You are on page 1of 21

European Congress on Computational Methods in Applied Sciences and Engineering

ECCOMAS 2004
P. Neittaanmäki, T. Rossi, S. Korotov, E. Oñate, J. Périaux, and D. Knörzer (eds.)
Jyväskylä, 24–28 July 2004

DESIGN OF GAS TURBINE ENGINES USING CFD

Leigh Lapworth? and Shahrokh Shahpar?


? Aerothermal Methods Group
Rolls-Royce plc, P.O.Box 31, Derby, DE24 8BJ, England
e-mails: leigh.lapworth@rolls-royce.com, shahrokh.shahpar@rolls-royce.com
web page: http://www.rolls-royce.com

Key words: Turbomachinery, Design, Optimisation, Adjoint.

Abstract. This paper describes a general purpose design system being developed at Rolls-
Royce plc. The key elements of the system are a parametric design and rapid meshing
capability; a state-of-the-art CFD solver with an adjoint capability; and, an advanced op-
timisation system consisting of a library of optimisers. A description is given of each
element in the design system. To illustrate its use and flexibility, five different applica-
tions of the system to a gas turbine are described. These are: optimisation of the guide
vanes in the bypass duct to minimise excitation of the fan rotor; the same bypass guide
vane optimisation using sensitivity gradients from the adjoint solver; optimisation of a
compressor stage to improve efficiency whilst constraining flow rate, pressure ratio and
outlet flow angle; minimisation of the forced excitation of a turbine rotor by modifying the
wake of the upstream nozzle guide vane; and, optimisation of a fan rotor to reduce tone
noise.

1
Leigh Lapworth and Shahrokh Shahpar

1 INTRODUCTION
The design of individual gas turbine components using CFD is now commonplace
[1]. Traditional design by analysis methods are increasingly being supplemented with
automated design systems [2] and the use of optimisation systems [3]. At the same time,
fluid machinery can now be modelled and analysed to an unprecedented level using CFD
on powerful multi-processor computers or clusters.
Although CFD can provide essential information to aid the physical understanding of
complicated flow fields, it is generally the requirement to design/modify geometry that
drives the application of CFD. When applied to an individual component, the physi-
cal understanding gained from CFD is often able to guide design improvements to that
component. However, fluid machinery is generally characterised by the interaction of a
number of components, such as multistage turbomachinery; the intake and the fan rotor;
the combustor and the upstream diffuser; and, so on. A truly optimal design can only
be achieved by accounting for all the component interactions. It is here that the design
by analysis approach becomes limited - a designer may know how he/she would like to
change the flow field, but changing the geometry to achieve that is much less intuitive
than in single component design. If the requirement to meet a number of constraints is
added, then design by analysis becomes a very crude tool.
This paper describes an automated design system that has been developed specifically
with multi-component fluid machinery in mind. Section 2 describes the main elements of
the design system. Section 3 describes five novel applications of the system.

2 ELEMENTS OF THE DESIGN SYSTEM


The design systems consists of the following processes:

• Parametric representation

• Geometry construction

• Mesh generation

• CFD solution

• Data extraction and functional evaluation

• Optimisation

The systems that implement these processes are described in the following sections.
An underlying theme of all the systems is that they have a batch execution mode that
allows them to be run automatically by the optimiser.

2
Leigh Lapworth and Shahrokh Shahpar

2.1 Parametric representation


Although the parametric representation, geometry construction and mesh generation
are logically separate processes, they are intimately linked with the objective of creating a
CFD mesh in the shortest possible time. With this in mind, an integrated design system
has been developed [4] called PADRAM (Parametric Design and Rapid Meshing System)
which provides a very efficient and robust system for parametric geometry and mesh
creation.
The parametrisation follows the strategy successfully adopted in previous work [2] of
using parameters that are already familiar to the designer. In the context of turbomachin-
ery blades, the parameters include: lean, sweep, stagger, inlet and outlet blade angles,
camber distribution and thickness distribution. The parameters can be independently
specified at a number of radial heights to produce a complete 3D design space. For
multi-passage and/or multi-stage turbomachinery applications, the parameters can be in-
dependently specified for every blade in the calculation domain. Figure 1 illustrates how
a single blade within a ring of blades can be designed independently of the other blades
in the ring. For multipassage simulations, both the pitch and the relative axial position
between adjacent blades are also design parameters.

Figure 1: Independent design of a single blade within a ring.

The philosophy of using familiar engineering parameters has been extended into other
application areas. For example, the design system for nacelles uses parameters such as
droop angle, scarf angle, highlight radius and intake duct area ratio.

2.2 Geometry construction


For constructing the geometry of turbomachinery blades, the approach taken is, typ-
ically, one of perturbing an existing base geometry. The blade geometry may be defined

3
Leigh Lapworth and Shahrokh Shahpar

using either a number of pre-defined blade-to-blade stream sections spanning the blade
from hub to tip; or, as a single three-dimensional NURBS entity. If a NURBS entity is
used then PADRAM creates its own set of pseudo stream sections on which to generate
the CFD mesh.
Since both the geometry and the meshes are constructed stream surfaces, an important
aspect of the design system is definition of these surfaces in three-dimensional space. In
PADRAM, the stream surfaces are defined in polar coordinates as Si (r, θ, z), where i is
the section index. Using non-dimensional parametric coordinates, m0 and θ, a typical
surface can be defined as:

rs = r(m, θ) (1)
θs = θ(m, θ)
zs = z(m, θ)

where, for axi-symmetric stream surfaces:


Zz √ 2
0 dr + dz 2
m = (2)
z
r(z)
0

r = r(z)

z0 is an arbitrary reference and m0 is the non-dimensional distance along a stream


section which will be zero at z0 .
PADRAM starts by transforming the stream sections into parametric two-dimensional
planes, using the co-ordinates θ and m0 . As r is greater than zero for all z coordinates, m0 is
a monotonic function of z, hence a unique inverse function exists to map the computational
coordinates back to the physical, three-dimensional polar coordinates. The advantage
of the above transformation is that the angles are preserved and the mesh-generation
procedure deals with plane sections only.
For non-turbomachinery geometries, PADRAM interfaces to CAD through IGES files.
PADRAM uses a library of stylised geometries, so that it has a number of pre-defined
rules for cleaning IGES files and constructing geometry.

2.3 Mesh generation


The PADRAM mesh generator can, very rapidly, produce good-quality viscous meshes
for multi-passage, multi-stage turbomachinery using 2D, quasi-3D or full 3D blade geome-
try. PADRAM makes uses of both transfinite interpolation and elliptic grid generators to
generate hybrid C-O-H meshes [4]. An orthogonal body-fitted O-mesh is used to capture
the viscous region in the vicinity of the blade(s). A C-mesh is used for any bifurcations
in the domain, such as pylons and drive farings in bypass duct applications. Elsewhere
H-meshes are used to patch adjacent blocks of O- or C- mesh together; and, to extend the

4
Leigh Lapworth and Shahrokh Shahpar

mesh to meet the upstream, downstream and periodic boundaries. The mesh is indepen-
dently generated for every stream section, hence three-dimensional meshes are produced
easily from stacked two-dimensional meshes with no mapping required to transfer the
meshes radially. This avoids mesh morphing, and ensures that good quality meshes are
created at every height even if the geometry varies considerably from hub to tip. For
blades with tip gaps, a separate H-mesh is generated in the gap. Mesh generation is
extremely fast, requiring no more than a few seconds to generate a single passage 3D
mesh. A mesh of several million nodes can be generated in approximately one minute
on a 3.0GHz PC. A typical PADRAM mesh for a single blade passage is shown in figure
2(a).

(a) Single blade mesh (b) Bypass duct mesh


Figure 2: PADRAM meshes for single and multi-passage applications.

As stated above, one of the key features of PADRAM is the fact that it is intimately
linked with the parametric geometry definition. For example, within a ring of bypass
guide vanes, the geometry of each vane can be specified independently of the others
either by using a different geometry definition file or by using one of the several design
parameters within PADRAM. Another important feature of PADRAM is that it is a true
multi-passage meshing system. The geometry of each blade in the annulus can be varied
independently and a good viscous mesh for the complete ring of blades can be generated
in one step. Figure 2(b) shows a PADRAM mesh for a bypass assembly consisting of 52
guide vanes (each defined independently of the others) a pylon and a radial drive faring.

2.4 CFD solution


CFD solutions are generated using the Rolls-Royce plc. HYDRA-CFD code. HYDRA-
CFD is a suite of non-linear, linear and adjoint solvers being developed collaboratively by
Rolls-Royce plc. and its University partners [5].
HYDRA-CFD is a general purpose code for hybrid unstructured meshes which uses an

5
Leigh Lapworth and Shahrokh Shahpar

efficient edge-based data structure [6]. The multi-block PADRAM meshes are converted
into this data structure using a pre-processor. The flow equations are integrated around
median- dual control volumes using a MUSCL based flux-differencing algorithm. The
discrete flow equations are predconditioned using a block Jacobi preconditioner [7] and
iterated towards steady state using the 5-stage Runge-Kutta scheme [8]. Convergence
to steady state is further accelerated through the use of an element-collapsing multigrid
algorithm [9]. The flow solver runs in parallel on both shared and distributed memory
machines using domain decomposition. The parallel multigrid capabilities are essential
for generating CFD solutions in the elapsed time needed for effective use of optimisation.
The HYDRA-CFD framework is novel in the fact that it has linearised unsteady [10]
and adjoint solvers [11] built on top of the non-linear solver. The linear and adjoint solvers
represent a full linearisation of the turbulent steady flow equations. Examples of the use
of both the linear and adjoint solvers within an optimisation process are described in this
paper.

2.5 Data extraction and function evaluation


In order to drive the optimisation process, the values of both cost and constraint func-
tions must be extracted from the CFD solution. The key element here is that the design
system must be able to extract these values as a batch process. Typically, design func-
tionals involve integrations over boundary planes. These integrals may be over the entire
boundary plane, as in the case of inlet or outlet flow rate; or, in the circumferential
direction at a series of radial positions, as in the case of prescribed inlet or outlet profiles.
Two approaches to data extraction have been adopted. The first is to export the
HYDRA-CFD solution into a multi-block PLOT3D flow file - this is only possible for
multiblock meshes and relies on a stored mapping generated by the HYDRA-CFD pre-
processor. Once the flow solution is in PLOT3D format, data can be extracted using
a range of structured mesh post-processors. The second approach is to process the un-
structured mesh directly - this relies on constructing a series of interpolating lines in the
circumferential direction which overlay the unstructured mesh.

2.6 Optimisation
Optimisation is performed using the SOFT (Smart Optimisation For Turbomachin-
ery) system [3] being developed by Rolls-Royce plc. and its university partners. SOFT
provides a library of different optimisers; design of experiments techniques; statistical
analysis of variations; and, advanced response surface models. It is well known that no
single optimisation technique performs better than others across a range of engineering
applications. Hence, SOFT provides a library of optimisers which fall into four broad
categories:

• Explorative methods, such as Simplex based methods.

6
Leigh Lapworth and Shahrokh Shahpar

• Gradient based methods, such as the Modified Feasible Descent and Sequential
Quadratic Programming techniques which are used in this paper.
• Evolutionary type methods, such as the Simulated Annealing technique used in this
paper.
• Hybrid methods, such as Tabu search which combine elements of the previous meth-
ods.
An issue that must often be addressed with CFD based optimisation is the run-time
associated with the performing a simulation. With access to a compute cluster, this
can be addressed through the parallel capabilities in the flow solver. In addition, SOFT
provides a number of approximation techniques for reducing the number of expensive
CFD simulations which must be performed [3].
The SOFT system provides the user with a graphical interface for constructing the work
flow, such as the one shown in figure 5, which defines the sequence of operations from the
setting of design parameters to the evaluation of the cost and constraint functions. The
GUI also provides a visualisation capability for plotting the progress of the optimisation.

2.7 Adjoint CFD


Consider an objective function I(U, α) which is a function of the flow variables U and
a set of design parameters α. The sensitivity of the cost function to the design variables
is:
dI ∂I ∂Uk ∂I
= + (3)
dαj ∂Uk ∂αj ∂α j
Using the fact that the flow variables satisfy the Navier-Stokes euqations R(U, α) = 0
and introducing the adjoint variables, v,
à !−1
∂I ∂Ri
viT = (4)
∂Uk ∂Uk
equation 3 can be rewritten as:
dI ∂Ri ∂I
= −viT + (5)
dαj ∂αj ∂αj
Equation 4 can be re-written to give the adjoint equation
à !T à !T
∂R ∂I
v= (6)
∂U ∂U
The advantage of the adjoint equation is that it depends only on the cost function I and
not on the design parameters α. Hence, for applications where there are many design
variables but only one, or a small number, of cost and constraint functions the adjoint
approach is a computationally efficient approach for computing design sensitivities.

7
Leigh Lapworth and Shahrokh Shahpar

3 DESIGN APPLICATIONS
In order to demonstrate the flexibility of the design system, the following applications
are presented:
• Optimisation of bypass guide vanes using steady state CFD,

• Optimisation of bypass guide vanes using adjoint CFD,

• Optimisation of a compressor stage using multi-stage CFD,

• Reduction of turbine forced response using linear unsteady CFD,

• Reduction of fan rotor noise using steady state CFD.


These applications make use of a range of capabilities from both the CFD solver and
the optmisation library. The use of both approximation techniques and PC clusters to
make the CFD simulation times tractable for the optimiser is also demonstrated.

3.1 Optimisation of bypass duct guide vanes


The layout of a modern high bypass ratio gas turbine engine is shown in figure 3(a).
Within the bypass duct there are large scale blockages due to the pylon (which attaches
the engine to the wing) and the faring around the radial drive shaft (RDS). Upstream of
these blockages is a ring of outlet guide vanes (OGVs) which remove the swirl from the
flow leaving the fan rotor, see figure 3(b). It is has been shown [12] that the pylon and
RDS generate a circumferential distortion in the static pressure field which can propagate
upstream, through the bypass OGVs, and excite the fan rotor. It has also been shown in
[12] that by circumferentially varying the OGV geometry the level of pressure distortion
reaching the fan, and hence its excitation, can be drastically reduced. Theoretically, the
distortion reaching the fan can be completely eliminated if every OGV is allowed to be
different. However, this is not a cost-effective solution and design options include: a single
OGV with a variable stagger pattern; and, a small number, typically 3 or 5, of different
OGVs arranged in blocks around the annulus.
The PADRAM-HYDRA-SOFT system has been applied to the design of a ring of 52
bypass OGVs to reduce fan forcing. In this example, all 52 OGVs are identical but their
stagger, or setting, angle is allowed to vary around the annulus up to a user specified
maximum value.
Notionally, each OGV stagger angle is independent of the others and the design space
has 52 degrees of freedom. In practice, the stagger angle should vary more smoothly
around the annulus its circumferential distribution can be represented by a Fourier series
of the form:
M ·
X µ ¶ µ ¶¸
2πi 2πi
ζi = A0 + Aj sin j + Bj cos j (i=1,N) (7)
j=1 N N

8
Leigh Lapworth and Shahrokh Shahpar

(a) Layout of a high bypass ratio gas turbine engine. (b) CFD solution for bypass duct.
Figure 3: Engine cut-away and bypass duct components: OGVs, pylon and radial drive faring.

where ζi is the stagger angle of the ith OGV; N is the number of OGVs; and, M is
the number of harmonics used in the Fourier expansion. In the current calculations,
seven Fourier harmonics have been used, leading to a design space with 15 parameters
(A0 , A1 , B1 , . . . , A7 , B7 ). In order to ensure that the all the stagger angles are less than
the user specified constraint, ζconst ≥ 0, the stagger angles from equation 7 are modified
by:
ζi
ζi = min (ζmax , ζconst ) (8)
ζmax

where ζmax = maxi=1,N (|ζi |) is the maximum, unmodified, stagger angle.


The objective function, I, for the design is the root mean square of the deviation in
static pressure from a uniform value at the inlet to the bypass duct:
v
uP
u N
u (p − p̄)2
t i=1 i
I= (9)
N
N
P
1
where p̄ = N
pi is the mean inlet static pressure, with N the number of mesh
i=1
nodes on the inlet boundary. Hence, I = 0 corresponds to a completely uniform pressure
downstream of the fan rotor and, therefore, a zero excitation.
Using SOFT, the OGV stagger pattern has been optimised using the dynamic hill
climbing (DHC) approach. Two constraint values, ζconst were used: 3◦ and 6.43◦ .
Figure 4(a) shows the convergence of the optimiser in the case when the stagger vari-
ation is limited to ±3◦ . Figure 4(b) shows the variation in static pressure around the
circumference at the inlet to the bypass duct. With uniform OGVs there is a peak to

9
Leigh Lapworth and Shahrokh Shahpar

! "
# $ % " " & "
# $ % " " & "

(a) Convergence of the optimiser. (b) Pressure variation at OGV inlet.


Figure 4: Optimisation of bypass OGV stagger variation.

trough variation in static pressure of approximately 6kPa. With the stagger variation
limited to ±3◦ , the optimiser achieves a 57% reduction in the objective function. With
the stagger variation limited to ±6.43◦ , a reduction of 85% is achieved.

3.2 Bypass OGV optimisation using adjoint CFD


The bypass OGV optimisation described in 3.1 is an example where the number of
design parameters is significantly larger than the number of cost and constraint functions.
In fact, the constraints apply directly to the design parameters and do not require the
evaluation of an auxiliary equation. This is precisely the situation where the use of
adjoint CFD, as described in section 2.7, should perform well. This section demonstrates
the adjoint capabilities of the PADRAM-HYDRA-SOFT system using the same bypass
OGV case as used in 3.1.
In order to utilise adjoint CFD, the chosen optimiser must be one that is able to use
gradient information effectively. In this example, the Sequential Quadratic Programming
(SQP) technique is used. This is well suited to the use of adjoint CFD since it uses a
pre-defined sequence of perturbations to each design parameter in turn, in order to build
an approximation to the Hessian matrix of second derivatives. The Hessian matrix is then
used to compute a new, more optimal, point in the design space.
When using adjoint CFD, one factor that must be addressed is that only first order
sensitivity gradients are computed. In a design space that is highly non-linear, the adjoint
gradients must be coupled with a procedure for re-evaluating the base and adjoint flow
fields at regular intervals in the optimisation cycle. The decision of whether to use the
existing adjoint gradients or recompute the steady and adjoint flow fields is based on a
simple Euclidean distance test:
v
uN ³ ´2
uX
dα = t αi − αiold (10)
i=1

10
Leigh Lapworth and Shahrokh Shahpar

where, in this instance, N is the number of design parameters, αi .


When the Euclidean distance is smaller than a user specified threshold, e.g. 0.01, the
adjoint sensitivity gradients are used to compute the new value of the objective function;
otherwise, the objective function is computed from a new non-linear steady flow solution
along with a new adjoint solution.
Figure 5 shows the work flow for the optimisation of bypass OGVs using adjoint CFD.
The modules setadj and rhslin are part of the HYDRA-CFD suite which compute the
adjoint boundary conditions and flow residuals on a perturbed mesh respectively.

! & & α' α & & ( ξ) !

! # $
" #

" #
# $ " #
# %
"

"
"

Figure 5: Flow chart for adjoint optimisation

In this example, the maximum stagger variation is limited to ±3◦ . Figure 6(a) shows
the convergence of the SQP optimiser, and figure 6(a) shows the solution after 280 design
cycles which achieves a reduction of 38% in the objective function. Comparing the con-
vergence paths of the DHC (figure 4) and the SQP (figure 6) optimisers, it is seen that
the SQP optimiser consists of a series of plateaux. Each plateau consists of the series
of perturbations to each design parameter - it is these sensitivities that are computed
using the adjoint solution. Once all the perturbations have been evaluated, the Hessien
matrix is assembled and the optimiser moves to a new point in the design space. At this
point, a new steady state solution is evaluated and if the Euclidean distance threshold
has been exceeded a new adjoint solution is also computed. Hence, in the 280 iterations
of the optimiser shown in figure 6 there are only 16 steady and adjoint CFD simulations.
Whereas, the 206 iterations shown in figure 4 each require a steady CFD solution.
The convergence of the adjoint SQP optimisation (figure 6) shows a number of over-
shoots in the latter stages of the convergence. These are due to the fact that part converged
steady and adjoint solutions were being used. This practice has been used successfully
in non-linear optimisation [13] - as the optimiser converges toward the optimum, the

11
Leigh Lapworth and Shahrokh Shahpar

! " #
! $

(a) Convergence of the optimiser. (b) Pressure variation at OGV inlet.


Figure 6: Optimisation of bypass OGV stagger variation using adjoint CFD.

convergence of the steady CFD becomes tighter because the solution from the previous
optimisation step is used as an initial guess. This practice appears to be less success-
ful when using adjoint CFD because the part convergence can lead to one or more poor
sensitivity gradients which are propagated forward to the next adjoint solution. This
also explains why the adjoint optimisation does not achieve as large a reduction in the
cost function as the non-linear approach. Further research is ongoing to understand how
best to use adjoint sensitivity gradients within a computationally efficient optimisation
scheme.

3.3 Optimisation of a compressor stage


Modern axial flow compressors consist of many stages of closely coupled blade rows
which must be modelled using multistage CFD techniques to obtain accurate performance
predictions. Similarly, design/optimisation of each blade row in turn is generally over-
constrained by the need to maintain the matching with the adjacent rows. More optimum
designs are, clearly, possible if the compressor is designed in its entirety as a multi-stage
machine. The capability of the PADRAM-HYDRA-SOFT system to design multistage
compressors is demonstrated in this section.
The example chosen is that of designing the 3rd stage of a 4-stage low speed compressor
rig. The datum stage is shown in figure 7. The rotor 3 inlet Mach number is approximately
0.2. The compressor is cantilevered with tip gaps of 1.18% span on the rotor and 1.12%
span on the stator. There are 101 rotors and 134 stators and the calculation is run with
a mixing plane between the rotor and stator.
The design space consists of lean, sweep, stagger, and camber angle at the trailing edge.
The same set of design parameters are used for both the rotor and stator but, otherwise,
the rotor and stator designs are independent. The design parameters are specified via

12
Leigh Lapworth and Shahrokh Shahpar

Figure 7: Datum stage 3 of a 4 stage low-speed compressor.

control points at 0, 25, 50, 75 and 100% span on both the rotor and stator. A cubic spline
is fitted through the control points to produce design perturbations at the intervening
radii. In this example, the sweep, stagger and camber angles are not allowed to vary at
either the hub or casing - this would allow existing fittings on the rig to be re-used. The
lean angle is also fixed at the casing since this is invariant to any bulk circumferential
offset. This reduces the design space to 26 independent parameters for the stage.
The objective function is
1 − ηstage
I= (11)
1 − ηbase
with the constraint that
¯ ¯ ¯ ¯ ¯ ¯
¯ ṁnew − ṁbase ¯ ¯ Rnew − Rbase ¯ ¯¯ βnew − βbase ¯¯
4 ¯¯ ¯+¯
¯ ¯
¯+¯
¯ ¯ ¯≤²
¯
(12)
ṁbase Rbase βbase
where ṁ is the inlet flow rate; R is the stage pressure ratio; β is the stator exit flow
angle; and, ² is a user defined threshold. By constraining the mass flow, pressure ratio
and exit whirl angle the matching between the stage being designed and the other 3 stages
should be maintained.
The optimisation is performed using the Simulated Annealing optimiser within SOFT.
Figure 8 shows the initial convergence of the optimiser together with snapshots of the
design at three points in the convergence history. After running the optimiser for 400
iterations a 0.72% improvement in stage efficiency was achieved.

3.4 Minimisation of turbine forced response using linear unsteady CFD


Forced response is the excitation of one blade row by another due to a periodic forcing
from either the wake from the upstream blade; or, the potential field from the downstream
blade. With trends towards increased blade loading and reduced axial gaps between blade
rows, the higher unsteady forces and hence vibratory stresses levels can lead to High Cycle
Fatigue (HCF).
A common approach to reducing forced response levels is the so called wake-shaping
technique. The rationale is that if the vane wake is completely in phase when it reaches
the rotor, the latter feels a large impulsive force at discrete instants in time. Whereas, if

13
Leigh Lapworth and Shahrokh Shahpar

Figure 8: Optimisation for stage efficiency

the vane wake is out of phase (i.e. leant circumferentially relative to the rotor), the force
experienced by the rotor is distributed over a longer time interval and the peak forcing
level, and hence likelihood of HCF, is lower. Leaning the vane wake can be achieved by
restacking the stream sections defining the vane geometry. Since the vane, typically, has
a high exit flow angle this is a very powerful design mechanism and the vane wake can be
leant significantly with relatively small changes in stacking.
The design objective is usually the HCF endurance limit, the evaluation of which is
a 2 step process. In the first step, the unsteady forces, i.e. pressures, on the rotor
are determined from a CFD simulation. In the second, a mechanical analysis of the
rotor is performed to determine the forcing levels. The mechanical analysis uses a modal
representation of the structure and solves an equation of the form:

M ẍ + C ẋ + Kx = F1 sin(ωt + φ) + F2 (13)

where x is the displacement; M , C and K are the mass, damping and stiffness matrices
respectively; F1 is the unsteady force from the CFD calculation at a frequency ω and
an inter-blade phase angle φ; and, F2 contains any additional forces such as non-linear
damping.
The result of the mechanical analysis is the amplitude of the maximum displacement,
usually at the rotor tip. This is fed into a database of material properties along with the
steady and alternating stress levels on the rotor and the metal temperature to produce
an estimate of the HCF endurance level.
The evaluation of the unsteady rotor pressures requires an unsteady CFD simulation,
which is expensive if a time-accurate calculation is used. Fortunately, time linearised
unsteady CFD methods have proved a successful means of capturing the primary forcing
mechanisms [14] and are computationally very efficient. Here, the unsteady flow in the

14
Leigh Lapworth and Shahrokh Shahpar

rotor is solved as a perturbation to its steady state flow field at a given frequency and
interblade phase angle (defined by the numbers of rotors and vanes). The unsteadiness in
the rotor is driven by the incoming wake. This is extracted from a steady CFD solution
of the upstream vane at the axial position corresponding to the rotor inlet. At each radial
height, the vane wake is decomposed into Fourier harmonics. The amplitude and phase
of the harmonic of interest provide the inlet boundary conditions for the linear unsteady
rotor calculation.
The capability of the PADRAM-HYDRA-SOFT system to combine aerodynamic and
mechanical analysis codes; and, to utilise advanced techniques such as linearised unsteady
CFD is demonstrated in this section. The example chosen is that of designing a High
Pressure (HP) turbine vane which minimises the HCF endurance level of the downstream
HP rotor. More exactly, it is the HCF limit of the 2nd edge (2E) mode which is to be
minimised; with the constraints that the HCF limits in the 1st torsion (1T) and 2nd flap
(2F) modes should not increase.
The design is broken into 2 phases:

• Phase 1: Taking the wake amplitudes from the base vane, use the linear unsteady
CFD solver to find the phasing, i.e. lean, of the wake which minimise the 2E HCF
limit.

• Phase 2: Taking the optimal phasing of the wake, use inverse design techniques to
find the stack of the vane that delivers the required wake.

The design space consists of perturbations to the phase of the wake from the base vane
specified via control points at 0, 25, 50, 75 and 100% height. A cubic B-spline is fitted
through the values at the control points to produce perturbations at the intervening radii.
Since, the wake phase is invariant up to an additive constant, the phase perturbation at
0% height is set to zero. The phase perturbations are also normalised to lie within ±2π/N
where N is the number of rotor blades. Hence, the wake can be leant by no more than
one rotor pitch in either circumferential direction.

3.4.1 Phase 1
The evaluation of the cost and constraint functions consists of: perturbing the incoming
wake according to the design parameters; generating the corresponding linear unsteady
solution; performing the forced response analysis; and, computing the HCF endurance
levels for the 2E, 1T and 2F modes.
This example is also used to demonstrate the approximation techniques available in
SOFT. The Response Surface Method (RSM) is used to construct an approximate model
of the design space which can be used to provide function evaluations in lieu of the simu-
lation codes. The RSM requires an initial population of simulations is order to construct

15
Leigh Lapworth and Shahrokh Shahpar

a first approximation to the response surface. The design parameters for the initial popu-
lation are determined using the Taguchi Design of Experiments (DoE) technique. In this
example, a three level Taguchi DoE was used which, for 4 design parameters, consists of 9
experiments. A further 10, manually prescribed, experiments were also performed. This
gave a population of 20 simulations (including the initial design) from which the initial
response surface was constructed. The RSM method proceeds by performing an optimi-
sation using the response surface to evaluate the HCF levels rather than the simulation
codes. Once an optimum has been obtained, a verification run of the simulation codes is
performed using the optimum wake shape parameters. If the results of verification run
are sufficiently close to the RSM value then the process has converged; otherwise, the new
results are added to the RSM population and the response surface is updated.

Structural mode Initial HCF level Minimum HCF level Redesigned Nozzle
1st Torsion 0.39467 0.0518 0.07001
2nd Flap 0.14444 0.0636 0.04815
2nd Edge 10.91056 1.7974 4.68130
Table 1: Initial and optimum HCF endurance levels.

For this case, the optimisation is performed using the Simulated Annealing (SA) tech-
nique and a total of 12 verification simulations are needed for the RSM to converge. The
2nd and 3rd columns of table 1 show the initial and optimised HCF endurance levels for
the three modes of interest The optimisation has not only achieved a reduction of 87%
in the objective function it has also bettered the constraints with reductions of 83% and
56% reductions in the 1T and 2F modes.

Figure 9: Inverse designed nozzle giving required wake shape.

16
Leigh Lapworth and Shahrokh Shahpar

3.4.2 Phase 2
The inverse design of the nozzle to deliver the required wake shape follows the process
described in [2]. The resulting nozzle is shown in figure 9. The HCF levels were recom-
puted using the wake from the redesigned nozzle and are shown in the 4th column of table
1. The rise in the 2E HCF level is due to the fact that the inverse design procedure is
not able to deliver a new nozzle which gives exactly the desired wake shape. Also, it can
be seen from figure 9 that the changes in the stack of the nozzle are somewhat extreme,
particularly from the point of view of internal cooling passages. This was not imposed
as a constraint in the design which is why such large reductions in the objective function
were achieved.

3.5 Low noise fan design


In a modern high bypass ratio gas turbine engine, a significant amount of noise is
generated by the fan rotor. The noise falls into 2 categories: tone noise which is generated
at discrete frequencies and generally associated with blade row interactions and shocks;
and, broadband noise which is generated over a broad spectrum and generally associated
with boundary layers and turbulence phenomena. Tone noise is much more amenable to
analysis by CFD and the steady and linearised unsteady solvers in HYDRA-CFD have
been successfully applied to a number of tone noise test cases [15]. Indeed, there is
sufficient confidence in the calculation of tone noise to consider using the optimisation to
design components specifically to reduce noise levels. To demonstrate this, the PADRAM-
HYDRA-SOFT design system has been applied to the reduction of tone noise generated
by a fan rotor.

Figure 10: High bypass ratio fan rotor with static pressure field and PADRAM generated mesh.

Figure 10 shows a typical high bypass ratio fan rotor along with its static pressure field

17
Leigh Lapworth and Shahrokh Shahpar

on the hub, casing and rotor surfaces. It is the strong shocks in outer part of the span,
as seen in the casing surface pressures, which are the source of the noise. Figure 10 also
shows a blade to blade section from the PADRAM mesh. This mesh has approximately 1.7
million nodes. This is much finer than would be needed for an aerodynamic optimisation
and is needed to resolve the acoustics with a sufficient number of points per wavelength.
The design space consists of sweep and lean of the rotor blade sections at 90% and 100%
height. Below 80% the blade sections are unmodified. Between 80-90% and 90-100% the
design parameters are linearly interpolated to the other sections. The objective function
is the amplitude of the 1st radial harmonic of the tone at one blade passing frequency (1
BPF). The objective functions is evaluated one chord upstream of the fan leading edge and
at aerodynamic conditions corresponding the fan working line. Multi-point simulations
are used to ensure the optimisation is performed at a point on the working line.

Figure 11: Optimisation of fan rotor tone noise.

Figure 11 shows the convergence of SOFT using the dynamic hill climbing optimiser.
The reduction in the amplitude of the first radial harmonic corresponds to a potential
reduction in noise of approximately 9dB. In this case, the HYDRA-CFD calculations were
run on a PC cluster. Each calculation used 60 processors and took approximately 2 hours
elapsed time. Hence, the majority of the predicted reduction in the 1 BPF tone noise was
achieved after 2 days of running the optimiser.

4 CONCLUSIONS
A new design system based around parametric design and mesh generation; and, ad-
vanced CFD and optimisation techniques has been presented. Five applications of the
system have been presented which demonstrate the following, novel, attributes of the
system:
• A true multi-passage, multi-row design and meshing capability.

18
Leigh Lapworth and Shahrokh Shahpar

• The use of gradient sensitivities from adjoint CFD.

• The simultaneous design of multiple components.

• The coupling of aerodynamic and mechanical simulation codes.

• The use of advanced CFD techniques, such as linearised unsteady methods.

• The use of approximation techniques to supplement expensive CFD simulations

• The effective use of parallel CFD and PC clusters to reduce CFD simulation times
to an acceptable level.

• The incorporation aerodynamic, aeromechanic and aeroacoustic design objectives.

5 ACKNOWLEDGEMENTS
The authors gratefully acknowledge the permission of Rolls-Royce plc to publish this
paper. We would also like to acknowledge the contribution of our colleague John Coup-
land to section 3.5; and, to Davide Giacche, Diego Benito and Piero Distefano for their
contributions to sections 3.1, 3.2, 3.3 and 3.4.

19
Leigh Lapworth and Shahrokh Shahpar

REFERENCES
[1] B.L. Lapworth. Challenges and Methodologies in the Design of Axial Flow Fans for
High Bypass Ratio Gas Turbine Engines using Steady and Unsteady CFD. Advances
of CFD in Fluid Machinery Design, Ed. Elder, Tourlidakis and Yates, Professional
Engineering Publishing, 2002.

[2] S.Shahpar and B.L. Lapworth. A Forward and Inverse Three-Dimensional Linear De-
sign System for Turbomachinery Applications. 4th ECCOMAS Computational Fluid
Dynamics Conference, Athens, 1998.

[3] S.Shahpar. SOFT: A New Design and Optimisation Tool for Turbomachinery. Evolu-
tionary Methods for Design, Optimisation and Control, Ed. Giannakoglou, Tsahalis,
Periaux, Papailiou, Fogarty, CIMNE, Barcelona, 2002.

[4] S.Shahpar and B.L. Lapworth. PADRAM: Paramtric Design and Rapid Meshing
System for Turbomachinery Optimisation. Paper GT-2003-38698, ASME Turbo Expo,
Atlanta Georgia, June 16-19, 2003.

[5] B.L. Lapworth. Hydra-CFD: A Framework for Collaborative CFD Development In-
ternational Conference on Scientific and Engineering Computation (IC-SEC), Sin-
gapore, June 30 - July 02, 2004.

[6] P. Moinier, J.-D. Müller and M.B. Giles. Edge-based multigrid and preconditioning
for hybrid grids. AIAA Journal, Vol. 40, No 10, 2001.

[7] P. Moinier and M.B. Giles. Preconditioned Euler and Navier-Stokes Calculations on
Unstructured Grids. 6th ICFD Conference on Numerical Methods for Fluid Dynam-
ics, Oxford, UK, 1998.

[8] L. Martinelli. Calculations of Viscous Flows with a Multigrid Method. Ph.D Thesis,
Dept. of Mech. And Aerospace Eng., Princeton University, USA, 1987.

[9] J.-D. Müller and M.B. Giles. Edge-Based Multigrid Schemes for Hybrid Grids. 6th
ICFD Conference on Numerical Methods for Fluid Dynamics, Oxford, UK, 1998.

[10] M.C. Duta, M.B. Giles and M.S. Campobasso. The harmonic adjoint approach to
unsteady turbomachinery design. International Journal for Numerical Methods in
Fluids, 40(3-4), p.323-332, 2002.

[11] M.S. Campobasso, M.C. Duta, M.B. Giles. Adjoint Methods for Turbomachinery
Design. Paper No. 1055, ISABE Conference, Bangalore, India, 2001.

[12] A.B. Parry. Optimisation of bypass fan outlet guide vanes, Paper No. 96-GT-433,
ASME Gas Turbine Conference, 1996.

20
Leigh Lapworth and Shahrokh Shahpar

[13] S. Shahpar, D. Giacche and B.L. Lapworth. Multi-objective Design and Optimisation
of Bypass Outlet-guide Vanes. Paper GT2003-38700, ASME Turbo Expo, Atlanta
Georgia, June 16-19, 2003.

[14] J.G. Marshall and M.B. Giles. Some Applications of a Time Linearized Euler Method
to Flutter and Forced Response in Turbomachinery. Unsteady Aerodynamics and
Aeroelasticity of Turbomachines, edited by T.H.Fransson, Kluwer Academic, Dor-
drecht, NL, 1998.

[15] A.G. Wilson. Application of CFD to Wake/Aerofoil Interaction Noise - A Flat Plate
Validation Case. Paper AIAA-2001-2135, 7th AIAA/CEAS Aeroacoustics Confer-
ence, Maastricht, 28-30 May, 2001.

21

You might also like