IOP Conference Series: Materials

Science and Engineering

OPEN ACCESS You may also like

- Assessment of groundwater well
Challenges and progress in turbomachinery vulnerability to contamination through
physics-informed machine learning
design systems Mario A Soriano, Helen G Siegel,
Nicholaus P Johnson et al.

- Probing the Small-scale Structure in

To cite this article: R A Van den Braembussche 2013 IOP Conf. Ser.: Mater. Sci. Eng. 52 012001 Strongly Lensed Systems via
Transdimensional Inference
Tansu Daylan, Francis-Yan Cyr-Racine,
Ana Diaz Rivero et al.

- Multilevel design optimization of hydraulic

View the article online for updates and enhancements. turbines based on hierarchical metamodel-
assisted evolutionary algorithms
E Kontoleontos, M Zormpa, S Nichtawitz
et al.

This content was downloaded from IP address 51.159.15.249 on 21/04/2024 at 19:44

6th International Conference on Pumps and Fans with Compressors and Wind Turbines IOP Publishing
IOP Conf. Series: Materials Science and Engineering 52 (2013) 012001 doi:10.1088/1757-899X/52/1/012001

Challenges and progress in turbomachinery design systems

R A Van den Braembussche

Von Karman Institute for Fluid Dynamics, Waterloose Steenweg, 72, B-1640 Sint-
Genesius-Rode, Belgium

E-mail: [email protected]

Abstract. This paper first describes the requirements that a modern design system should meet,
followed by a comparison between design systems based on inverse design or
optimization techniques. The second part of the paper presents the way these challenges
are realized in an optimization method combining an Evolutionary theory and a Metamodel.
Extensions to multi-disciplinary, multi-point and multi-objective optimization are illustrated by
examples.

1. Design System Requirements

Navier Stokes solvers and Finite Element Stress Analysis are now routinely used to predict the
performance and verify the mechanical integrity of new geometries. However they do not specify what
modifications are needed to improve the performance or to minimize weight while keeping stress and
vibrations below some limit values. Although they provide very detailed information, the designer will
often base his decisions on overall parameters such as efficiency, pressure ratio and mass flow, leaving
huge amounts of information unexploited. Present paper discusses some design systems for
turbomachinery applications that have been developed over the years in order to assist the designer in
finding the optimal geometry by making better use of the available information.
Further progress in performance requires incorporating all 3D designs features, such as lean and
sweep, that may help to improve the performance or to reach other design targets. Limitations or
simplifications of the geometry are acceptable only if they are needed to satisfy other design
requirements such as manufacturing or in service cost. Design systems for advanced turbomachinery
should define the final geometry. Any post design geometry modification may result in a suboptimal
geometry.
The quality of a design depends on the accuracy of the analysis methods that have been used.
Approximate solvers or surrogate models can only used for a first approximation in order to speed up
the design procedures. The use of accurate solvers is mandatory to verify the final geometry because
any inaccuracy of the flow solver could drive the design system towards a suboptimal geometry.
The outcome of a design should not only be optimum in terms of aerodynamic/hydraulic
performance but should also respect all other objectives such as cost and manufacturing limitations
while assuring a safe operation over the preset lifetime. This requires a multi-disciplinary approach
(fluids, stress vibration, economics etc) and a delicate balance between the different sometimes
contradicting targets.
Guaranteeing a stable operation over a sufficiently large operating range requires a multi-point
approach. Designing for maximum efficiency or large operating range may result in different

Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution
of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
Published under licence by IOP Publishing Ltd 1
6th International Conference on Pumps and Fans with Compressors and Wind Turbines IOP Publishing
IOP Conf. Series: Materials Science and Engineering 52 (2013) 012001 doi:10.1088/1757-899X/52/1/012001

geometries or impact on the manufacturing cost. Multi-objective design systems should be able to
find a compromise between these sometimes conflicting objectives.
Computerized designs have the tendency to have peak performance but to be very sensitive to
geometrical imperfections. All design parameters are stressed to their limit and even small variations
may result in a rapid deterioration of performance. Robust design systems provide geometries that are
less sensitive to geometrical imperfections and to inherent inaccuracies of the evaluation programs.
An affordable computer effort is a prerequisite for a design system to be economically acceptable.
Combining all previous requirements in a design system is the major challenge for the developer.
Following discusses how advanced design systems respond to these challenges

2. Inverse design versus Optimization

Inverse design systems define the geometry that corresponds to a pre-defined Mach number or
pressure distribution. This requires a very good insight in fluid dynamics of the designer to find out
how an optimum distribution should look like. This is more or less understood for 2D flows and has
resulted in controlled diffusion blades. The main problem is to find out how such an optimum
distribution is influenced by secondary and tip leakage flows and how to guarantee the geometrical
constraints such as minimum or maximum blade thickness (existency problem). The geometry is the
outcome of the inverse design so that the mechanical constraints can be verified at the end of the
procedure. The latter is often avoided by specifying a thickness- and loading distribution whereby the
velocity is then a consequence of the meridional contour [1]. Hence there is no direct control of the
local velocity deceleration which according to Lieblein is a major factor influencing the losses.
Inverse methods are often based on simplified (inviscid) flow equations eventually corrected for
boundary layer blockage and neglecting secondary flows. The outcome will be different from what
would be obtained by solving the real flow equations and the performance cannot be guaranteed. In
what follows one will limit our self to optimization systems that are based on the accurate analysis
methods commonly used now in industry.
Optimization systems find the geometry that best satisfies the design objectives (OF) expressed in
terms of performance, cost etc. while respecting the constraints (max. stress, lifetime). This is
illustrated on figure 1 showing the iso-loss contours in function of the two design parameters X1 and
X2.
Objective is to find the combination of the
design parameters X1 and X2 that result in a
minimum loss coefficient ω while respecting
the constraints. Most systems make use of an
X2,max iterative procedure and start from an existing
X0 geometry (X0). Simple mechanical and
X X X
geometrical constraints (X1<X1max and X2<
X 2max) can already be verified before any time
X1
consuming flow analysis is started. First order
X X X3 X methods find the optimum geometry by
following the direction of steepest descend.
X2 This requires the calculation of the derivatives
X X X  / X 1 and  / X 2 , and the calculation of
X1,max the optimum step length to reach the point X1.
Figure 1. Optimization with two New steps are calculazted until the optimum
parameters geometry is found (zero gradient) or the path is

blocked by a constraint.
Most optimization systems make use of existing and well proven solvers to predict the OF of
different geometries so that the outcome is very trustworthy. The large number of analyses that are
needed to calculate the gradients and the large number of steps that may be required to reach the

2
6th International Conference on Pumps and Fans with Compressors and Wind Turbines IOP Publishing
IOP Conf. Series: Materials Science and Engineering 52 (2013) 012001 doi:10.1088/1757-899X/52/1/012001

optimum, results in a computational effort that may be prohibitive for most real cases. A first
challenge is to reduce this effort by reducing the number of required iterations and/or reducing the
computational effort for each iteration.
Adjoint methods allow calculating the steepest gradient with a computational burden that is
comparable to the one of an analysis. This requires a modification of the flow solver, excluding the
use of “off the shelf” solvers. It also complicates the extension to multisdisciplinary designs.
An alternative are the zero-order or stochastic search mechanisms requiring only OF evaluations.
The systematic exploration of the design space, indicated by “X” on figure 1, requires only nine OF
evaluations to obtain a rather good idea of the optimum geometry. However the number of evaluations
increases exponentially with the number of design parameters, leading to prohibitive computer efforts
for more complex geometries. Zero order methods have fewer chances to get stuck in a local minimum.
Evolutionary strategies such as Genetic Algorithms (GA), Simulated Annealing (SA), Kriging and
many others 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.
A way to reduce the computational burden is by working on different levels of sophistication,
combining approximate but fast prediction methods with accurate but time consuming ones [2,3].
Such a system is illustrated on figure 2. The
fast but less accurate optimization loop is to the
right; the expensive but accurate one is to the left.
The OF driving 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 OF to the geometry
similar to what is done by the accurate analyzers.
Surrogate models have the same input and output
as the analysis methods 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
Figure 2. Flowchart of optimization system the accurate solvers. The optimized geometry is
then verified by the accurate one. The procedure
is stopped when the accurate solver confirms that the surrogate model makes accurate predictions i.e.
confirms that the optimizer was driven by accurate predictions. Otherwise a new GA optimization is
started after a new learning of the metafunction considering also the new optimized geometries.
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 possible approximations and errors when transferring
the geometry from one discipline to another.
 The possibility to shorten the design time by making all expensive analyses in parallel
 The existence of a global OF accounting for all disciplines. This allows a concurrent optimization
driving the geometry to a compromise between all requirements without iterations between the
aerodynamically optimum geometry and the mechanically acceptable one.

3. Multidisciplinary Optimization of a Radial Impeller

Previous approach can easily be extended to multidisciplinary optimization by calculating the different
contributions to the OF (performance, heat transfer, stress, etc.) in parallel. The method 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 [3] The objective is a maximum efficiency while respecting the stress
limits.

3
6th International Conference on Pumps and Fans with Compressors and Wind Turbines IOP Publishing
IOP Conf. Series: Materials Science and Engineering 52 (2013) 012001 doi:10.1088/1757-899X/52/1/012001

The first step in an optimization is the parameterization of the geometry. This is a very important
issue as it should be sufficiently general, not to exclude the optimal geometry, without increasing the
number of design parameters beyond a limit where it starts to slow down or prevent convergence. The
3D radial impeller is defined by the meridional contour (figure 3), in combination with the blade
camber line at hub and shroud (figure 4).

Figure 3. Meridional contour defined by Figure 4. Definition of the blade camber line
Bézier control points by β angle.

The meridional contours are defined by fourth order Bezier curves. Design variables are the six
coordinates of the control points that can be varied and indicated by arrows on figure 3. The
camberlines are defined by third order polynomials specifying the distribution of the angle β(m)
between the meridional plane and the camberline (figure 4).
 u    0 1  u 3  31u 1  u 2  3 2 u 2 1  u 2   3 u 3
u is the non-dimensional meridional length. β0 and β3 are the blade angles at leading- and trailing edge.
The splitter blades are a short version of the full blades, with the leading edge cut back. The design is
completed by a prescribed thickness distribution normal to the camber line and the number of blades.
The latter could also be a design parameter to be optimized, but has been fixed to 7 for manufacturing
reasons. A parameterization with more degrees of freedom is described in [4].
The OF to be minimized is a weighted sum (weight factor w) of three penalties:

 

 

 

 
 

OF G  wstress  Pstress G  w  P G  wmassflow  Pmassflow G  wMach  PMach G

The first one concerns the mechanical stresses and starts increasing when the maximum von Mises
stress in the impeller σmax exceeds a prefixed value σallowable.
   allowable 
Pstress  max  max ,0.0
  allowable 
Expressing this constraint as a penalty does not guarantee that it is respected but has the advantage that
all geometries that have been analyzed provide information that leads the GA towards the optimum
geometry.
The mass flow penalty increases when the error exceeds .3% of the required mass flow ( m  req ) and
with the difference in mass flow between the blade channels on both sides of the splitter blade. The
latter favours an equal blade loading between splitter and full blade and improves the periodicity of the
impeller exit flow

4
6th International Conference on Pumps and Fans with Compressors and Wind Turbines IOP Publishing
IOP Conf. Series: Materials Science and Engineering 52 (2013) 012001 doi:10.1088/1757-899X/52/1/012001

2
  m req  m  2
 1    m upper  m lower 
Pmassflow   max   ,0 . 0     
  m 300     m upper  m lower 
  req
   

The penalty on the Mach number penalizes non-optimal loading distributions. The first part
increases with negative loading. The second part increases with the loading unbalance between main
blade and splitter blade. It 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.
2
1  Abl  Asp 
PMach 
  max M ps s   M ss s ,0.0  ds    
 Abl  Asp 
0  

Figure 5. Negative loading and loading Figure 6. Aero - versus stress penalty for
unbalance in a compressor with splitter vanes. baseline, database- and optimization geometries.

Figure 7. von Mises stresses due to centrifugal loading in the baseline (left) and optimized (right)
impeller.

5
6th International Conference on Pumps and Fans with Compressors and Wind Turbines IOP Publishing
IOP Conf. Series: Materials Science and Engineering 52 (2013) 012001 doi:10.1088/1757-899X/52/1/012001

The optimization starts from the outcome of a simple aerodynamic optimization without stress
computation, called “Baseline” impeller. Although this geometry has a good efficiency, it cannot be
used because a FEA stress analysis predicts von Mises stresses in excess of 750 MPa.
The initial database contains 64 geometries selected by the DOE technique. About 40
optimization cycles are needed to obtain a good agreement between ANN predictions and the NS and
FEA analyses. It means that the optimization is achieved with only 100 NS and FEA analyses.
The aero penalty is plotted versus the stress penalty in figure 6. The geometries created during
the optimization cycles are all in the region of low penalties. Most of them outperform the geometries
of the Database. From all geometries at zero stress penalty, the one with minimum aero penalty is the
optimum. The reduction of the maximum stress level with 370 MPa is at the cost of a 2.3 % decrease
of efficiency.
Figure 7 compares the von Mises stresses in the baseline geometry with the optimized one. 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

4. Robust Design
Robustness expresses the in-sensitivity of the design to manufacturing noise, variations in the
operating conditions and inaccuracies of the OF calculations. It is normally verified a posterior by
perturbing the geometrical and operating parameters. Doing this by means of the accurate
metafunction, available at the end of the design, allows very large time savings.
Analysing the Database of the iterative methods, as the one presented in section 3, provides also
valuable information about the robustness. This is illustrated on figure 8 showing the variation of the
stress and efficiency in function of the leading edge lean. The latter is defined as the angle between the
blade leading edge and the meridional plane (positive in the direction of rotation). Designers
intuitively try to keep it zero in order to minimize bending stresses at the hub. Figure 8 however shows
that the lowest stresses occur at -12.0º. This unexpected result is a major outcome of the optimization.
The small variations of the stress and efficiency in function of lean further illustrate the robustness of
this design.

Figure 8. Blade lean versus stress and efficiency for database and optimization geometries.

5. Multi-point optimization of a transonic radial impeller

Multipoint designs aim to achieve sufficient range between choking and surge, to have maximum
efficiency at design mass flow and a sufficiently negative slope of the pressure rise versus mass flow
curve for stability reasons (figure 9). Defining the whole operating line requires a large number of
expensive NS calculations that are of little use if the choking mass flow is different from the required
value.
The optimization algorithm has therefore been adapted by adding a geometry scaling to the
optimizer which allows recuperating the optimized geometries that do not satisfy the choking
requirements [4]. Geometries that show a potential in terms of operating range and performance are

6
6th International Conference on Pumps and Fans with Compressors and Wind Turbines IOP Publishing
IOP Conf. Series: Materials Science and Engineering 52 (2013) 012001 doi:10.1088/1757-899X/52/1/012001

scaled to satisfy the choking requirements. The scaled geometry is then input to a final series of NS
calculations and one FEA stress analysis. Using the results of the first series of NS analyses of the
unscaled impeller one can define the boundary conditions corresponding to a uniform distribution of
operating points between surge and choke. The use of databases and metafunctions dedicated to
predict surge and choking further allows speeding up the convergence because fewer expensive
accurate calculations will be required.
The outcome of a transonic radial impeller optimization by this method is shown on figure 10. The
final geometry satisfies the range requirements and the efficiency is increased by more than 2 points.
The latter is due to the increased degree of freedom allowing splitter vanes that are different from the
main blades. All mechanical constraints in terms of maximum stresses are respected.

Figure 9. Procedure for multipoint compressor

optimization

Figure 10. Performance of the redesigned versus

baseline impeller

6. Multi-objective optimization
The outcome of an optimization with 2 objectives can be visualized by plotting them in the 2D fitness
space. The non dominated solutions i.e. the geometries G for which one objective cannot be decreased
without increasing the other one, define a Pareto front. The choice is then left to the designer to select
at the end of the optimization the geometry on the Pareto front that has the right balance between both
objectives. A Pareto front is quite useful for problems with 2 OF as long as it remains convex.
Visualization becomes more complicated when more than 2 OF are specified and special techniques
are required to come to a motivated decision [5].
Defining the Pareto front is a time consuming activity requiring a large number of geometry
analyses of which many will be of no interest. An alternative is a combination of the penalties
corresponding to the different objectives into one pseudo-OF. Much less geometry analyses may be
needed to reach the optimum.
OF (G )  w1.P1 (G )  w2 .P2 (G )
This approach was already used in previous sections and is illustrated here by the optimization of
the cooling system of a HP turbine blade [6]. The optimization aims to lower P1 (increasing with the
amount of cooling mass flow) and P2 (increasing when the required life time is not reached). The
lifetime depends of the equivalent stress, function of the material parameters, stress and temperature
in each point of the blade. [7]. During the optimization process driven by a pseudo OF, the
optimization follows a path in the design space towards the point where the lines of constant pseudo
OF become tangent to the Pareto front.

7
6th International Conference on Pumps and Fans with Compressors and Wind Turbines IOP Publishing
IOP Conf. Series: Materials Science and Engineering 52 (2013) 012001 doi:10.1088/1757-899X/52/1/012001

The main advantage of this approach is that only

30 geometry analyses are needed 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.
Increasing the weight on P1 emphasizes on
minimum cooling mass flow and hence cycle
efficiency. Increasing the weight on P2
emphasizes on lifetime. The choice of the relative
weights is rather obvious when one objective
must be satisfied without compromise. This was
the case in section 4 when optimizing the radial
compressor, because the stress penalty has to be
satisfied at all cost. The balance between the
different penalties may be less clear in other
cases. However perturbing the design parameters
Figure 11. The evolution of P1 and P2 t
around the optimum geometry provides
towards the Pareto front when using a pseudo OF.
information about the interesting part of the
Pareto front with minimum extra effort.

7. Conclusions
Turbomachinery design systems based on optimization have seen a large development in recent years.
The method presented here allows an important gain in design time and performance while respecting
the requirements. Extensions to multipoint and multi-objective are straightforward and result in
advanced and realistic geometries

References
[1] Watanabe H and Zangeneh M 2003 Design of the blade geometry of swept transonic fans by 3D
inverse design ASME-GT 38770.
[2] Pierret S and Van den Braembussche R A 1999 ASME Journal of Turbomachinery 121 326-332.
[3] Verstraete T, Alsalihi Z and Van den Braembussche R A 2007 ASME Journal of
Turbomachinery 132 03104.
[4] Van den Braembussche R A, Alsalihi Z, Verstraete T, Matsuo A, Ibaraki S, Sugimoto K and
Tomita I 2012 Multidisciplinary Multipoint Optimization of a Transonic Turbocharger
Compressor ASME-GT 695645.
[5] Sugimura K, Jeong S, Obayashi S and Kimura T 2008 Multi-Objective Robust Design
Optimization and Knowledge Mining of a Centrifugal Fan that takes Dimensional
Uncertainty into Account ASME-GT 51301.
[6] Verstraete T, Amaral S, Van den Braembussche R A and Arts T 2008 ASME Journal of
Turbomachinery 132 021014.
[7] Verstraete T, Amaral S, Van den Braembussche R A and Arts T 2008 ASME Journal of
Turbomachinery 132 021013.