10th AIAA/ISSMO Multidisciplinary Analysis and Optimization Conference AIAA 2004-4427

30 August - 1 September 2004, Albany, New York

3d-Duct Design Using Variable Fidelity Method

D. Ghate , A. Isaacs†, K. Sudhakar‡, P. M. Mujumdar‡ and A. G. Marathe‡

Center for Aerospace Systems Design & Engineering (CASDE), Department of Aerospace Engineering, Indian
Institute of Technology Bombay, Powai, Mumbai, India 400 076.

Low Fidelity Analysis (LFA) modules, Design of Experiment (DOE), and surrogate
modeling are brought together with High Fidelity Analysis (HFA) based on
Computational Fluid Dynamics (CFD) to provide an optimization methodology for the
design of an intake duct of a fighter aircraft. Issues such as time scales, gradient
information, and automation, related to the use of HFA in optimization cycles, have been
successfully addressed in the study.

Nomenclature
y& z = Mean Flow Line parameters
1& 2 = Area parameters
d-avg = Average diffusion angle
w = Local wall angle
A = Area of duct cross-section
Cf = Skin friction Coefficient
DOE = Design of experiments
HFA = High fidelity analysis
LFA = Low fidelity analysis
Lm = Length of duct from entry to merger cross-section
M = Mach number
Po = Total pressure
p = Area centroid of cross-section
RSM = Response surface method
R = Radius of curvature of MFL
r = Equivalent duct radius = A /
To = Total temperature
x = X coordinate along duct length

Subscripts:
en = Duct entry
ex = Duct exit
m = Merger section

I. Introduction
Fighter aircraft with single engine embedded in the fuselage typically have Y-shaped intake ducts which are
symmetrical about their central plane. The two arms of the Y-shaped duct merge at a section (henceforth
referred to as merger section), beyond which the duct has single arm symmetric about the vertical plane. The
two arms of the Y are typically ducts with double bends (S-shaped). Examples of aircrafts with Y-shaped intake
ducts include most fighter aircrafts. The present work focuses on the design of Y-ducts. In discussion that
follows the word “duct” will be used in this context and will refer to one half of the symmetric Y-duct. During
the subsonic cruise, the main sources of pressure loss are skin friction along the duct walls, mixing loss at the
merger section, and loss arising out of local pockets of flow separation inside the duct. An S-shaped duct
typically also generates flow distortion and swirl at the exit section1. It is desirable to reduce the pressure loss
across the duct while keeping the distortion levels to acceptable levels for good performance of the engine and
for avoiding instabilities.

Research Assistant.
†
Research Scientist.
‡
Professor.

1
American Institute of Aeronautics and Astronautics
Copyright © 2004 by the American Institute of Aeronautics and Astronautics, Inc. All rights reserved.
 With demands on higher performance and lower uncertainties in the design, a need for HFA in design is felt.
High fidelity analysis has been made possible in the last few decades because of the significant increase in the
computational resources and the maturity of CFD packages2. Though high fidelity analysis offers accurate
analysis of the system, it suffers from the drawbacks of huge computation requirements, need of accurate initial
& boundary conditions, and overall increase in the complexity of the design process in terms of data
management and inter-disciplinary communication. With this perspective, a case study has been taken up to
demonstrate the use of a commercial high fidelity module in the optimization cycle while tackling the issues of
gradient information, limited computational resources and inter-module communication. The problem of
optimization of an inlet duct of a typical fighter aircraft for minimum pressure loss and exit distortion has been
discussed in this paper.
The paper discuses the problem statement followed by the description of the mixed-fidelity optimization
methodology adopted. Following this, various steps of the optimization process along with the associated issues
have been presented. The paper concludes with the results and a brief discussion about the possible future work.

II. Problem Statement

An intake duct has its area constrained at the entry and exit sections from low speed operation and engine
size considerations respectively. Its length is prescribed from the aircraft layout considerations. Duct design is
thus an exercise in arriving at the geometry of the duct between prescribed entry and exit stations – i.e. entry and
exit station cross-section shape and their longitudinal locations are given. As stated earlier, the objective is to
maximize total pressure recovery for one or more operating conditions under the constraints of acceptable
distortion and swirl coefficients at the compressor face.
Duct design using low fidelity analysis typically involves use of empirical relations and thumb rules
developed over years through experimentation and observations. Pressure loss can be computed with good
accuracy using low fidelity analysis for a duct that is free from pockets of flow separation. But the calculation of
distortion and swirl for arbitrary 3D-ducts requires CFD simulations. Hence, this is an optimization problem that
necessarily requires high fidelity analysis at the same time providing an opportunity to demonstrate the use of
mixed fidelity optimization methodology.
FLUENT® is selected as the commercial CFD package for this study for reasons elaborated later. One issue
to be tackled in the use of commercial CFD packages is that gradient information is not available directly.
Options such as adjoint gradient calculation, automatic differentiation and complex step method get eliminated
because of the unavailability of the source code of the CFD Solver. One option for gradient calculation is the
finite difference method. But, as finite difference requires (N+1) HFAs for gradient calculation at a given point
in design space with N design parameters, it proves to be a prohibitively expensive method for design
optimization using CFD simulations. Moreover, the finite difference method also demands identification of an
appropriate step size3 for each parameter. Hence, RSM in conjunction with DOE is considered. However, due to
the large computation time required per CFD analysis of such ducts4, a direct application of RSMs like
sequential local RSM (SLRSM) is also not viable within acceptable time scales with available computer
resources. Therefore to reduce overall time for the total optimization, various strategies have been demonstrated
in the rest of the paper.
Literature survey on all the experimental work done on the S-shaped ducts can be found in reference 2.
Independent studies have been reported by Harloff4 and Loeffer and Jenkins5 for calculation of the distortion at
the exit of a S-shaped duct using Navier-Stokes solvers with turbulence models. However, these studies do not
employ any systematic optimization methodology using CFD. Doyle Knight et. al.2,6 have performed an
optimization study for a S-shaped duct with circular entry and exit sections using GASP CFD solver and CFSQP
optimizer. This group has reported gradient calculations using finite difference methods.
Present study explores the possibility of reduction in computational requirements of the optimization process
using variable fidelity methods and surrogate models.

III. Optimization Methodology

The adopted optimization methodology uses a mixture of low fidelity analysis and high fidelity analysis.
Figure 1 illustrates the overall optimization process. Parameterization of the duct geometry is performed in the
first step. This step is crucial in reducing the overall computation costs at all stages of optimization.
Parameterization is followed by a constrained optimization of the duct geometry using low fidelity analysis for
pressure recovery and the thumb rules based on existing design practices over the full design space. The low
fidelity analysis is explained later.
Optimization using HFA starts now. The optimum located using LFA becomes baseline configuration for
this stage. Low fidelity rule based constraints are retained to provide a reduced design space. In the next step,
DOE is performed in this reduced space in order to identify points at which CFD analysis is performed. A
surrogate model is generated using the DOE points for the distortion coefficients obtained from CFD analyses.

2
American Institute of Aeronautics and Astronautics
In the final step, a gradient-based optimizer is used in conjunction with the surrogate model to obtain the
optimized solution. Various steps of the optimization process are discussed hereafter in greater detail.
This two-step multi-fidelity optimization methodology is computationally cheaper than a single stage high
fidelity optimization process using CFD analysis without any dilution of the quality of results.

CFD analysis
LFA Optimum at DOE points

RS for
DOE in the PR & DC60
reduced space

LFA
Optimization HFA Optimization
Constraints

Parameterization

Figure 1. Block diagram illustrating complete optimization methodology

IV. Duct Geometry Parameterization

The duct entry and exit cross-section shape is shown in Figure 2. Duct design aims to arrive at the duct
geometry, i.e., definition of the duct surface from entry to exit station, S(x, y, z), xen x xex. The X-axis is
along the longitudinal axis of the aircraft and the entry and exit sections are normal to these axes. Duct geometry
is defined by parameterization of the mean flow line (MFL) followed by parameterization to define the cross-
sectional area distribution with reference to the MFL. MFL is the line joining the centroids of cross-section. All
the cross-sections are normal to the MFL. For the purposes of duct parameterization, the duct has been logically
divided into two parts as: S-section (from entry to merger
section), and post-merger section.

A. MFL Design Variables

The MFL of the duct is a space curve. The shape of the
curve is derived from two cubic splines as shown in Figure
3, after imposing conditions that the two cubic spline
segments are C0, C1 and C2 continuous and the slope of the
MFL at the entry is zero and at the merger it is defined by
the merger cross-section. In Figure 3, pen and pm are the
centroid of the entry and merger cross-sections respectively
Figure 2. Entry and Exit cross-sections. and Lm denotes the length of the duct from entry to merger
along x-axis. At x=Lm/2, the y and z coordinates of MFL are
controlled by parameters y and z respectively as,

y(Lm/2) = y(pen) + (y(pm) – y(pen)) * y 0< y<1

z(Lm/2) = z(pen) + (z(pm) – z(pen)) * z 0< z<1 (1)

to yield a representation of the MFL entirely in terms of y and z.

B. Area Design Variables

The cross-section area is varied monotonically from entry to exit sections as:

A(x) = Aen + (Aex – Aen) * (x) 0 < (x) < 1 (2)

3
American Institute of Aeronautics and Astronautics
 y2, z2 pm
y, z y(Lm/2), z(Lm/2)

y1, z1
pen

0 Lm/2 Lm x

Figure 3. Mean Flow line parameterization

The parameter varies along the length of the duct by treating it as a function of the x coordinate of the
centroid at that cross-section. A piecewise cubic variation has been chosen. At one-thirds length and two-thirds
length respectively, the value of is prescribed as 1 and 2, which represent the area design variables. Three
piecewise cubic splines represent variation of (x), xen x xex. (x) at any x is known in terms of 1 and 2.
Area of cross-section at x is obtained using equation 2.
The duct is represented as a set of cross-sections at discrete x. A piecewise linear curve describes cross-
section contour at each discrete x. Number of points describing cross-section is same at all x. The shape of the
cross-section at any x is obtained by the linear interpolation of coordinates of corresponding points between
entry and exit curves. Cross-section so obtained is then scaled to the area at that location. This is followed by the
rotation of the cross-section to keep it normal to the MFL.

Figure 4. Sample duct shape generated with the parameterization

( Y = 0.61, Z = 0.31, 1 = 1.0, 2 = 1.0)

The geometry of the duct in the S-section and the post-merger section is developed separately. At the
merging section, the shape of the cross-section is assumed semi-circular with the area as prescribed by the area
parameters 1 and 2. Since exit section is semi-circular, post-merger section has semi-circular cross-section
with prescribed area variation. As the symmetry plane is fixed in the post-merger section, the location of
centroid of the area is calculated after the area calculation at any cross-section. Sample duct geometry with the
corresponding parameter values is shown in Figure 4.

V. Low Fidelity Analysis

For any given values of duct parameters the MFL, the distribution of cross-section area normal to the MFL,
average diffusion angle ( d-avg), radius of curvature of MFL (R), equivalent duct radius (r) and local wall angle,
( w) can be estimated. A procedure based on these geometry parameters is used in industry for quick design of
ducts that are free from undesirable phenomenon like flow separation, and rapid boundary layer growth due to
excessive diffusion rates or acute bends, both locally and globally7.

d-avg –3 0; 6r– R 0; w – 6 0 (3)

These criteria are based on past experience. For estimating the total pressure loss in the duct, a simple
procedure based on skin friction losses, taking into account duct internal wetted area, average diffusion rate and
Reynolds number is used. These procedures are collectively called low fidelity analyses.

4
American Institute of Aeronautics and Astronautics
A. Total Pressure Recovery
The input parameters required for computing total pressure recovery are duct geometry parameters (cross
sectional area A and perimeter P) along the duct length and operating conditions (Mach number, To, Po at entry).
The rate of change of Mach number (averaged over the cross-section) with respect to axial distance is a function
of area, perimeter and skin friction coefficient, Cf and is given by:

1 1
M 3 1 + ( 1) M 2 P 1 + ( 1) M 2 M
dM 2 2 dA (4)
= Cf
dx 2(1 M 2 ) A (1 M 2 ) A dx

Skin friction coefficient, Cf is calculated as a function of Mach number and Reynolds number based on
distance along duct axis. Equation 4 can thus be integrated using a numerical scheme (the Runge-Kutta-Gill
fourth order method is used in this study). Solution of the equation yields Mach number along the duct length.
The pressure recovery between the exit and entry stations is then computed as given in Equation 5, where the
subscripts en and ex refer to the entry and exit planes.

3
1
1+ ( 1) M 2
ex
Pressure Recovery = Pex = Aen M en 2 (5)
3
Pen Aex M ex 1
1+ ( 1) M 2
en
2

B. Low fidelity Design Problem

Low fidelity design problem is now to find the duct geometry that will maximize the pressure recovery
subject to

d-avg –3 0; 6r– R 0; w – 6 0 (6)

Note that R, r, w are functions of the longitudinal distance x from duct entry to duct exit. Further, for w at
station x is also a function of circumferential location around the section.

C. Low Fidelity Design Process, Results & Discussion8

CFSQP (C - Feasible Sequential Quadratic Programming)9 is used as the optimizer. A parameterization and
geometry generator module is developed at CASDE to compute all geometry related quantities required to
evaluate constraints. The design point is taken as cruise, M = 0.8 at 11000 meter pressure altitude at ISA+15oC.
The entry conditions for the duct corresponding to this are Mentry = 0.52, Po = 34.5 kPa and To = 261.3 K.
CFSQP requires the initial design to be specified, from where the optimizer starts its search. Several initial
designs were investigated to see where the optimization leads. Table 1 summarizes the findings.
Initial designs at rows 1 to 7 all terminate at same optimum design. This optimum design has 1 = 2 = 1.0,
indicating that the duct has its entire diffusion over the first one third of its length (L/3) and has constant area
over its last two-thirds length. Increase in cross sectional area reduces dynamic pressure but increases perimeter.
The former effect is dominant and hence an increase in area improves pressure recovery. 1 = 2 = 1.0 at
optimum indicates that increase in area over the first 1/3 length does not violate any constraint on diffusion.
Initial Design Optimized Design
y z 2 2 Pressure y z 2 2 Pressure Nobj NCon
Recovery Recovery
1 0.9 0.3 0.5 0.5 0.9738 0.61 0.31 1 1 0.9797 267 323
2 0.3 0.3 0.1 0.5 0.9689 0.61 0.31 1 1 0.9797 61 114
3 0.9 0.1 0.1 0.1 0.9652 0.61 0.35 1 1 0.9797 337 345
4 0.9 0.9 0.9 0.1 0.9746 0.61 0.33 1 1 0.9797 166 181
5 0.3 0.3 0.3 0.1 0.9687 0.61 0.35 1 1 0.9797 185 136
6 0.7 0.3 0.9 0.7 0.9781 0.61 0.35 1 1 0.9797 86 190
7 0.7 0.3 0.5 0.9 0.9756 0.61 0.31 1 1 0.9797 152 157
8 0.3 0.1 0.1 0.1 0.9652 0.62 0.36 1 1 0.9796 966 967

Table 1. Optimization results from the low fidelity Analysis

5
American Institute of Aeronautics and Astronautics
Variation in the values of Z for various optimum designs indicate a region over which pressure recovery is very
flat along the direction of z (or insensitivity of pressure recovery to z).
The optimization process has yielded improvement of up to 1.5% (row 3 of Table 1) in pressure recovery in
some cases and only 0.2% (row 6 of Table 1) at the other extreme case. If maximum Specific Excess Power
(SEP) condition is considered instead of cruise, then M = 0.8, sea level and ISA+15o C, with entry conditions
as Mentry = 0.55, Po = 154.5 kPa and To = 342.0 K. If this is used as design input, the actual values of pressure
recovery improve, compared to cruise condition. But levels of improvement realized by the design process are
similar, i.e. 1.4% improvement in pressure recovery. The maximum SEP condition has (T-D)/T approximately
0.7. A 1.4% improvement in pressure recovery will mean 3.2% improvement in SEP, which is significant.

VI. Design Space Reduction

Based on the low fidelity design optimization exercise, it is possible to narrow down the design space to a
subspace of feasible designs to be investigated further for exit distortion. Analysis of constraints showed that all
designs with 1 > 2 and/or z > y are infeasible. Further, best pressure recovery requires 1 and 2 to be large
and have close values and the wall angle and curvature constraints dictate values of y in the range 0.45 to 0.65.
1 and 2 are selected in the range of 0.85 to 1.0 so as keep the pressure recovery above the acceptable value of
0.978. Thus a subspace defined by y in the range 0.45 to 0.65, 2 in the range 0.85 to 1.0, z < y and 1 < 2 is
considered with five levels of each variable in these ranges. From the 625 designs so obtained, 203 were found
to be feasible from low fidelity constraints.

VII. CFD Analysis Of 3D Duct

The low fidelity analysis based optimization explained above has used pressure recovery as the measure of
goodness of the duct. Distortion as a measure of goodness of a duct is brought in now. Distortion is
characterized through the DC60 parameter. DC60 refers to the ratio of difference between the average total
pressure over the worst 60 degrees sector and the overall average total pressure, and the dynamic pressure q.

ro

DC 60 =
{
P (0,360) min P ( , + 60) } where ,
P( r , ) rdrd
(7)
P( , ) =
ri
q ro

rdrd
ri

ri is the internal diameter of the section of interest, ro is the external diameter of the section of interest and -
varies from 0o to 360o.
DC60 can only be captured through a CFD flow analysis using an appropriate turbulence model. This pre-
supposes an ability to generate volume grids for the CFD software and then the post-processing to evaluate DC60
to be passed on to the optimizer. A methodology for achieving automation, especially by using commercial CFD
software is briefly explained below. The commercial CFD software used is FLUENT® and its pre-processor
GAMBIT®. Various steps in the CFD analysis along with the related issues are discussed in the light of the
proposed optimization henceforth.

A. Grid Generation
Grid generation process proves challenging because of the following requirements of the design process and
the CFD solver:
• Parametric grid generation, which renders the straightforward use of third party grid generation
packages difficult.
• Automated grid generation, without user intervention when design parameters are changed.
• Quality of grids in terms of clustering and smoothness consistent with high fidelity CFD analysis
(turbulence modeling) to generate quality solution.
• Need to adhere to specific file formats of CFD packages.

An automated grid generation methodology is developed. As shown in Figure 5, journal files were written
for the automatic surface mesh generation for the entry and exit sections. The required level of clustering is
decided based on y+. The entry and exit sections were generated in GAMBIT® using journal files. The journal
files take the number of grid points and the clustering parameters as their input. A four block transfinite grid is
generated by GAMBIT®. But, these grids are stored by GAMBIT® in the form of a single block unstructured

6
American Institute of Aeronautics and Astronautics
mesh file. The wall and symmetry boundary conditions are also specified in GAMBIT® and are stored in the
mesh files. The complete process can be invoked in a batch mode of operation using a journal file.
In the next step, these mesh files are converted to CFD General Notation System11 (CGNS) file format for
further processing. The CGNS grid files for entry and exit sections are given to a parametric grid generator
(developed at CASDE) to generate volume grid for the duct. The grid generator generates grids specific to duct
geometry parameterization as given by equation 1 & 2. Other inputs for the grid generator are the number of
sections required between the entry and exit sections and the grid parameterization. The grid generator adds no
connectivity information.
In the last step, the connectivity and the boundary condition information is added to the volume grid thus
generated. As the journal files for the entry and exit sections are identical except for the boundary vertices, the
surface grid for the entry and exit sections have one-to-one correspondence. This special property is used to
write the connectivity information of the volume grid. It can be clearly seen here that the complete process is
automated with full control over the number of grid points and clustering.

Duct Parameters
( 1, 2, y, z)
Clustering
Parameters
Generation of structured volume
Generation of entry and exit grid using parameterization
sections using GAMBIT

Mesh file
Conversion of structured grid
Conversion of file format to to unstructured format
CGNS using FLUENT Entry &
Exit sections Unstructured
CGNS file
CFD Solution using FLUENT
Initial guess from
previous solution
CFD Solution

DC60 Objective/Constraints evaluation

Using UDFs (FLUENT)

Figure 5. Block diagram illustrating a complete CFD cycle

B. Integration With FLUENT

The process of grid generation is integrated with the solver through the CGNS files and FLUENT® journal
files. CGNS files contain the grid and the boundary conditions. The journal file specifies input file format, flow
model specification and parameters, boundary conditions, simulation details and the output format. Unix shell
scripts are used to automatically generate journal files depending on the grid size, required convergence levels
and the turbulence model. Parameters specific to the turbulence model and solution procedure are passed on to
these scripts, which generate the journal file accordingly. Journal files for various models are available that
could be executed in batch mode without the graphical user interface.
Some of the above mentioned steps could have been realized more elegantly by using functionalities
provided by FLUENT12. But since these modifications are unlikely to speed up the execution of the process, the
grid generation process as described above is retained.

C. Choice of Turbulence Model

Selection of an appropriate turbulence model is critical for obtaining desired level of accuracy in the
solution. As the final aim of the study is to find the distortion at the exit section of the duct, following aspects of
the flow were of main interest:

7
American Institute of Aeronautics and Astronautics
 • boundary layer development,
• turbulence development, and
• flow separation (if any).
Literature survey of the similar studies in the past shows that various groups: Doyle Knight et al.2,6, Harloff,
Smith et al.4 and Jenkins & Loeffer5 have used Baldwin-Lomax model. Baldwin-Lomax model is an algebraic
turbulence model. Algebraic models are known to require fine-tuning of parameters for every problem. As the
duct design problem is treated in this study as a show case problem for the use of CFD in design optimization,
this is not acceptable. The aim of the study is to use a commercial analysis module and integrate it with the
overall optimization framework without any problem specific fine-tuning. This is the first drawback of Baldwin-
Lomax model.
Besides this, the algebraic models are not very reliable for complex three-dimensional flows with boundary
layer separation. It should be noted here that all the studies mentioned above have been carried out on S-ducts
with a circular cross-section throughout the duct whereas the present study has a kidney shaped entry section
that merges into a semi-circular cross-section at the exit. Also, additional flow complications can be expected
from the merging of the two ducts in the merging part of the duct. Because of the above considerations, the
Baldwin-Lomax model is not considered adequate for the present optimization study. Another study conducted
by Devaki Ravi Kumar and Sujata Bandyopadhyay13 on a similar S-duct problem has used the k- realizable
model14 and Reynolds Stress Model (RSM). The study also reports good agreement with the NASA validation
case15. Considering the availability of time for the study and the fact that FLUENT® is being used for the present
study, it is decided that the k- realizable turbulence model will be used for turbulence modeling.

D. Grid Sensitivity Analysis

A grid sensitivity analysis is performed to determine the minimum grid size required to capture the distortion
coefficient appropriately. This is essential to strike a proper balance between the conflicting requirements of
lower simulation times and the desired accuracy levels. A literature survey of the grid sizes employed by similar
studies was performed. Kumar et al.13 have reported a grid size of around 440,000 cells whereas Jenkins et al.5
have used a grid size of 450,000 cells. Knight et al.16 have used nearly 300,000 grid points in an optimization of
a S-shaped duct having a bump.
The solution is obtained at four levels of grid sizes by successively increasing the number of sections along

DC_60 Variation
6.9
Series1
Duct-1
Duct-1
6.85 Duct-2
Series2
Distortion Coefficient

6.8

6.75

6.7

6.65

6.6
0 10 20 30 40 50 60 70 80
Number of Sections

Figure 6. Grid sensitivity analysis for DC60 parameter

the length of the duct. The clustering of the grid points near wall is kept constant because it is found adequate
from the observed y+. The aim of the grid sensitivity study was to ensure that the shape of the duct is captured
with adequate accuracy along the duct length. Figure 6 shows the variation of the DC60 parameter plotted against
the number of sections along the length of the duct for two ducts that are close with respect to their parameters.
There are two observations to be made from this. First, the DC60 parameter is converging as the grid size is
increased. It is also observed that though the duct with 45 sections is adequate, the grid with 60 sections will be
employed in the optimization process to ensure proper accuracy. Another possibility is to refine the grid
successively during the optimization process as we progress. Secondly, it may be observed that the results for

8
American Institute of Aeronautics and Astronautics
two different ducts show similar trends in the DC-60 variation with the grid size further increasing our
confidence on the solution scheme.

E. Initial Guess
The strategy of reducing the computation time by giving a better initial guess is implemented here. Typically
in a design process, CFD solutions are obtained for various ducts that differ slightly from each other. Hence, a
technique in which the converged solution of a duct is considered as an initial starting point for the flow analysis
of a neighboring duct, with incremental change in the parameters, is tested and the gains were quantified. The
solution of the first duct is used to start the iterations of the second duct. The number of iterations for the
solution process of the second duct is compared with and without this method. The results of the study are as
follows:
First Duct Parameters: (0.61, 0.31, 1.0, 1.0) Second Duct Parameters: (0.10, 0.31, 1.0, 1.0)

3-decade-fall 6-decade-fall
No. of iterations for arbitrary initial guess 4996 9462
No. of iterations with first duct results as initial guess 1493 6588
% time savings 70.12 30.38

Table 2. Comparative study of arbitrary and close initial guess

It should be noted that in case of the 3-decade-fall, the second duct solution is started with a 3-decade-fall
solution of the first duct and in the 6-decade-fall case; the second duct solution is started with a 6-decade-fall
solution of the first duct. This methodology is implemented using different sets of grid and solution files of
FLUENT®. Substantial timesaving is achieved using this method requiring little efforts from the part of the
designers. It is also observed here that the gains are substantially less for 6-decade-fall because of the
exponentially decreasing behavior of the residues.

F. Post Processing
DC60 parameter for the exit section at various iteration levels is evaluated using User Defined Function
(UDF) provided by FLUENT®. A program in C is written using the standard library functions provided by
FLUENT® for accessing the flow variables. The distortion coefficient is calculated from the flow variables and
is stored in an output file, which may be accessed by the optimizer. The iterations may also be stopped when the
convergence of the DC60 parameter is achieved. This control is achieved by using the scheme files in
FLUENT®.
The present study has successfully demonstrated some of methods for achieving a single CFD run (from grid
generation to post processing) without any user intervention. This has been achieved by automating each of the
steps followed by the integration of various steps. A simple script may be generated in any scripting language to
call all these processes sequentially with all the appropriate parameters.
The overall process of the CFD analysis is captured in Figure 5. The complete process is automated
requiring no user intervention and can be run in a batch mode. The DC60 parameter as obtained at the end may
be directly fed to the optimizer or for the response surface generation. A maximum y+ of 62.6 is observed in the
near wall cells all along the duct confirming the adequacy of the grid clustering and ensuring proper resolution
of boundary layer.

VIII. Surrogate Modeling

Surrogate models can be non-interpolating fits like RSM technique or interpolating fits. RSM is primarily
used for experiments where independent random error is associated with every experiment. As we are
performing deterministic computer experiments, these fits are not appropriate. Hence we have used Design and
Analysis of Experiments (DACE)17 for fitting an interpolating surface. DACE uses Kriging model.
A literature survey on sampling strategies for these models revealed that space-filling designs are
required18,19. Various space filling sampling strategies like latin hypercube, stratified sampling20 and LP-
sampling21 are tested. Only 16 initial points were proposed for building the surrogate model because of the time
intensiveness of the CFD runs. It was found that these strategies do not give good uniformity characteristics
because of the irregular nature of the design space obtained after applying constraints from LFA.
Hence, a different method for constructing the surrogate model is adopted. Only the corner points (12 points)
of the feasible design space are provided to the DACE model initially and successive points are added at the
points with maximum predicted error by the DACE model. Surrogate model is considered accurate once this

9
American Institute of Aeronautics and Astronautics
 maximum predicted error is reduced to acceptable levels. Successive points are added till the relative error
(error/value) was brought below 5% everywhere in the design space. 21 points (as tabulated in Table 3) were
required to achieve this.
Two design points were added to the surrogate model at each stage because of the facility of performing two
CFD runs simultaneously. A criterion based on the maximum error and the spatial distance between the points
was used to select these two points. The points with maximum error that are adequately separated from each
other in the design space were selected at each stage.

Sr. y z 1 2 DC60 Pressure

No. Recovery
1 0.45 0 0.85 0.85 4.4728 0.9856
2 0.65 0 0.85 0.85 1.7051 0.9850
3 0.45 0.45 0.85 0.85 10.8753 0.9867
4 0.45 0 0.85 1 4.5688 0.9859
5 0.65 0.65 0.85 0.85 11.2684 0.9858
6 0.45 0.45 0.85 1 11.2114 0.9869
7 0.45 0 1 1 4.5617 0.9864
8 0.65 0 0.85 1 1.8372 0.9859
9 0.65 0.65 0.85 1 11.3537 0.9867
10 0.45 0.45 1 1 11.1159 0.9872
11 0.65 0 1 1 1.7547 0.9856
12 0.65 0.65 1 1 11.9412 0.9863
13 0.65 0.21667 1 1 4.4704 0.9865
14 0.65 0.21667 0.85 1 4.1753 0.9861
15 0.65 0.21667 0.85 1 4.3572 0.9859
16 0.60556 0.33642 1 1 7.3432 0.9868
17 0.60556 0.33642 0.85 0.85 7.1980 0.9863
18 0.65 0.50556 0.9241 0.9333 9.4487 0.9864
19 0.45 1 0.9241 0.9333 5.8021 0.9866
20 0.5389 0 0.9167 0.9167 2.6260 0.9857
21 0.5611 0 0.9333 0.9333 2.3335 0.9857
Table 3. Points used for constructing the DACE Model
IX. High Fidelity Optimization
The duct is optimized for DC60 coefficient using the DACE model
HFA Optimum Duct constructed above. Pressure recovery was found to be almost constant in the new
design space obtained from the low fidelity analysis. The minimum value of
pressure recovery was more than 98.57% throughout the reduced design space.
Hence, a single criteria optimization using DC60 as the objective function is
adopted. The DACE model is used as the surrogate for DC60 with CFSQP for the
optimization. Several runs with different initial points were tried to probe for
multiple local minima. Only one minimum was detected in the entire design
space with parameters as (0.65, 0, 0.9027, 0.9027) with DC60 parameter of
LFA Optimum Duct 1.6857. A HFA at this point gave a DC60 parameter of 1.7032 and pressure
recovery of 98.52%.
It should be noted here that the optimum duct from LFA had a DC60
parameter of 6.19. Also, pressure recovery has been consistently over predicted
by HFA as compared to LFA. Figure 7 shows the total pressure contours for the
exit section of the final optimized duct and the optimum duct from LFA. As can
be seen from the plots, the total pressure loss has become more symmetric in the
radial direction resulting in a lower DC60 in the HFA optimum duct.
Figure 7. Total pressure A Linux based cluster of 8 nodes (each node with PENTIUM IV 1.6GHz
profiles on the exit sections processor with 1 GB RAM) was used for this study. A single run of FLUENT
using 4 nodes took around 30 hours for each run. Overall design took 630 hours
of CPU time.

10
American Institute of Aeronautics and Astronautics
 X. Conclusion
A complete end-to-end duct optimization methodology has been presented. Considerable time saving has been
achieved through the use of mixed-fidelity optimization and the surrogate model. Special attention was given to
develop automation techniques that do not need user intervention.

References
1
Seddon, J., and Goldsmith, E. L., “Intake Aerodynamics,” Collins Professional and Technical Books, 1985.
2
Knight, D., Zhang, W., and Smith, D., “Automated Design of a three-dimensional subsonic diffuser,” Propulsion and
power,Vol. 16, No.6, Nov.-Dec. 2000.
3
Jameson, A., and Nadarajah, S., “A comparison of the continuous and discrete adjoint approach to automatic
aerodynamic optimization,” AIAA-2000-0667.
4
Harloff, G. J., Smith, C. F., Bruns, J. E., and DeBonis, J. R., “Navier-Stokes Analysis of three dimensional S-ducts,”
Journal of Aircraft, Vol. 30, No. 4, July-Aug. 1993.
5
Jenkins, R. C., and Loeffler Jr., A. L., “Modeling of subsonic flow through a Compact Offset Inlet diffuser,” AIAA
Journal, Vol. 29, No. 3, March 1991.
6
Knight, D., “Automated optimal design of supersonic and subsonic diffusers using CFD,” European Congress on
Computational Methods in Applied Sciences and Engineering, ECCOMAS 2000.
7
Jolly, R., Sudhakar, K., and Damodran, K.A., “A Synthesis of Procedures for Selecting the Geometric Features of an
Air Intake for M=0 to 1.8,” Journal of Aeronautical Society of India, Vol.38 No.4, November 1986.
8
Isaacs, A., Mujumdar, P M, Sudhakar, K, Jolly, R., and Pai, TG., "Low Fidelity Analysis Based Optimization of 3D-
Duct using MDO-Framework," International Conference & Instructional Workshop on Industrial Mathematics, IIT
Bombay, December 7-9, 2002.
9
Zhou, J. L., Tits, A. L. and Lawrence, C. T., “User's Guide for FFSQP Version 3.7: A Fortran Code for Solving
Optimization Programs, Possibly Minimax, with General Inequality Constraints and Linear Equality Constraints, Generating
Feasible Iterates,” Institute for Systems Research, Univ. of Maryland, Technical Report SRC-TR-92-1075, College Park,
MD, 1997. (http://www.aemdesign.com/)
10
Montgomery, D.C., "Design and Analysis of Experiments," 5th ed., John Wiley and Sons Inc., New York, 2001.
11
Poirier, D. , Allmaras, S. R., McCarthy, D. R., Smith, M. F., and Enomoto, F. Y., "The CGNS System," AIAA Paper
98-3007.
12
Discussions with FLUENT India Pvt. Ltd., November 2003.
13
Kumar, D. R., and Bandyopadhyay, S., “Computation of three dimensional compressible turbulent flow through a non-
diffusing S-duct,” 5th Annual CFD Symposium, CFD Division, Aeronautical Society of India, Bangalore, 9-10 August, 2002.
14
Shih, T. H., Liou, W. W., Shabbir, A., Yang, Z., and Zhu, J., "A New k- eddy viscosity model for high Reynolds
number turbulent flows," Computers & Fluids, Vol. 24, No. 3, pp. 227-238.
15
Vakili, A., and Wu, J. M., “Comparison of Experimental and Computational Compressible Flow in a S-duct,” AIAA
nd
22 Aerospace Sciences Meeting, January 9-12, 1984.
16
Lefantzi, S., and Knight, D. D., “Automated design optimization of a three-dimensional S-shaped subsonic diffuser,” J.
of Propulsion & Power, Vol. 18, No. 4, July-August 2002.
17
Schonlau, M., “Computer Experiments and Global Optimization,” Doctoral Dissertation, University of Waterloo,
Canada, 1997.
18
Barton, R. R., “Metamodelling: A State of the Art Review,” Proceedings of the 194 Winter Simulation Conference,
edited by J. D. Tew, S. Manivannan, D. A. Sadowski and A. F. Seila, 1994, pp. 237-244.
19
Jones, D. R., Schonlau, Matthias, and Welch, W. J., “Efficient Global Optimization of Expensive Black-Box
Functions,” Kluwer Academic Publishers, Netherlands.
20
Conover, W. J., Mckay, M. D., and Beckman, R., J., “A comparison of three methods for selecting values of input
variables in the analysis of output from a computer code,” Technometrics, vol. 21, no. 2, May 1979.
21
Statnikov, R. B., and Matusov, J. B., “Multicriteria Analysis in Engineering,” Kluwer Academic Publishers, 2002.

11
American Institute of Aeronautics and Astronautics