 5) linked by four curves. The
points LE and 2 are linked by a Bézier curve defined by three additional points. The lasts ones are located
in such a way to ensure continuity up to the third derivative at point 2 and continuity up to the first
derivative in point 3. A Bézier Curve with three additional polygon points is used to define the pressure

3-6 RTO-EN-AVT-167
 Global Optimization Methods: Theoretical Aspects & Definitions

side in the same way as the first part of the suction side. This curve is fully defined by β1blade, Rle, L4, αps,
L3 and the tangent to point 2. Using the same Rle and L4 for both the suction and pressure side guarantees
the continuity of the curvature radius at the leading edge. The trailing edge is defined by part of a circle
whose radius Rte is specified by the Trailing edge thickness. The 2D blade geometry is thus fully defined
by means of 15 parameters represented by G(n) : n=1,15

Fig. 8 Parameterized definition of 2D turbine blade

Fig. 9 shows 4 different types of turbine blades generated with this geometry model. For each blade
one parameter is changed and its influence on the blade shape is shown. This figure demonstrates that the
method is capable of representing the large variety of turbine blades encountered in industrial designs.

Fig. 9 Different parameterized blade geometries

3.2 Objective Function (OF)

The OF based on Navier-Stokes results predictions measures in how far the geometry satisfies the
Aero-requirements and reaches the performance goals that have been set forward. The same OF, but based
on ANN results, drives the GA towards the optimum geometry.
High efficiency however is not the only objective of an aerodynamic shape optimization. A good

RTO-EN-AVT-167 3-7
Global Optimization Methods: Theoretical Aspects & Definitions

design must respect the mechanical and manufacturing constraints. Some constraints must be satisfied
without any compromise (i.e. maximum stress level) whereas others tolerate some margin (i.e. cost or
weight) or can be corrected for after the design (adjusting the blade height to achieve the required mass
flow).
A possible way to satisfy objectives and constraints is by defining a pseudo-OF by summing up the
penalty terms that are increasing when the constraints are violated [7]. This does not guarantee that each
individual constraint will be satisfied but contributes to a easier convergence to the constrained optimum.
Following lists some possible contributions to the pseudo OF.

OF2 D = w perf ⋅ Pperf + wa ⋅ PaeroBC + wM ⋅ PMach + wG ⋅ PGeom + wS ⋅ PSide

Pperf is the penalty for non optimum performance i.e. low efficiency (η) or high losses.
Pperf = max[1 − η ,0.0]

PAeroBC is the penalty for violating the aerodynamic boundary conditions. The purpose of this
penalty is to enforce the design targets, such as the outlet flow angle (β2) or the mass flow etc. These
penalties start increasing when the actual value differs from the target value by more than a predefined
tolerance. Following is a typical expression for mass flow penalty:

[ ( )]
Pmass = wi . max m& act − m& req / m& req − .02 , 0. 2

i.e. the penalty starts increasing when the error exceeds 2% of the required mass flow. The rate of increase
is defined by wi .

PMach is the penalty for non-optimum Mach number distribution. Analyzing the Mach number
distribution may help to make a selection between blades that have nearly the same loss coefficient by
decreasing the uncertainty due to transition predictions, or to favor Mach number distributions that are
likely to perform better at off-design (see section 4. Multidisciplinary optimization).

PGeom is the penalty for violating the geometrical constraints. These are the constraints that do
not influence the mechanical integrity but restrict maximum length and camber or assure dimensional
agreement with other components. Another reason to introduce geometrical constraints may be to favor
geometrical features that are known to improve the design or off-design performance i.e. progressive
change of curvature, some prescribed lean or sweep laws, limiting camber of the uncovered turbine
suction side, etc .

PSide is the penalty for violating any other constraint that might be imposed depending on the
application i.e. weight, manufacturing and maintenance cost, etc

3.3 Results
The best blade of the initial database is used as starting geometry. The Mach distribution predicted
by the NS and ANN are compared on Fig. 10a. The agreement is not perfect but the main features such as
a shock at mid chord are predicted by the ANN.
Fig. 10b compares the value of the OF predicted by the ANN with the one predicted by the NS
solver during the design process. The value of the OF computed by the approximate model decreases until
iteration 13 after which only very small improvements are found. The value predicted by the NS solver
shows large discrepancies between both predictions at iteration 2, 5 and 9. It indicates that during these
first design iterations, the ANN predictions are not very accurate because the Database does not
sufficiently cover the relevant design space. However this shortcoming is remediated by adding new

3-8 RTO-EN-AVT-167
 Global Optimization Methods: Theoretical Aspects & Definitions

geometries to the Database. As these new blades are close to the desired operating point they provide very
valuable information and the ANN becomes more and more accurate. This convergence to the same OF
value illustrates the self learning capacity of the proposed procedure. Starting from iteration 13 the ANN
predictions are very reliable. The whole procedure could have been stopped after 15 iterations but has
been continued to verify the good convergence.

Fig. 10 Comparison of (a) Mach number distribution predicted by Navier-Stokes solve and
ANN, trained on initial database and (b) between the ANN and the NS predicted OF

Figure 11 shows the variation of the Mach number distribution and blade shape during the design
process. The small constant velocity region on the suction side close to the leading edge and the low
velocity on the pressure side close to the leading edge indicate that the incidence angle on the initial blade
is too large. After the first modification (one GA and NS verification), this incidence angle has been
partially reduced by decreasing the stagger angle. The shock intensity is also smaller but the suction side
Mach number distribution is still wavy. The shock has completely disappeared after 13 design iterations.
The stagger angle has decreased in order to adapt the blade geometry to the prescribed inlet flow angle.
The smooth shock free Mach number distribution is reflected in the low loss coefficient (Table 1).

Fig. 11 Evolution of the Mach number distribution and geometry during optimization

RTO-EN-AVT-167 3-9
Global Optimization Methods: Theoretical Aspects & Definitions

4. MULTIDISCIPLINARY OPTIMIZATION.
Most mechanical constraints such as maximum stress and deformation have a direct impact on the
turbomachinery integrity and must therefore be rigorously respected. Hence they cannot be imposed by a
weak formulation as done for the geometrical constraints. Some of those constraints can easily be
respected by a simple limitation of a design parameter. Bird ingestion resistance is often expressed by a
minimum leading edge radius (RLE) . Corrosion may define the minimum trailing edge radius (RTE) and
blade thickness (Bthickness). However most of the mechanically unacceptable geometries result from a
combination of different parameters and cannot be avoided by reducing the feasible range of the
individual design parameters. Hence, a large percentage of the design space will consist of
geometries violating the constraint.
A possible approach is a verification of the mechanical constraints by a Finite Element Analysis
(FEA) before starting the Aero analysis on every geometry proposed by the GA. This sequential analysis
is very time consuming and is easily be replaced by a parallel analysis shown on Fig. 12. It is an extension
of the flow chart shown in Fig. 5. The GA, searching for the optimum geometry, gets its input from the
Finite Element stress Analysis (FEA) as well as from the NS flow analysis. The same type of extension
could also be made for the constraints related to heat transfer, acoustics, weight limitations, etc.

Fig.12 Multidisciplinary optimization flow chart

The main advantages of such an approach are:

• The existence of only one “master” geometry i.e. the one defined by the geometrical parameters used
in the GA optimizer. This eliminates all possible approximations and errors when transmitting the
geometry from one discipline to another.
• The existence of a global OF accounting for all disciplines. This allows a more direct convergence to
the optimum geometry without iterations between the aerodynamically optimum geometry and the
mechanically acceptable one.
• the possibility to do parallel calculations. The different analyses can be made in parallel if each
discipline is independent i.e. if stress calculations do not need the pressure distribution on the vanes or
flow calculations are not influenced by geometry deformations.

3 - 10 RTO-EN-AVT-167
 Global Optimization Methods: Theoretical Aspects & Definitions

The computational effort increases proportional with the number of different analyses that are
needed for the performance evaluation of the GA proposed geometries. It can be drastically reduced if one
can eliminate the unfeasible geometries before any expensive flow analysis is started i.e. if, in analogy
with the aero analysis, one can formulate an approximate prediction model for mechanical characteristics,
heat transfer, etc. to drive the GA
The multidisciplinary optimization is illustrated by the design of a radial compressor impeller for a
micro-gasturbine application with a diameter of 20 mm rotating at 500,000 rpm..

4.1 Geometry Definition

The 3D radial impeller is defined by the meridional contour at the hub and shroud (Fig 13), the
camber line of the main and splitter blade (Fig 14), the thickness distribution normal to the camber line
(Fig. 15) and the number of blades.
The hub and shroud meridional contours from the leading to the trailing edge are defined by third
order Bézier curves (Fig 13). The coordinates of the control points are geometrical parameters that can be
changed by the optimization program. Only 6 parameters are needed to define the meridional contour.
Each of them has a limited range in which it can vary. The possible variations of the individual Bézier
control points are shown by arrows in Fig. 13. The control point at the hub trailing edge is fixed by the
prescribed outlet diameter and axial length. Most control points have only one degree of freedom because
they are linked to other parameters in order to guarantee an axial inlet or radial outlet. The shroud leading
edge diameter defines the variable impeller inlet height. Third order Bézier curves define also the inlet
duct. Their control points are automatically adjusted to obtain a smooth link between the given radial inlet
and impeller.

Fig. 13 Meridional contour defined by Bézier control points.

The blade camber lines at hub and shroud are defined by the distribution of the angle β(u) between
the meridional plane m and the blade camberline (Fig. 14). The β distributions at hub and shroud are
defined by third order polynomials:
β (u ) = β 0 (1 − u )3 + 3β1u (1 − u )2 + 3β 2 u 2 (1 − u ) + β 3u 3
with u the non-dimensional meridional length ( u ∈ [0,1], 0 at the leading edge and 1 at the trailing edge).
The camber line circumferential position θ (Fig. 14) is then defined by integration of:
R dθ = dm tan β
β0 and β3 are the blade angles at leading- and trailing edge. This definition is used for both the main-
and splitter blade, at hub and shroud. The splitter trailing edge angles are the same as the full blade values
at hub and shroud. This results in 14 design variables for the blade camber line definition.
The streamwise position of the splitter blade leading edge is also a design parameter. It is defined as
a percentage of the main blade camber length and can vary between 20% and 35%.

RTO-EN-AVT-167 3 - 11
Global Optimization Methods: Theoretical Aspects & Definitions

Fig. 14 Definition of the blade camber line by β angle.

The blade thickness distributions at hub and shroud are functions of two parameters: the thickness
“LE” of the ellipse defining the leading edge and the trailing edge thickness “TE” (Fig. 15). The blade
thickness is kept constant at the shroud (LE=TE=0.3 mm). The two parameters defining the blade
thickness at the hub are design parameters and can vary between 0.3 to 0.6 mm. The same values are used
for the main and splitter blade.

Fig. 15 Thickness distribution along the camber line of the blade (not to scale).

The number of blades could also be a design parameter to be optimized, but has been fixed to 7 for
manufacturing reasons. This brings the total number of design parameters to 23.

4.2 Analysis programs

The TRAF3D Navier-Stokes solver [8] is used to predict the aerodynamic performance of the radial
compressors. The computational domain starts at constant radius in the radial inlet (Fig. 13) and ends in
the vaneless diffuser at r/r2 = 1.5 .
A structured H-grid with 2x216x48x52 or approximately 1,080,000 cells is used for all
computations to guarantee a comparable accuracy for all the samples stored in the database. The total inlet
temperature is 293°K and the total inlet pressure is 1.013E+5 Pa. The design mass flow is 20 g/s. The wall
temperature of the impeller is fixed at 400º K, as found in a previous study on the heat transfer inside the
entire micro gas turbine [9].
The commercial code SAMCEF [10] is used for the stress calculation. Quadratic tetrahedral
elements are used as a compromise between element quality and automatic meshing. Similar grids with
250,000 nodes and 160,000 elements are used for all samples. The grid is refined in areas of stress
concentrations.
The impeller tip speed of 523.6 m/s results in very high centrifugal stresses. Titanium TI-6AL-4V
has been selected for its high yield stress over mass density ratio (σyield/ρ). The characteristics are:
Elasticity modulus = 113.8E+9 Pa, Poisson modulus = .342 and mass density = 4.42E+3 kg/m3.
A fillet radius of 0.25 mm is applied at the blade hub to limit the local stress concentrations. The
unshrouded impeller has a tip clearance of 0.1 mm, which is 10% of the exit blade height. This is typical
for these small impellers and one of the reasons for the moderate efficiencies.

3 - 12 RTO-EN-AVT-167
 Global Optimization Methods: Theoretical Aspects & Definitions

4.3 Objective Function

The OF is the weighted sum of several penalties:
r r r r r
( ) ( ) ( ) ( ) ( )
OF G = wstress ⋅ Pstress G + wη ⋅ Pη G + wmassflow ⋅ Pmassflow G + wMach ⋅ PMach G

The first penalty concerns the mechanical stresses.

⎡σ − σ allowable ⎤
Pstress = max ⎢ max ,0.0⎥
⎣ σ allowable ⎦
where σmax is the maximum stress in the impeller. This penalty is zero when the stresses are below the
allowable limit σallowable and increases linearly when the von Mises stresses exceeding that value. This
weak formulation of the constraints does not guarantee that they are fully respected. However, it has the
advantage that all geometries that have been analyzed provide information that leads towards the optimum
geometry.
The efficiency and Mass flow penalties are similar to the ones used for the 2D turbine. The penalty
on the Mach number aims to favour Mach number distributions that are expected to be good at design
point and remain good at off-design operation. It also has two contributions. The first one penalizes
negative loading and is proportional to the area between the suction and pressure side when the pressure
side Mach number is higher than the suction side one (Fig. 16).
1
[
PMach = ∫ max M ps (s ) − M ss (s ),0.0 ⋅ ds ]
0

Fig. 16 Negative loading and loading unbalance in a compressor with splitter vanes.

The second Mach penalty increases with the loading unbalance between main blade and splitter
blade. This penalty compares the area between the suction- and pressure side Mach number distribution of
main blade Abl and splitter blade Asp, corrected for the difference in blade length (Fig. 16):
2
⎛ Abl − Asp ⎞
Ploading unbalance =⎜ ⎟
⎜A +A ⎟
⎝ bl sp ⎠
The weight factors of the OF are determined based on the knowledge gained in previous
optimizations. The values used in present design are such that an efficiency drop of 1% is as penalizing as
an excess in stress limit of 6.668 MPa.

RTO-EN-AVT-167 3 - 13
Global Optimization Methods: Theoretical Aspects & Definitions

4.4 Results
The optimization starts from the outcome of a simple aerodynamic optimization without stress
computation. It is called the “Baseline” impeller. Although this geometry has a good efficiency, it cannot
be used because a mechanical stress analysis predicts von Mises stresses in excess of 750 MPa. It serves as
a reference for further optimizations.
An initial database containing a total of 53 geometries is used at the start. 13 geometries out of the
64 initial ones defined by the DOE technique [5] could not be analyzed because of geometrical constraints
(intersection of the main blade with the splitter blade). Two additional geometries have been added,
namely the baseline geometry and the central case. The latter one is a geometry with all parameters at 50%
of their range.
Fig. 17 shows the convergence history of the optimization. The “aero penalty”, based on efficiency,
Mach number and mass flow, the “stress penalty” and “total penalty”, obtained from the Navier-Stokes
and FEA calculations, are compared to the ones predicted by the ANN. One observes a decrease of the
discrepancy between both prediction methods with iteration number. This is the consequence of an
increasing number of samples in the database, resulting in a more accurate ANN.
Only 10 iterations are needed to obtain a very good agreement for the aero penalties. The ANN
stress penalty is zero for every geometry proposed by the GA. However it takes more than 15 iterations
before the FEA confirms that the proposed geometries satisfy the mechanical requirements.
The good agreement in both stress- and aero penalties, over the last 18 iterations, indicates that the
ANN predictions are reliable. It means that the same optimum geometry would have been obtained if the
GA optimization had been driven by the more sophisticated NS and FEA analyses. Hence no further
improvement can be expected. The optimization procedure could already have been stopped after 35
iterations.

Fig. 17 Convergence history of the optimization.

The aero penalty is plotted versus the stress penalty in Fig. 18. The geometries created during the
optimization process are all in the region of low penalties. Most of them outperform the geometries of the
database. Only a few geometries of the optimization loop have penalties of the same order as the database
samples. Those geometries are the ones created during the first 10 iterations where the ANN is still
inaccurate.

3 - 14 RTO-EN-AVT-167
 Global Optimization Methods: Theoretical Aspects & Definitions

Fig. 18 Aero penalty versus stress penalty for baseline, database- and optimization geometries.

Fig. 19 is a zoom on the low penalty region of Fig. 18. A large number of geometries have zero
stress penalties but with different aero penalty. The geometries corresponding to iteration 17, 49 and 25
have the lowest aero penalty.

Fig. 19 Zoom on the low penalty region of Fig. 18.

From all geometries created during the optimization, iteration 25 performs best. It has a little lower
efficiency than iteration 17 but less loading unbalance and the stresses are 33 MPa below the limit. In spite
of its high efficiency the baseline impeller shows a high aero penalty because of a too high mass flow. The
influence of the stress penalty on the optimization is clear by comparing the values of the baseline impeller
with the ones of iteration 25. The reduction of the maximum stress level with 370 MPa is at the cost of a
2.3 % decrease of efficiency (21).
Figure 20 compares the von Mises stresses in the baseline geometry with the ones of iteration 25.

RTO-EN-AVT-167 3 - 15
Global Optimization Methods: Theoretical Aspects & Definitions

The drastic reduction in stress is the consequence of:

• the reduced blade height at the leading edge, resulting in lower centrifugal forces at the leading
edge hub
• the increase of blade thickness at the hub
• the modified blade curvature resulting in less bending by centrifugal forces

Fig. 20 von Mises stresses due to centrifugal loading in the baseline (left) and optimized (right) impeller.

The blade lean is defined as the angle between the blade leading edge and the meridional plane
(positive in the direction of rotation). It is a result of the integration of the β distribution at hub and shroud,
while limiting the rake at the outlet. Its impact on stress and efficiency is shown in Fig. 21. One observes a
rather clear relation between lean and stress. It shows that the lowest stresses can be expected around -
15.0º. Several geometries with good efficiency are found for lean angles between -40.0º to -5.0º. The drop
in efficiency for lean angles above -5.0 º suggests that in present application the best impellers have some
negative lean. This unexpected result is a major outcome of the optimization.

Fig. 21 Blade lean versus stress and efficiency for database and optimization geometries.

3 - 16 RTO-EN-AVT-167
 Global Optimization Methods: Theoretical Aspects & Definitions

5. ROBUSTNESS
Robustness characteerizes the insensitivity of the performance to small changes in operating
conditions or geometrical changes (manufacturing inaccuracies).
Small variations of incidence may trigger flow separation in turbomachinery components operating
at the verge of separation (limited diffusion factor) and be at the origin of large performance variations. As
will be demonstrated later this problem can be avoided by a multipoint optimization. Small changes in
operating conditions or flow characteristics may result in large performance changes due to a large shift of
the transition point. This can be avoided by adding penalties on the predicted Mach number distribution to
account for the expected changes at off-design operation.
Small geometrical changes or manufacturing inaccuracies should not influence the performance.
The results of the many geometries that have been analysed during the optimization process provide an
indication of robustness. Figure 21 shows that a change in blade lean around the optimum value has
almost no effect on stresses and efficiency. The design is robust in this respect.
Fig. 22 shows the stress and efficiency versus blade thickness at the leading edge for all geometries.
Non-dimensional values 0 and 1 correspond to a blade thickness of respectively 0.3 and 0.6 mm. The
database geometries are at 25% or 75% of the range suggest that the blade hub thickness does not have
much impact on efficiency. Hence thicker blades are selected because it lowers the stresses. The figure
also shows that thicker blades are more likely to have a lower stress than thin ones and that the stresses are
not very sensitive to variations near the maximum thickness.

Fig. 22 Stress and efficiency versus blade thickness for database and optimization geometries.

6. MULTIPOINT OPTIMIZATION

Multipoint optimization aims for a design that performs well in more than one operating point. The
simplest straight forward approach is to analyze every candidate geometry at the different operating
conditions and to calculate a weighted value of the performance. This is not only expensive in terms of
flow analysis but often compromised by practical problems.
When doing multipoint optimization one should make distinction between cases with varying inlet
conditions or varying outlet conditions (back pressure). In the latter case one cannot a priori guarantee that

RTO-EN-AVT-167 3 - 17
Global Optimization Methods: Theoretical Aspects & Definitions

the proposed component can operate steadily at each point i.e. that the NS solver can provide a converged
solution at the required operating conditions. In fact, it is not a priori known at what pressure ratio a
compressor will surge because the change in mass flow with pressure ratio depends on the still unknown
performance curve of the compressor. Less problems occur at the maximum mass flow side because any
low back pressure provides a good idea of the choking or maximum mass flow and does not create
stability problems.
The stability problem is less likely to occur for turbines. Operating with a favorable pressure gradient
facilitates the convergence of the N.S. calculations and changing inlet angle of back pressure is less of a
problem.
Variation of the inlet conditions is less risky in terms of calculation stability than change in outlet
conditions.
Following describes an application where the off-design corresponds to a known variation of the inlet
conditions and where the outlet conditions have less impact on the procedure. It is followed by a
description of the procedure that has been put in place to do a multipoint optimization for compressors
where the surge point prediction is a major issue.

6.1 Multipoint optimization of a Low Solidity Diffuser

Radial compressors with vaned diffusers provide high pressure recovery and efficiency but the
operating range is limited by stall, at positive incidence, and diffuser throat choking, at negative incidence.
Vaneless diffusers do not limit the maximum mass flow but have lower efficiency and minimum mass
flow may be limited by vaneless diffuser stall. Low Solidity Diffusers (LSD) are characterized by a small
number of short vanes and do not show a well defined throat section. They intend to stabilize the flow at
low mass flow without limiting the maximum mass flow by choking. The solidity (chord/pitch) is
typically of the order of 1 or less (Fig. 23). A multipoint optimization is mandatory for the LSD design
because achieving a wide operating range is the major target.

Fig. 23 Low Solidity Diffuser

The optimization of the LSD described here is done for the 3 operating points listed in Table 2 .
Inlet conditions are different for each operating point because they result from a different operation point
of the impeller. It is not possible to impose a pressure ratio corresponding to the impeller mass flow

3 - 18 RTO-EN-AVT-167
 Global Optimization Methods: Theoretical Aspects & Definitions

because maximizing the diffuser design- and off-design pressure ratio is the target of the optimization
process. Hence the mass flow will be different for every geometry. However it is expected that the
performance at the low Mach number flow with fixed inlet angle will not noticeable be influenced by a
modest change in mass flow.
Table 2 Diffuser inlet conditions at the 3 operating points

surge design choke

Α (flow angle) 62.5o 52.8 o 37.5 o
Mass flow 162.8 210.0 267.5

The blade geometry is defined by a NACA thickness distribution superposed on a camber line defined
by a 4 parameter Bezier curve The 5th design parameter is a scale factor for the NACA thickness
distribution (between .7 and 1.3). The 6th parameter is the number of blades (between 6 and 21). The last
design parameter is the blade height. It is constant from leading edge to training edge but can be different
for the prescribed impeller outlet width.
The main performance parameters are the static pressure rise- and total pressure loss coefficient
P3 − P2 P o 2 − P o3
Cp = ω=
P o 2 − P2 P o 2 − P2
They are calculated from the 3D Navier Stokes results obtained by means of the TRAF3D solver on a
grid with 400 000. cells.
Making the Database is quite costly because it requires analyzing every geometry at three operating
points. The initial Database is therefore limited to only 10 randomly generated samples requiring 30
Navier Stokes calculations.
The outlet over inlet pressure ratio defines the outlet boundary condition and cannot be used as an
objective for the optimization. Losses only are also not sufficient to measure the performance because
minimum losses may be obtained without any pressure rise. One wants to achieve the maximum pressure
rise with minimum loss to keep the maximum kinetic energy available for the downstream components.
Hence one must also consider the losses that are generated to reach that pressure rise.
The optimizer therefore aims to maximize following OF
OF = (1 − ( wlow .Cplow + wmean .Cpmean + whigh .Cphigh )) + wlow .ωlow + wmean .ωmean + whigh .ωhigh
This corresponds to a maximization of Cp while minimizing the losses. Taking into account that
Cp + ω = Cpisentropic
it is clear that minimizing the losses helps to reach the isentropic Cp. The latter is geometry dependent.
The outcome of the optimization is illustrated on Fig. 24. All diffusers show low losses with an
increase of Cp at all operating points up to Cp+ω=.74. Higher values of the latter give rise to increasing
losses and a decrease of Cp at low mass flow. Iteration 17 is considered as the optimum because of only a
small decrease of Cp at minimum mass flow.
The Mach number distribution around the optimized vanes is shown on Fig. 25. One observes a small
flow separation at minimum mass flow corresponding to the increase of the losses shown on Fig. 24.

RTO-EN-AVT-167 3 - 19
Global Optimization Methods: Theoretical Aspects & Definitions

Fig. 24 Performance criteria of LSD

low m& max m&

med m&
Fig. 25 Mach number distribution in the optimized geometry at the three operating points

3 - 20 RTO-EN-AVT-167
 Global Optimization Methods: Theoretical Aspects & Definitions

6.2 Multipoint optimization of a radial compressor impeller

The mass flow in compressors is an outcome of a compressible flow calculation whereby the pressure
ratio is imposed. Verifying if the target surge and choking mass flows are reached therefore requires
knowing the corresponding pressure ratio before the calculation is made. The procedure developed at the
VKI therefore calculates the flow at three predefined pressure ratios. A low one to find the choking mass
flow, one corresponding to what is estimated the pressure ratio at design point and one at higher pressure
ratio. This allows drawing a performance curve (indicated in bleu on Fig. 26) which is unlikely to satisfy
the required choking mass flow. A simple scaling of the inlet section allows defining a new geometry that
does satisfy this requirement. This first performance curve also allows a better definition of the design
point pressure ratio. A last information is the pressure/mass flow slope allowing a guess of the pressure
ratio at the required surge point mass flow. This is verified by an analysis of the scaled geometry and the
resulting performance curve (red one) allows verifying if the targets are reached: i.e. choking mass flow,
best efficiency at design mass flow and minimum pressure mass flow slope. A minimum of six flow
analyses are required for each optimized geometry proposed by the GA.

Fig. 26 Procedure for multipoint compressor optimization

7. MULTIOBJECTIVE OPTIMIZATION

7.1 Pseudo Objective Function versus Pareto front

Cooling HP turbine blades allows increasing the thermal efficiency by operating at higher Turbine
Inlet Temperature (TIT) while maintaining the lifetime of the turbine. However the cooling air does not
contribute to the work output and should be minimized because of its negative impact on thermal
efficiency. Hence minimizing coolant flow and increasing lifetime are conflicting objectives. One way to
satisfy both requirements, already explained in section 4.3, is by defining a pseudo OF that increases with
insufficient lifetime and the amount of coolant mass flow.
OF (G ) = wl .Pl (G ) + wm .Pm (G )
The balance between the different objectives is defined by the respective weight factors. The task of the
optimization algorithm thereby consists in finding the geometry that minimizes this pseudo OF.
However the balance between the different objectives may not be clear from the beginning. The
different OF can then be plotted in the fitness space allowing a tradeoff between the two goals (Fig. 27).
The non dominant solutions define a Pareto front. i.e. the collection of the geometries G for which one

RTO-EN-AVT-167 3 - 21
Global Optimization Methods: Theoretical Aspects & Definitions

objective cannot be decreased without increasing the other one. The choice is then left to the designer to
select at the end of the optimization one geometry out of the non-dominated ones, that has the right
balance between both objectives.

Fig. 27 Definition of Pareto front

The relation between the pseudo OF and the Pareto front approach is illustrated on Fig. 28.
Optimization driven by a pseudo OF will follow a path in the direction of the red arrow on the figure; i.e.
perpendicular to the slope defined by OF (G ) = wl .Pl (G ) + wm .Pm (G ) . Convergence is reached when the
line of constant OF is parallel to the Pareto front. The main advantage of using a pseudo OF is that fewer
geometries need to be analyzed to find this optimum. The disadvantage is that the pseudo OF approach
requires a rather good idea of the relative weights to be given to both penalties. The choice of the relative
weights is rather obvious when one objective must be satisfied without compromise. This was the case
when optimizing the radial compressor, shown in section 4, i.e. where the stress penalty had to be satisfied
at all cost.
The advantages and disadvantages of both approaches are illustrated by the optimization of the
cooling system of a HP turbine blade

Fig. 28 Pseudo Objective Function versus Pareto front

3 - 22 RTO-EN-AVT-167
 Global Optimization Methods: Theoretical Aspects & Definitions

7.2 Optimization of a cooled turbine blade

The optimization of an internal cooling system for an HP turbine aims for a reduction of the coolant
flow while assuring the required lifetime. The mechanism is by uniformizing the metal temperature and in
particular by limiting the metal temperatures in areas where the stresses are high [11].
Geometry parameterization
The cooling system to be optimized consists of five straight cylindrical channels. Design parameters
are the position of their center at hub and shroud and their diameter. The centers are defined by the local
curvilinear (η, ε) coordinates. The η ∈ [0,1] coordinate represents the length along the camber line, while ε
∈ [−1,1] defines the position perpendicular to the camber line (Fig. 29). The maximum value of ε (±1)
corresponds to half the blade thickness at each η location. This facilitates the definition of a valid set of
design parameters, i.e. for which the cooling channels do not intersect the blade wall. The local
coordinates are the same at hub and shroud in order to reduce the number of design parameters.

Fig. 29 Parameterization of the location and diameter of the cooling channel

Five individual cooling channels, parameterized by three numbers (η, ε and D), result in 15 design
variables. Although the choice of relative coordinates reduces the number of invalid geometries, cooling
channels can still be too close to the blade surface. Table 3 shows the individual range for all parameters
of each cooling channel to avoid it.

Table 3 cooling hole parameter range

Lifetime Prediction
The prediction of the lifetime requires the calculation of the stresses and metal temperature inside
the blade. The heat transfer from the external flow in the solid blade and cooling channel is calculated by a
coupled method (CHT) [12]. It is a combination of a 3D NS solver for the external flow, a FEA method

RTO-EN-AVT-167 3 - 23
Global Optimization Methods: Theoretical Aspects & Definitions

for the internal heat transfer and stress calculation and a 1D non-adiabatic flow model for the cooling
channel. Hence the method requires three grids (Fig. 30). Solid and fluid flow solvers alternate with an
exchange of boundary conditions until the continuity of temperature and heat flux is obtained at their
interfaces. The drawback of this approach is the need for sequential iterations between the two platforms
and an interpolation of the boundary conditions from one grid to the other. The main advantage of the
coupled approach is that one can make use of standard grid generators, NS and FEA solvers. Those codes
have been extensively validated and their limitations and capabilities are well known. A FEA calculation
is anyway needed to calculate the stresses.

Fig. 30 Superposition of grids used in the coupled calculation

Upon completion of the heat transfer analysis, the solid temperature is known at each node of the
FEA grid which allows a straight forward calculation of the thermal stress. The stresses due to the
centrifugal forces and blade bending, resulting from the pressure difference between pressure and suction
side, can be computed on the same grid with a temperature dependent material model. Assuming linearity,
the total stresses are the sum of thermal, centrifugal and pressure stresses. Knowing the temperature and
stress in each node of the FEA grid allows a computation of the blade lifetime.
The lifetime of the blade is assumed to be proportional to the creep-to-rupture failure according to
Hill’s anisotropic equation [13]. The equivalent stress γ ranges from 0 to 1. The material is considered to
fail when γ= 1.
γ = H (σ xx − σ yy ) 2 + F (σ yy − σ zz ) 2 + G (σ zz − σ xx ) 2 + 2 N (τ xy ) 2 + 2 L(τ yz ) 2 + 2M (τ zx ) 2
H, F, G, N, L and M are material properties which depend on the Larson-Miller parameters in the
longitudinal and transverse directions. They are specified by (Fig. 31)
LMP = T .[log10 (l ) + 23]
The parameters for the directionally solidified nickel superalloy material GTD-111 are used for the present
analysis.

3 - 24 RTO-EN-AVT-167
 Global Optimization Methods: Theoretical Aspects & Definitions

Fig. 31 Larson-Miller plot for longitudinal directional solidified GTD-111

The algorithm starts with a first guess of l and computes the Larson-Miller parameter in each node
of the solid grid. These parameters are input for the calculation of the material properties (H, F, G, N, L
and M). The material failure is checked in each node. Depending on the result, a new estimation of l is
made, i.e. l is lowered if the maximum value of γ is larger than 1, or vice versa. This computation is
repeated until a value of l is found for which the largest γ equals 1, plus or minus a tolerance.
Performance
The optimization method is the extension of the aerodynamic optimization tool for axial and radial
impellers and stators and schematically shown on Fig. 12. The performance of the geometry is assessed
after each CHT and thermal stress analysis or after the prediction by the metamodel. In order to achieve
the target lifetime with the prescribed coolant mass flow, a penalty related to an insufficient lifetime and a
too large coolant mass flow is given to each design. The task of the optimization algorithm thereby
consists in finding the parameters that minimize the pseudo OF defined by
OF (G ) = wl .Pl (G ) + wm .Pm (G )
The penalty on the lifetime is proportional to the difference between the calculated l (G ) and target ltar
lifetime. The latter is set to 20000h, which is very high considering the high TIT(1400K) and the absence
of a thermal barrier coating. This penalty is zero only if the target lifetime ( ltar ) is exceeded (Fig. 32).
Pl (G ) = max(ltar − l (G ),0.)

Fig 32 Lifetime and cooling mass flow penalties

RTO-EN-AVT-167 3 - 25
Global Optimization Methods: Theoretical Aspects & Definitions

A higher coolant mass flow results in a lower overall cycle efficiency and the penalty increases when the
& offset = 3gr / s , as shown on Fig. 32.
mass flow is higher than an offset value m
Pm (G ) = max(m& (G ) − m& offset ,0)
Considering that the lifetime penalty is more important than the mass flow penalty, the weight given to the
latter is 400. This results in a penalty of 133.33 for each g/s exceeding the limit of 3g/s. The weight given
to the lifetime is 1/h.
Results
The starting geometry is a blade with 5 equidistant cooling holes. The surface temperature, von
Mises and effective stresses are shown on Fig. 33. The lifetime based on thermal stresses of the cooled
blade (220h) is hardly longer than the one of the uncooled blade (161h). The reason is that the solid
temperature has decreased over most of the blade but not at the trailing edge hub where the largest stresses
occur.

Fig. 33 Non optimized cooled blade (equidistant holes)

Two distinct optimizations are performed starting from the same initial database. One uses the ANN
while the other one uses the RBF to predict the same quantities. Both optimizations are run for 30
iterations, after which a synchronization of the databases is made i.e. all existing samples are put together
in one unified database. An additional 20 optimization iterations are then performed restarting with this
extended database. The purpose of this synchronization is to exchange information between both
optimizers and see if they can profit from it.
Fig. 34 shows, for both optimizations, the evolution per iteration of the mass flow, lifetime and total
penalty. The metamodel predictions are compared with the results of the CHT and lifetime calculation.
The mass flow in the individual cooling channels is immediately well predicted by the ANN whereas the
method is too optimistic in terms of lifetime. The lifetime is better predicted by the RBF with a monotonic
convergence towards the optimum except for 2 stepwise increases in lifetime. They correspond to a new
choice by the optimization algorithm, based on the newly acquired information. It takes longer to
accurately predicted the massflow by the RBF.
An increased discrepancy is observed after the synchronization of the databases at iteration 30. This
is probably due to the larger number of training samples, requiring more hidden neurons, respectively
RBF-centers to remain accurate. The best geometry proposed by the ANN is geometry 18. It has a total
penalty of 14 058 corresponding to a lifetime of 6872h and a 10g/s cooling mass flow. The best RBF
geometry is found at iteration 47 with a lifetime of 6758h for 10.4g/s cooling flow.

3 - 26 RTO-EN-AVT-167
 Global Optimization Methods: Theoretical Aspects & Definitions

Fig. 34 Convergence with ANN and RBF

Fig. 35 Optimized cooling geometry

The optimal solutions found by both optimizations are compared on Fig. 35. A tendency towards
small non-turbulated cooling channels, near the pressure side of the blade is observed. The trailing edge
cooling channel should be positioned as far downstream as possible (h = 0.95). Hole number four has the
largest diameter (4.4mm) with turbulators.

RTO-EN-AVT-167 3 - 27
Global Optimization Methods: Theoretical Aspects & Definitions

Fig. 36 Convergence to the optimum with pseudo OF

Figure 36 shows the mass flow penalty versus the lifetime penalty for the initial database samples and
both optimizations. It is clear that all geometries created by the optimization system outperform the initial
database ones. Whereas the maximum lifetime of the samples contained in the initial database does not
exceed 1685h with an average lifetime of only 300h, most optimized geometries have a lifetime above
5000h. This illustrates the capability of the optimization system to rapidly improve the performance
starting from only a limited information. The envelope of all the results form a Pareto front. It is a very
incomplete one because the optimizer has been targeted towards a particular combination of the two
penalty functions and a large part of the design space has not been investigated.

Fig. 37 Effective strength distribution at rupture

The effective stresses at failure on the surface of the uncooled (161h) , cooled (220h) and optimized
(6872h) blade surface are shown on Fig. 37. The first two show no change in the critical area. The large
improvement of the optimized one is the consequence of a more uniform stress distribution over the whole
blade. The critical area has shifted from the hub trailing edge towards the blade tip. All parts of the blade
are aging in almost the same

3 - 28 RTO-EN-AVT-167
 Global Optimization Methods: Theoretical Aspects & Definitions

Fig. 38 Cooling channel 4 diameter versus lifetime and mass flow

Figure 38 shows the influence of the diameter of the 4th channel on lifetime and mass flow. A sudden
variation in the mass flow is observed at 3mm due to the activation of the turbulators. The longest lifetime
is obtained with a turbulated cooling channel of 4.4mm diameter. One also observes that the result is
rather robust i.e. the lifetime does not change very much for small changes in the hole diameter.

7.3 Self Organizing Maps

The Pareto front is quite useful for problems with 2 objectives as long as it remains convex. A
visualization becomes more complicated when 3 objectives are specified (Fig. 39).

Fig. 39 Multi-objective Pareto front

RTO-EN-AVT-167 3 - 29
Global Optimization Methods: Theoretical Aspects & Definitions

Self Organizing Maps (SOM) [14] allow a balanced evaluation of the different geometries when more than
three objectives are specified. The high dimensional maps are mapped into lower dimensional spaces by
grouping geometries with similar characteristics.

Fig. 40 Data mapping in organized maps

This is illustrated by the results of an airplane wing optimization where minimum drag at transonic
and supersonic speed, bending and pitching moments are the four objectives to be reached. The different
types of wing sections are grouped on Fig. 41. The corresponding values of the four objectives are shown
by color on Fig. 42.

Fig. 41 Map of similar geometries

3 - 30 RTO-EN-AVT-167
 Global Optimization Methods: Theoretical Aspects & Definitions

Fig. 42 Self organized maps

AKNOWLEDGMENTS
The contributions of Dr. S Pierret, Dr. T. Verstraete and Dr. Z Alsalihi are gratefully
acknowledged.

BIBLIOGRAPHY
[1] Peter J., 2010. “Local optimization”, in: Strategies for Optimization and Automated Design of Gas
Turbine Engines, RTO-MP-AVT-176, paper

[2] Aarts E.H.L. and Korst J.H.M., 1987, ``Simulated annealing in Boltzmann machines'', Wiley
Chichester.

[3] Bäck Th., 1996, ``Evolutionary Algorithms in Theory and Practice'', Oxford University Press,
N.Y.

[4] Rechenberg, I., 1973, “Evolutionsstrategie: Optimierung technischer Systeme nach Prinzipien der
biologischen Evolution”, Stuttgart: Fromman-Holzboog.

RTO-EN-AVT-167 3 - 31
Global Optimization Methods: Theoretical Aspects & Definitions

[5] Van den Braembussche R.A., 2010. “Tuning of Optimization Strategies”, in: Strategies for
Optimization and Automated Design of Gas Turbine Engines, RTO-MP-AVT-176, paper .

[6] Pierret S. and Van den Braembussche R.A., 1999, ``Turbomachinery Blading Design Using
Navier Stokes Solver and Artificial Neural Network'', ASME Journal of Turbomachinery, Vol. 121 pp.
326-332.

[7] Vanderplaats G.N., 1984, ``Numerical Optimization Techniques for Engineering Design'',
McGraw-Hill

[8] Arnone A., 1994, ``Viscous Analysis of Three-dimensional Rotor Flow Using a Multigrid
Method'', ASME Journal of Turbomachinery, Vol. 116, pp. 435-445.

[9] Verstraete T., Alsalihi Z. and Van den Braembussche R.A., 2007, ``Multidisciplinary
Optimization of a Radial Compressor for Micro Gas Turbine Applications'', ASME GT2007-27484.

[10] SAMCEF FEA code by: SAMTECH group, www.samcef.com .

[11] Verstraete, T., Amaral, S., Van den Braembussche, R. A., and Arts, T., 2008. “Design and
Optimization of the Internal Cooling Channels of a HP Turbine Blade — Part II, Optimization”. In
ASME. Paper No. GT2008-51080.

[12] Amaral, S., Verstraete, T., Van den Braembussche, R. A., and Arts, T., 2008. “Design and
Optimization of the Internal Cooling Channels of a HP Turbine Blade — Part I, Methodology”. In ASME.
Paper No. GT2008-51077.

[13] Dowling, N., 2007. Mechanical Behaviors of Materials. Prentice Hall.

[14] Sasaki S. and Obayashi S., 2004. “Adaptive Range Multi-objective Genetic Algorithm and Self
Organizing Map for Multi-objective Optimization Problems”, in Optimization Methods for Multi-
criteria/multidisciplinary Design Applications to Aeronautics and Turbomachinery, VKI-LS-2004-7

3 - 32 RTO-EN-AVT-167