Title: Evolutionary Optimization Method for Thermal Protection System Design
1Evolutionary Optimization Methodfor Thermal
Protection System Design
Department of Mechanical and Materials
Engineering Wright State University, Dayton, OH,
45435
Air Vehicles Directorate WPAFB, OH, 45433
2Thermal Protection System (TPS)
- Responsible for protecting the spacecrafts
components from melting due to high re-entry
temperatures. - Key technology that enables a spacecraft to be
lightweight, fully reusable, and easily
maintained. - Consists of different types of materials that
are distributed all over the spacecraft, such as
felt blankets, ceramic tiles, carbon-carbon
leading edge, and metallic TPS.
3Metallic TPS
- The metallic TPS makes maintenance and
replacement easier
- Its inherent ductility and design flexibility
offers the potential for a more robust system
with lower maintenance costs than competing TPS
systems
lt ARMOR TPS model gt
lt X-33 metallic TPS panel gt
4Metallic TPS
PLATE (Unremovable region)
FRAME
z
x, y
SUPPORT
CHALLENGES IN DESIGN
- Weight is a key disadvantage in its use
- Thermal and acoustic loading conditions tend to
be at odds with one another
It is difficult to develop an optimum TPS model
5Metallic TPS
REMEDIES
- The support height is increased to reduce the
heat transfer to the fuselage by inserting more
insulators.
- A frame for thin plate is attached to prevent
flutter due to acoustic loading.
6Structural Optimization Method
Topology design is for determining the
distribution pattern of material and void.
This 0-1 discrete problem is transferred into a
continuous one
- Homogenization method (Bendsoe and Kikuchi, 1988)
The design region is assumed consisting of the
porous microstructure and material density
distribution is optimized by taking size and
direction of the hole of the micro-structure as
design variables.
- Density-based method (Yang and Chuang, 1999)
This method heuristically designs the material
properties such as Youngs modulus and density
for each finite element directly to find optimal
material distributions.
- Cellular automation generation method (Inou,
1994)
Evolution of the organism is considered cell
automaton based on the local field rule and it is
the evolution makes the structure form that
adapted to the dynamics environment.
7Continuous approaches to topology optimization
Hard to derive sensitivity functions
(Mathematically difficult!!)
Make Final design to be different by a penalty
rule for gray regions
Hard to apply a sensitivity for stress
(Usually, strain energy is used)
The evolutionary structural optimization (ESO)
method
or
The bidirectional evolutionary structural
optimization (BESO) method
Discontinuous approach, which deals with only
0-1 distribution (No penalty function)
very simple concept !!
Stress, as well as strain energy can be easily
utilized
8Evolutionary Structural Optimization (ESO)
Methodfor Thermal Protection System Design
9Evolutionary Structural Optimization (ESO) Method
- Based on a simple concept that the residual
structure evolves toward an optimum by gradually
removing inefficient material.
STATIC OPTIMIZATION PROBLEM
low stress values low strain energy values
A number of elements with
are removed from structure
10CONVENTIONAL ESO FOR EIGENVALUE PROBLEM
- General eigenvalue problem
(1-1)
(1-2)
- Change of the i th frequency
(1-3)
(1-4)
(1-5)
11CONVENTIONAL ESO FOR EIGENVALUE PROBLEM
Using a dynamic control parameter based on the
Rayleigh quotient,
(1-6)
A number of elements with high positive values
are eliminated to increase the natural frequency
of interest.
12Eliminating a large number of elements from a
structure through many iterative steps
Results in very weak modal stiffness !
Drastic alteration in the natural frequency of
interest !
13Derivation of New Dynamic Control Parameter
The response of the equation of motion (1) is
described in modal coordinates as
(1)
(2)
where ,
,
The displacement of each nodal point is computed
by implementing the concept of virtual static
displacement for each mode shape.
By assuming 0,
(3)
14The displacement of each nodal point is computed
by implementing the concept of virtual static
displacement for each mode shape.
By assuming ?0,
(3)
modal displacement (generalized displacement)
generated by external force F .
The natural mode can be treated as a
response (displacement) by applying the external
force whose modal displacement is 1 for the r th
mode and 0 for the other modes.
When Fr is the external force satisfying this
condition,
(4)
15- Force magnitude for each natural mode is
different from others
- A higher natural mode requires a larger external
force due to the complexity of the mode shape
The magnitude of Fr is scaled to be that of the
external force of the interested natural mode.
(5)
The scaled external force vector can estimate the
displacement of the r th natural mode by the
force magnitude of the interested natural mode
(6)
16Dynamic Control Parameter by von Mises Stress
(7)
Eq. (7) considers not only the i th natural
frequency, but also its neighboring natural
frequencies by choosing maximum stress of r th
natural mode
If , , as well as
are selected because
approaches to that of the natural mode of
interest.
A number of elements with the smallest stresses,
that is, the most inefficient elements, are
removed from the structure.
17Cantilever Structure
- 2 of the total number are eliminated in each
iteration.
Weighting Factor Change
Evolutionary History of Frequencies
18Comparison between conventional ESO and new ESO
19TPS Design with Heat Transfer Problem
subject to
- Using weighting objectives method
(8)
Four corners are simply supported
Initial Metallic TPS Model
20Optimization Challenge for TPS Model with a Thin
Plate
Plate (Unremovable)
- A local mode can be observed at the plate region
due to its thinness
Frame
Z
Support
X
To avoid the local mode,
Fixed region
Modified
- Dense-meshed structure
- Few elements removal in one iteration
Too much computational cost !
21TPS Design Process
Design optimization of the support and frame
regions are conducted separately
TPS initial model with dense mesh
TPS support design by topology optimization
Modified TPS initial model based on the topology
result
TPS support design by shape optimization
TPS frame design
Optimum TPS model
22Research Approach
Research Approach
- The support region is designed by topology
optimization using ESO algorithm
Initial structure modified from topology result
23TPS Support Design
- The initial TPS model is set up as a full-meshed
structure.
- Plate and frame region are set as unremovable.
- Topology optimization is applied to the
conventional ESO and the new ESO with the
proposed control parameter.
Unremovable
24Change of the Fundamental Natural Frequency
- Fundamental natural frequency is improved up to
certain percentage of material removal.
- As the removed volume percentage increases, the
conventional method decreases the fundamental
natural frequency very quickly because there are
no direct considerations of mode-switching
phenomenon and modal stiffness.
- The proposed method keeps the fundamental
natural frequency higher.
25Change in Maximum Thermal Stress
- In the conventional method, the maximum thermal
stress increases at an early stage because the
control parameter doesnt consider thermal stress.
- As the removed volume percentage increases, the
maximum thermal stress decreases because the heat
transfer to the bottom side is decreased by
eliminating elements. - ( Elements that connect between the plate and
the support region )
- The proposed method restrains the increase in
maximum thermal stress from increasing at an
early stage, as well as at increased volume
reduction.
26Resultant Models with the Fundamental Frequency
at 900 Hz
Conventional Method
Proposed Method
27TPS Support Redesign by Shape Optimization
- The initial TPS model with hollow cube is chosen
from the previous result.
- Shape optimization (Called Nibbling ESO) is
applied.
- Plate and frame region are set as unremovable.
- In dynamic analysis, additional stiffness kX,
kY, kZ 0, 0, 108 (N/m) are set up at four
edges of the plate.
The local mode will have a higher natural
frequency when compared with the fundamental
natural frequency of the supports first bending
mode.
28Modified TPS Support Model
- Fundamental frequency at 77.7 901.5 Hz
- Maximum stress at 77.7 0.236 GPa
- When the additional stiffness at the plate edge
is removed from the structure, the fundamental
frequency becomes 871.9 Hz due to the increase in
modal mass.
Fundamental frequency may be increased by
reducing the frame weight.
29TPS Frame Design
- Shape optimization is applied by eliminating
elements from the bottom surface of the frame
region.
- Plate and support regions are set up as
unremovable regions.
- No thermal stress analysis is used.
- TPS model is designed to be lightweight with the
fundamental frequency greater than 900 Hz.
30Evolutionary Process of the TPS Frame Model
- Up to the 41st iteration, fundamental frequency
increases.
- In the end iterations, the frequency drastically
decreases.
- Modified TPS model satisfies the frequency
constraint at the 45th iteration.
- A transient heat transfer analysis is conducted
on the TPS model of 45th iteration.
View from the Bottom Side of the Frame
31Optimized TPS Model
- Fundamental natural frequency 919.8 Hz (gt900
Hz) - Maximum thermal stress0.228 GPa (lt0.3 GPa)
- TPS mass 76.50 kg (cf. For the full meshed
structure, 470 kg)
This model reduced its weight until 84 in
comparison with the full-meshed structure
32Summary
1. A multi-objective optimization problem for
thermal stress and fundamental natural frequency
was conducted to make a lightweight TPS model by
using Evolutionary Structural Optimization (ESO)
algorithm.
2. New control parameter based on static analysis
was proposed for the TPS design concerned with
dynamic analysis.
3. An efficient way to obtain a metallic TPS
model was shown by designing the support and the
frame region separately using the ESO method with
the proposed control parameter.
33Bidirectional ESO Methodfor Thermal Protection
System Design
34Conventional Bidirectional Evolutionary
Structural Optimization Method
applies element addition, as well as element
removal (e.g. New elements are attached around
the elements with overly stress to reduce
localized high stress regions in static
optimization problem)
starts from a simple structure satisfying
boundary conditions, not a full-meshed one
uses control parameters based on Rayleigh
quotient for both element addition and removal in
eigenvalue optimization problem
35Control Parameter Based on Rayleigh Quotient
IN THE ADDITION PROCESS
(A)
This equation expresses the change of the i-th
eigenvalue due to attaching the l-th virtual
element to the original structure.
Some elements whose are most
positive among all virtual elements are
converted into real elements !
Eq. (A) is also applied to following situation
- The i-th natural frequency increases by adding
the l-th element of 0 - The i-th natural frequency decreases by removing
the l-th element of 0
There is no drastic change of natural frequency
of interest even if mode-switching occurs in
iterative steps because there is no sudden
changes of stiffness of structure.
36Control Parameter Based on Rayleigh Quotient
IN THE REMOVAL PROCESS
(B)
This equation expresses the change of the i-th
eigenvalue due to removing the l-th element from
the original structure.
- Only the changes in natural frequencies are
evaluated. Nothing is considered about the
stiffness of any part of the structure. - The sensitivities of higher-order natural modes
are incorrect due to the inaccuracy of the
natural frequency calculations. - In the case of mode-tracking, comparison between
the current and previous modes is impossible if a
new mode appears at the current step, or if a
previous mode disappears at the current step. - The change of the natural mode of interest due to
the mode-switching phenomenon induces a
significant alteration of the structural
stiffness in the next evolutionary process.
1 and 2 are the main challenges for the
evolutionary method with the fundamental natural
frequency optimization.
37Challenge in Conventional BESO Method
ADDITION PROCESS
REMOVAL PROCESS
(A)
(B)
No significant decrease in structural stiffness
Significant decrease in structural stiffness due
to mode-switching
Sudden drop of the natural frequency of interest
Unstable eigenvalue optimization process
- Static control parameter using strain energy or
von Mises stress is applied to consider the modal
stiffness
- The static control parameter is expanded to
include the neighboring natural modes, as well as
the natural mode of interest
38Static Control Parameter Using Strain Energy
Modified natural mode
(6)
Strain energy
Static control parameter in the removal process
39Proposed Bidirectional Evolutionary Structural
Optimization Process
- Virtual element is newly applied.
- In element removal, structural stiffness is only
considered to avoid sudden drop of the stiffness.
- In element removal, new control parameter is
applied to consider mode-switching phenomenon.
40Algorithm
The objective of this method is
to shift the fundamental natural frequency of a
three-dimensional structure to a target
frequency as closely as possible.
Because of starting from a simply small structure,
The number of FE to be added gt The number of FE
to be removed
Stage 1
the growth stage to grow the structure until
raising the fundamental natural frequency to a
little bit higher than the target frequency (
upper limit).
( upper limit target frequency a, where a
1.051.50)
The number of FE to be removed gt The number of FE
to be added
Stage 2
the weight reduction stage to control the
fundamental natural frequency until it reaches to
lower limit ( target freq ß, where ß1.001.02).
Stage 3
The number of FE to be removed The number of FE
to be added
the alternation stage to change the positions
of the elements
41Algorithm
The evolutionary iterative process is continued
until the i th natural frequency ( fundamental
natural frequency) converges after the
application of the three kinds of stages.
The number of finite elements, which are added
and removed in an iterative process, is set
within 12 of the number of elements of the
structure .
42A Bridge Model
1st natural freq. 2349 Hz
Upper limit frequency 4250 Hz (target
frequency 1.05)
1st mode (2349Hz)
2nd mode (2506Hz)
3rd mode (3658Hz)
43A Bridge Model
V shape
Z
X
1st mode (2349Hz)
2nd mode (2506Hz)
Reverse-V shape
Z
Y
3rd mode (3658Hz)
44Availability of Proposed Method
ltlt ADDITION gtgt
Control parameter using Rayleigh quotient
ltlt REMOVAL gtgt
ltlt REMOVAL gtgt
Control parameter using strain energy
Control parameter using Rayleigh quotient
METHOD 1 (Conventional method)
METHOD 2 (Proposed method)
COMPARISON
45A Bridge Model
In the case of METHOD 1,
In the case of METHOD 2,
46A Bridge Model
1st mode (2349Hz)
2nd mode (2506Hz)
3rd mode (3658Hz)
1st mode (4028Hz)
2nd mode (4176Hz)
3rd mode (4299Hz)
METHOD 2 can select appropriate finite elements
to be removed, and can evolve the bridge model
to the optimum design stably under the condition
that Mode-switching happens in.
47A Thermal Protection System Model
Adaptable, Robust, Metallic, Operable, Reusable
(ARMOR) TPS panel
X-33's Innovative Metallic Thermal Shield
1st natural freq. 104.8 Hz
ltlt Inconel 693 alloy gtgt
E 196 GPa Density 7770 kg/m3 Poisson 0.32
Target freq. 1000 Hz
Initial model for designing TPS
48A Thermal Protection System Model
At the 100th iteration, the fundamental natural
frequency is 1011.2 Hz
Although "mode-switching" occurs irregularly in
the iterative processes between the bending mode
in X-axis and the bending mode of Y-axis, which
are mutually orthogonal directions, the evolution
process is executed very smoothly.
Evolutionary history of the first three natural
frequencies
49A Thermal Protection System Model
Initial model
Modified model
ltThe fundamental natural mode for the initial
model and the modified model gt
50Summary
1. Two kinds of control parameters are introduced
in the bidirectional evolutionary structural
optimization.
2. It is shown that the static control parameter
can make a natural frequency of interest
convergence stable or improved by removing
elements with low values.
3. Static control parameter is modified to
consider adjacent natural modes, as well as the
targeted mode to alleviate the problem of the
mode-switching phenomenon.
4. Using the proposed control parameter and new
addition process, two representative models are
analyzed.