Issues in Optimization
Jaroslaw Sobieski
NASA Langley Research Center
Hampton Virginia
NASA Langley Research Center
LaRC/SMC/ACMB
Copyright NASA, Jaroslaw Sobieski, 2003
How to know whether
optimization is needed
How to recognize that
the problem at hand needs
optimization.
General Rule of the Thumb:
there must be at least two opposing trends
as functions of a design variable
Analysis
x
f1
f1
f2
f2
f1
f2
x
Power Line Cable
tout cable
slack cable
h
Length(h)
Given:
A(h)
Ice load
Volume(h)
self-weight small
h/span small
A
L
V
min
tout
h
slack
Wing Thin-Walled Box
Lift
Top cover panels
are compressed
b
thickness t
Buckling stress
= f(t/b)
2
b
many
Rib total weight
min
few
Cover weight
Wing box weight
ribs ribs
Multistage
Rocket
drop when
burned
number
segment
junctions
weight
rocket weight
2 3
min
fu
el
fuel weight
of segments
More segments (stages) = less
weight to carry up = less fuel
More segments = more junctions =
more weight to carry up
Typical optimum: 2 to 4.
Saturn V
more weight
fore
nacelle
aft
Under-wing Nacelle
Placement
shock wave
drag
nacelle
wing
underside
Inlet ahead of wing max. depth =
shock wave impinges on forward
slope = drag
Nacelle moved aft = landing gear
l
drag
weight
Range
max
moves with it = larger tail (or
onger body to rotate for take-off =
National Taxation
tax paid on $ earned
revenue collected
max
incentive to work
0 % average
100 %
tax rate
More tax/last $ = less reason to strive to earn
More tax/$ = more $ collected per unit of economic activity
National Taxation
revenue collected
0 %
max
tax paid on $ earned
incentive to work
average
100 %
tax rate
More tax/last $ = less reason to strive to earn
More tax/$ = more $ collected per unit of economic activity
What to do:
If we are left of max = increase taxes
If we are right of max = cut taxes
Nothing to Optimize
Rod P Newton
A cm
2
Monotonic trend
No counter-trend
σ
σ allowable
Nothing to optimize
N/cm
2
A
Various types of design optima
Design Definition: Sharp vs.
constraints - 0 contours Shallow
- bad side of
1
2
constraints - 0 contours
1
2
band
point
X
X
Constraint
descent
Objective
Near-orthogonal intersection
defines a design point
Tangential definition identifies
a band of of designs
X
Multiobjective Optimization
trade-
both
Q = 1/(quality &
f
1
off
both
performance &
f
2
comfort)
$
1
4
$
4
pareto-frontier
2
3
3
2
design & manufacturing
sophistication
1
Q
pareto-optimum
V&W R&R
A Few Pareto-Optimization
Techniques
Reduce to a single objective: F = Σ w
i
f
ii
where w s are judgmental weighting factors
Optimize for f
1
; Get f*
1;
;
Set a floor f
1
>= f*
i
; Optimize for f
2
; get f
2
;
Keep floor f
1
, add floor f
2
; Optimize for f
3
;
Repeat in this pattern to exhaust all f s;
The order of f s matters and is judgmental
Optimize for each f independently; Get n optimal designs;
i
Find a compromise design equidistant from all the above.
Pareto-optimization intrinsically depends on judgmental
preferences
Imparting Attributes by
Optimization
Changing w
i
in
F = Σ
i
w
i
f
i
modifies the design within broad range
Example: Two objectives
setting w
1
= 1; w
2
= 0 produces design whose F = f
1
setting w
1
= 0; w
2
= 1 produces design whose F = f
2
setting w
1
= 0.5; w
2
= 0.5 produces design whose
F is in between.
Using w as control, optimization serves as a tool
i
to steer the design toward a desired behavior or
having pre-determined, desired attributes.
Optimum: Global vs. Local
X2
Why the problem:
Objective
contours
Nonconvex
objective or
constraint
constraints
(wiggly contours)
X1
L
G
resonance
d Spring k N/cm
Disjoint design
mass
space
d
P
P = p cos (ωt)
k
Local information, e.g., derivatives, does not distinguish
local from global optima - the Grand Unsolved Problem in Analysis
Use a multiprocessor computer
Start from many initial designs
Execute multipath
optimization
Increase probability of locating
global minimum
Probability, no certainty
Multiprocessor computing =
analyze many in time of one =
new situation = can do what could
not be done before.
What to do about it
A shotgun approach:
F
Start
M1
Opt.
Tunnel
M2<M1
X
Tunneling algorithm
finds a better minimum
A shotgun approach:
Use a multiprocessor computer
Start from many initial designs
Execute multipath
optimization
Increase probability of locating
global minimum
Probability, no certainty
Multiprocessor computing =
analyze many in time of one =
new situation = can do what could
not be done before.
What to do about it
F
Start
M1
Opt.
Tunnel
shotgun
Multiprocessor
computer
M2<M1
X
Tunneling algorithm
finds a better minimum
What to do about it
F
Start
M1
Opt.
Tunnel
A shotgun approach:
Use a multiprocessor computer
Start from many initial designs
Execute multipath
M2<M1
optimization
X
Increase probability of locating
global minimum
Tunneling algorithm
Probability, no certainty
finds a better minimum
Multiprocessor computing =
analyze many in time of one =
new situation = can do what could
not be done before.
Using Optimization
to Impart Desired Attributes
Larger scale example: EDOF = 11400;
Des. Var. = 126; Constraints = 24048;
Built-up, trapezoidal, slender transport aircraft wing
Design variables: thicknesses of sheet metal, rod cross-sectional
areas, inner volume (constant span and chord/depth ratio
Constraints: equivalent stress and tip displacement
Two loading cases: horizontal, 1 g flight
with engine weight relief, and landing.
n
p
a
s
f
t
7
0
Four attributes:
structural mass
1st bending frequency
tip rotation
internal volume
Case : F = w
1
(M/M
0
) + w
2
(Rotat/Rotat
0
)
Normalized
Mass M/M
0
variation:
52 % to
180 %
weight factor
Mass
weight factor
Broad
Rotation
Rotat = wingtip twist angle
Optimization Crossing the
Traditional Walls of Separation
Optimization Across
Conventional Barriers
data
Vehicle design
Fabrication
Focus on vehicle physics
Focus on manufacturing
and variables directly
related to it
process and its variables
E.g, range;
E.g., cost;
wing aspect ratio
riveting head speed
Two Loosely Connected Optimizations
Seek design variables Seek process variables
to maximize performance to reduce the fabrication cost.
under constraints of:
Physics
Cost
Manufacturing difficulty
The return on investment (ROI) is a unifying factor
ROI = f(Performance, Cost of Fabrication)
Integrated Optimization
Required: Sensitivity analysis on both sides
?Range/ ?(AspectRatio) ?Cost/ ?(Rivet head speed)
?(Rivet head speed)/ ?(AspectRatio)
ROI = f(Range, Cost of Fabrication)
?ROI/ ?AspectRatio = ?ROI/ ?Cost ?Cost/ ?(Rivet h.s.) ?(Rivet h.s)/ ?(AspectRatio) +
+ ?(ROI)/?Range ?Range/?(AspectRatio)
Integrated Optimization Design < --- > Fabrication
Given the derivatives on both sides
Design
Fabrication
Unified optimization may be constructed to seek
vehicle design variable, e.g., AspectRatio, for
AR
maximum ROI incorporating AR effect on Range and on
Opt.
fabrication cost.
ROI
ROI
Range
Cost
Range; Cost
Optimization Applied to Complex
Multidisciplinary Systems
Multidisciplinary Optimization
MDO
Coupling
Decomposition
What to optimize for at the discipline level
Approximations
Sensitivity
Wing drag and weight both influence the flight range R.
R is the system objective
direct
i
P
P
Displ
a
l
i
that affect drag
Displ
Wing - structure
Wing - aerodynamics
Loads
acements
a = sweep angle
Structure influences R by
y by weight
ndirectly by st ffness that
affect displacements
Loads &
acements
must be consistent
R = (k/Drag) LOG [( W
o
+ W
s
+ W
f
)/ (W
o
+ W
s
)]
Dilemma: What to optimize the structure for? Lightness?
Displacements = 1/Stiffness?
An optimal mix of the two?
Trade-off between opposing objectives
of lightness and stiffness
Weight
Displacement
Weight
Displacement ~ 1/Stiffness
Thickness
limited by
stress
Wing cover sheet thickness
Lightness Stiffness
What to optimize for?
Answer: minimum of f = w1 Weight + w2 Displacement
vary w1, w2 to generate a population of wings
of diverse Weight/Displacement ratios
Let system choose w1, w2.
Approximations
Why Approximations:
Analyzer
Analyzer
Approximate
Model
Human
judgment
problems
ice
for large problems to
reduce and control cost
$$
cents
a.k.a. Surrogate Models
Optimizer
Optimizer
OK for sma
Now-standard pract
ll
Design of Experiments(DOE) & Response
Surfaces (RS)
RS provides a domain guidance , rather than
local guidance, to system optimizer
DOE
Placing design points in
design space in a pattern
Example: Star pattern
(shown incomplete)
RS
X1
X2
F(X)
F(X) = a + {b}
{X} + {X} [c]X
quadratic polynomial
hundreds of variables
Response Surface Approximation
A Response Surface is an n-
dimensional hypersurface relating n
Design of Experiments
(DOE) methods used to
disperse data points in
design space.
More detail on RS in
section on Approximations
inputs to a single response (output).
V
a
r
i
a
b
l
e
1
V
a
r
i
a
b
l
e
2
Respon
s
e
BLISS 2000: MDO Massive Computational Problem
Solved by RS (or alternative approximations)
or di
or di
or di
System
optimization
X1
X2
)
X1
X2
)
X1
X2
)
RS
RS
i
in parallel
I
n
s
t
a
n
t
a
n
e
o
u
s
r
e
s
p
o
n
s
e
MC
D
A
T
A
B
A
S
E
Optimization of subsystem
scipline
Analysis of subsystem
scipline
Optimization of subsystem
scipline
F(X
F(X
F(X
Precompute off-l ne
cloud
Radical conceptual simplification at the price of a lot
more computing. Concurrent processing exploited.
Coupled System Sensitivity
Consider a multidisciplinary
Y
A
system with two subsystems
A and B (e.g. Aero. & Struct.)
system equations can be
written in symboli
A[( X ,Y ),Y ] = 0
A B A
B[( X ,Y ),Y ] = 0
B A B
rewrite these as follows
Y
A
= Y
A
( X ,Y )
A B
Y
B
= Y
B
( X ,Y
A
)
B
A
B
A
X
B
X
B
Y
B
Y
A
Y
c form as
these governing equations
define
as implicit functions.
Implicit Function Theorem applies.
Coupled System Sensitivity
-
Equations
These equations can be represented in matrix notation as
Y dY
dX
?
?
? ?? ?
?
?
?
=
?
?
?
A A
A
I
?
Y
A
X
?
?
A
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
Y
B
Y dY?
?
?
?
?
?
?
?
?
?
?
?
B B
I?
0
Y dX
different
A A
same
Y dY?
?
? ?? ?
Right Hand Sides
A A
I ?
0
=
?
?
?
?
?
?
matrix
?
?
?
?
?
?
?
?
?
?
?
?
?
Y dX ?
?
?
Y?
?X
B B
B
Y dY
dX
?
?
?
?
?
?
?
?
?
?
?
?
B B
B
I
?
B
Y
A
Total derivatives can be computed if partial sensitivities
computed in each subsystem are known
Linear, algebraical equations with multiple RHS
Example of System Derivative
for Elastic Wing
Example of partial and system sensitivities
Angle of
atta
ck
deg
10
id wing partial derivative
7.0
Based on rig
Based on elastic wing system derivative
4.0
-40 -30 -20 -10 0
… chord sweep angle -deg
In this example, the system coupling reverses the
derivative sign
X
Flowchart of the System
Optimization Process
System Analysis
α β
γ
System Sensitivity Analysis
α β γ
Start
Sensitivity solution
Approximate Analysis
Optimizer
X
Y
γ
Y
α
β
Y
β
Stop
System Internal Couplings
Co
u
p
l
i
n
g
B
r
e
a
d
t
h
Quantified
All-in-One
Decompose
(
(
D
e
c
o
m
p
o
s
e
)
)
(
D
e
c
o
m
p
o
s
e
)
Strength: relatively large
? YO/ ?YI
Breadth:
{YO} and {YI} are long
[? YO/ ?YI] large and full
Coupling Strength
A Few Recent Application Examples
Multiprocessor Computers create
a new situation for MDO
Supersonic Business Jet Test Case
Structures (ELAPS)
)
)
)
Aerodynamics (lift, drag, trim
supersonic wave drag by A - Wave
Propulsion (look-up tables
Performance (Breguet equation for Range
Some stats:
Xlocal: struct. 18
aero 3
propuls. 1
X shared: 9
Y coupl.: 9
Examples: Xsh - wing aspect ratio, Engine scale factor
Xloc - wing cover thickness, throttle setting
Y - aerodynamic loads, wing deformation.
System of Modules (Black Boxes) for
Supersonic Business Jet Test Case
Struct.
Perform.
Aero
Propulsion
Data Dependence Graph
RS - quadratic polynomials, adjusted for error control
0
1
1 10
Flight Range as the Objective
Normalized
Cycles
0.2
0.4
0.6
0.8
1.2
1.4
2 3 4 5 6 7 8 9
Series1
Series2
RS
1
10
1
0
Analysis
Histogram of RS predictions and actual analysis for Range
12
Air Borne Laser System Design:
another application of the similar scheme
Beam Control System
System Level Design
Turret Assembly
Boeing
Large Optics
CDR 25-27 April
Four Axis gimbals
Transfer optics
747F Aircraft -t
Beam Transfer Assembly
Boeing
Sensor Suite
Active Mirrors
CDR 29 Feb - 3 Mar
Illuminators
Electronics
BMC
4
I
Boeing
Software/Processors
Chemical Oxygen Iodineen
Laser (COIL)
TRW
21-23 March
8-10 March
500000
A Candidate for Shuttle Replacement:
Two-stage Orbital Transport
Collaborated with GWU,
2
nd
stage separates and continues
and ASCAC Branches: System
to destination
Analysis and Vehicle Analysis
LB
x
UB
900000
810000
720000
630000
450000
360000
270000
180000
90000
0
RS
True
Result sample: System Weight (lb) Fly-back
Variance over MDO iterations.
booster
Initial design was infeasible
NVH Model
A Body-In-Prime (BIP) Model - Trimmed Body Structure
without the powertrain and suspension subsystems
MSC/NASTRAN Finite Element Model of 350,000+ edof;
Normal Modes, Static Stress, & Design Sensitivity analysis
using Solution Sequence 200;
29 design variables (sizing, spring stiffness);
Computational Performance
Fine grain parallelism of Crash Code was an important factor
in reducing the optimization procedure total elapsed time:
291 hours cut to 24 hours for a single analysis using 12
processors.
Response Surface Approximation for crash responses
that enabled coarse grain parallel computing provided
significant reduction in total elapsed time:
21 concurrent crash analysis using 12 processors
each over 24 hours (252 processors total).
For effective utilization of a multiprocessor computer, user
has to become acquainted with the machine architecture.
255 days of elapsed computing time cut to 1 day
Computer Power vs. Mental Power
Quantity vs Quality
Invention by Optimization?
P
A
I
b
P
{X} = {A, I, b}; Minimize weight; See b Zero
Optimization transformed frame into truss
A qualitative change
Why:
structural efficiency is ranked:
Tension best
Compression
Bending worst
If one did not know this, and would not know the concept of
a truss, this transformation would look as invention of truss.
Optimizing Minimum Drag/Constant Lift Airfoil
for Transonic Regime
New
(he use a fil i
Base
If this was done before Wh
e & w
Drag minimized while holding
constant lift by geometrically
adding the base airfoils.
Each base airfoil had some
aerodynamic merit
Result: a new type, flat-top
Whitcomb airfoil .
itcomb invented the flat-top airfoil
nd tunnel), this would look like an invention.
Continuous quantitative transformation
vs. conceptual quantum jump
Common feature in both previous examples:
Variable(s) existed whose continuous change
enabled transformation to qualitatively new design
X
no seed
for 2nd wing
OK
Second wing may
Counter-example:
wither away
Optimization may reduce but cannot grow what is not there,
at least implicitly, in the initial design.
Technology Progress:
Sigmoidal Staircase
piston/
vacuum tube/transistor
/digital camera
Performance
jet
film
Time
exhaustion
rapi
inception
d advance
optimizat
Optimization assists
in rapid advance phase
;
ion Human creativity shifts gears
to next step
Augmenting number crunching power
of computer with good practice rules
members
In compression
constraints because the slender members
are not defined until the end.
Subtle point: it is difficult to keep the analysis valid when the
imparted change calls for new constraints.
Topology Optimization
Modern version of what Michelangelo said 500 years ago:
(paraphrased)
to create a sculpture just remove the unnecessary material
QuickTime and a
Base
material
TIFF (Uncompressed) decompressor
are needed to see this picture.
This optimization cannot include buckling
Topology optimization removes pixels from base material
Topology Optimization - 2
Base
material
theoretical
as built
QuickTime and a
TIFF (Uncompressed) decompressor
are needed to see this picture.
members
This optimization can not include buckling
In compression
constraints because the slender members
do not emerge as such until the end.
Subtle point: it is difficult to keep the analysis valid when the
imparted change requires new constraints.
Design by Rules
Tension
Bending
Compression
Structural efficiency
ranking
Structural
weight
String
Truss
Problem Solution Problem
Solution
Problem
Solution
narrow
Problem
obstacle
Complications
Solution 1
Solution 2
.things are getting
too complicated
Human eye-brain apparatus excels in handling
geometrical complexities amplified by abundance of choices
By some evidence, eye-brain apparatus may process
250 MB data in a fraction of a second.
Optimization in Design Process
feedback
Need
or
Concept
Design
Design
Proto-
Production
ld
Oppor-
tunity
Preliminary
Detailed
type
Qualitative
Quantitative
Firm footho
research
extension trend
Optimization most useful where quantitative content is high
Closure
Optimization became an engineer s partner in design
It excels at handling the quantitative side of design
It s applications range from component to systems
It s utility is dramatically increasing with the advent of
massively concurrent computing
Current trend: extend optimization to entire life cycle
with emphasis on economics, include uncertainties.
Engineer remains the principal creator, data interpreter,
and design decision maker.
LaRC/SMC/ACMB
Copyright NASA, Jaroslaw Sobieski, 2003