Engineering Design
under Uncertainty
Xiaoping Du
Department of Mechanical and Aerospace Engineering
University of Missouri – Rolla
April 2004
Our Research Sponsors
n National Science Foundation
n University of Missouri Research Board
n UMR Intelligent Systems Center
Outline
n Optimization
n Why design under uncertainty?
n Challenges in design under uncertainty
n Optimization design under uncertainty
n Robust design
n Reliability based design
n New development
Analysis Model for Design
Analysis model: relationship between
product performance and design variables
Input Analysis Output
Model
Design variables Performance
Length, thickness, height P Weight, stress, deflection
factor of safety
Design Optimization
Initial design Design
variables Performance
Analysis
Model
Analysis
Analysis (simulation) model
Yes
New design
Satisfactory? Optimal results
No
Update
Modify the design
Optimization Design Model
Minimize f(x) Objective function
x=(x1,x2,…,xn) Design variables
Subject to
gi(x)=0, i=1,2,…,ni Inequality constraint
hj(x)=0, j=1,2,…,ne Equality constraint
xk,l=xk=xk,u Side constraint
Task:
Task: find
find an
an optimal
optimal set
set of
of design
design variables
variables that
that
minimizes
minimizes thethe design
design objective
objective while
while satisfies
satisfies all
all
the
the design
design constraint.
constraint.
Why Design Under Uncertainty?
n Uncertainties exist anywhere
n Uncertainties have a great impact on
system performance
n Simulation models are not perfect
n Uncertainty consideration can improve
quality and reliability and
n Assist designers make reliable decisions
under uncertainty and risk.
Examples of Uncertainty
n The width b2 = 10 ± 0.02mm.
n The actual value could be 10.015, 9.99, or
10.01.
P
A A-A
C B
h1 h2
b1
l2 l1 A
b2
n External force P
Uncertainty Analysis
n What is the impact of input uncertainties on
the system performance?
n What is the distribution of y given the
distribution of x?
x y ???
System
Effects of Uncertainty
Quality Loss – Small variations of system input may
lead to large variations of system output.
y Solution – Robust Design
Robust
Optimized Solution x2
Solution Failure (infeasible)
Region
Constraint Boundary
xopt x g=0
Deterministic
optimal point
Risk – Low constraint satisfaction
Solution – Reliability-Based Design Safe (feasible)
Region
x1
Challenges of in Uncertainty
Response y Analysis
Computational Burden
PDF of y
Capture the probabilistic
Input x2
characteristics of system
1 response
utx
Inp
PDF of Input x2
Design Point
PDF of Input x1 Perform a number of
deterministic analyses
Methods of Uncertainty Analysis
n Given distributions of x, find distributions
of y
n Sampling based method
n Monte Carlo Simulation
n Improved Monte Carlo Simulation
n Reliability based method
n Design of Experiments
Design Optimization under
Uncertainty
n Mitigate the effects of uncertainty without
eliminating the causes of uncertainty
n Robust design
n Reliability based design
Robust Design
The design is insensitive to the variation of input
by minimizing
• the mean performance (objective ) and
• its variation
pdf of Shift the distribution
OBJ µ
µ
Before
After
OBJ
Shrink the distribution
Reliability Based Design
• The probability of constraint satisfaction g<0 is
greater than the required reliability.
• Reliability = probability of constraint satisfaction
= Prob{g(x)≤0}
Prob{g(x)≤0}≥R pdf of g
0 g
Integrated Design Model
Objective robustness
n min f ( µobj , σ obj )
n Design Variable DV = {d , ?x }
n Subject to: Prob{gi (d, X, P ) ≤ 0} ≥ Ri
n i=1, 2, …, m Reliability constraint
n d – deterministic design variables
n X – random design variables
n P – random parameters
n R - required reliability
A Framework of Probabilistic Design for
Industrial Applications
Design Model Physical Experiments
Empirical Model Simulation Model
• CAE Model
Simulation Model • FEA Model
• CFD Model
Mathematical Model • ......
Design of Experiments
Probabilistic Design
Surrogate Model
Robust Design
Reliability-Based Design
Uncertainty Analysis
Design for Six Sigma
Monte Carlo Simulation
Reliability Assessment Methods
• Optimization
Sensitivity Analysis • Trade-Off Analysis
• ......
Research Issue – How to
Improve Design Efficiency?
Initial Design Optimal Design
Optimization
Double Loop Optimization Loop
Procedure Design Variables Reliability Constraints
Uncertainty Analysis
Computationally random reliability random reliability
…
Expensive variables constraint variables constraint
Engineering Simulation Models
Uncertainty Analysis Loop
Sequential Single Loop
Procedure
Perform design in sequential cycles
1 cycle = Optimization + Uncertainty Analysis
Cycle 1 Cycle 2
UA UA
Opt Opt
UA UA
Number of Uncertainty Analyses = Number of cycles
Example - Reliability-Based Design for Vehicle
Crashworthiness of Side Impact
Dummy’s responses
• Head injury criterion
• Abdomen load
• Pubic symphysis force
• Viscous criteria
Reliability-Based Design Model
11 Random Variables
Minimize Mean of Weight (W)
9 Design Variables:
Subject to P{FAbdomen_load ≤ 1.0 (KN)} ≥ R 1 sizes of the structure
P{VC1 ≤ 0.32 (m/s)} ≥ R 2 material types
P{VC2 ≤ 0.32 (m/s)} ≥ R 3 2 Parameters:
P{VC3 ≤ 0.32 (m/s)} ≥ R 4 barrier height and its
P{Dlow ≤ 32 (mm)} ≥ R 5 position
P{Dmiddle ≤ 32 (mm)} ≥ R 6 Normally Distributed
P{Dup ≤ 32 (mm)} ≥ R 7
P{FPublic ≤ 4.01 (KN)} ≥ R 8
P{Vbp ≤ 9.9 (m/s)} ≥ R 9
P{Vd ≤ 15.69 (m/s)} ≥ R 10
Where D=Deflection
VC=Viscous criteria
V - Velocity
P - Probability
F - Force
Result of Reliability-Based
Baseline: Design
Weight=29.17 kg
Convergence History of the Weight
30
Worst Reliability=0.05
29
Optimum result: Opt: Optimization
RA: Reliability Assessment
28
Weight=24.49 kg Cycle 1 Cycle 2 Cycle 3
27
Worst Reliability =0.9
Weight
Opt 1
26
Efficiency:
25 RA2: Rmin=0.9
# of Function 24 RA1: Rmin=0.5 Opt 3
Evaluations Opt 2
RA3: Rmin=0.9
23
Single loop: 1051
22
Double loop: 26984 0 50 100 150 200 250
Number of Function Evaluations
300 350 400 450
Our Current Research
n New efficient uncertainty analysis method
n Saddlepoint Approximations
n For many problems we tested, the method is more
accurate and efficient than existing reliability based
uncertainty analysis method
n Complex engineering systems design
under uncertainty
n Simulation model validation
Saddlepoint Approximation
Existing Reliability Method
0.45
FORM
data2
•Transforms non-normal variables to
data3
0.4
data4
normal variables
0.35 SORM
•Increases the degree of nonlinearity
7
0.3
6.5
FOSPA
Probability of failure
0.25 6
Exact g(x1, x2)=0
0.2
Saddlepoint 5.5
x2
5
0.15
4.5
0.1 4
FORM
3.5
0.05 0 0.5 1 1.5 2 2.5 3 3.5
x1
0
0 0.5 1 1.5 2 2.5 3 3.5 4
a
3.5
2.5
g(u1, u2)=0
2
1.5
u2
1
0.5
-0.5
-1
-3 -2 -1 0 1 2 3
u1
Uncertainty Consideration for
Complex Systems Design
n Complexities
n Large number of uncertain variables
n Multidisciplinary systems
n Large amount of uncertainties
n Multilevel details of information
n Computationally expensive simulations
n Model structure uncertainty
Multidisciplinary Systems
Design
x s, x 1
n Disciplines are Subsystem 1
F1(xs, x1, y21, …, yn1)
z1
ε1(xs, x1, y21, …, yn1)
coupled through
y12 y21
linking variables
n Uncertainties x s, x 2
Subsystem 2 y1n yn1
F2(xs, x2, y12, …, y n2)
propagate z2
ε2(xs, x2, y12, …, y n2)
through linking y2n yn2
variables
x s, x n Subsystem n
Fn (xs, xn y1n, …, yn-1n)
z3
εn (xs, xn, y1n, …, yn-1n)
Hierarchical Approach to Collaborative
Multidisciplinary Robust Design
System-level objective
Compatibility of subsystems
System Level Optimizer System-level D.V.
min F = [µf i , s f i ]
s.t. di = 0 i = 1, n
Subsystem D.V . x0 = [µxs , µ0y , s 0y ]
µxs , µ0y , s 0y µf 1, s f 1, d1 µf n , s f n , dn µxs , µ0y , s 0y
constraints
µxs , µ0y , s 0y µf 2 , s f 2 , d2
Subsystem Level Optimizer 1 Subsystem Level Optimizer 2 Subsystem Level Optimizer n
min d1 = (µy1 − µ0y1)2 + (s y1 − s 0y1)2 min d2 = (µy 2 − µ0y 2 )2 + (s y 2 − s 0y 2 )2
…… min dn = (µyn − µ0yn )2 + (s yn − s 0yn )2
s.t. µg1 + ks g1 ≤ 0 s.t. µg2 + ks g2 ≤ 0 s.t. µgn + ks gn ≤ 0
D.V . x1 = [µx1] D.V . x2 = [µx2 ] D.V . xn = [µxn ]
Subsystem-
0 0 0
level D.V. µf 1, y1, g1, s f 1, y1, g1 µxs, , µy , s y µf 2, y 2, g 2 , s f 2, y 2, g 2 µ0x s , µ0y , s 0y µf n, yn, gn, s f n, yn, gn µ0xs , µ0y , s 0y
Subsystem Uncertainty Subsystem Uncertainty Subsystem Uncertainty
Analysis 1 Analysis 2 …… Analysis n
µf 1, y1, g1, s f 1, y1, g1 µf 2, y 2, g 2 , s f 2, y 2, g 2 µf n, yn, gn, s f n, yn, gn
Robust Design for an Electronic
Packaging Problem
x5,x6,x7,x8 x1,x2,x3,x4
y6,y7
Thermal
xi (i=1~8) and all the model
Electronic
Subsystem 1 Subsystem 2 errors are normally
y11,y12,y13 distributed.
f,h,g1 ,g2
Find: µ1~ µ8
Minimize: w1µf1+ w2σf1 Watt density
s.t. µg1+kσg1 < 0
Component temperature
µg2+kσg2 < 0
µh+kσh < 0
Method µf σf µh σf µg1 σf µg2 σf
SUA -1.847×103 5.363×103 3.901×10-6 1.340×10-2 -44.110 4.046 -44.10 4.390
CSSUA -1.847×103 5.363×103 3.901×10-6 1.340×10-2 -44.110 4.046 -44.10 4.390
MCS -1.829×103 5.319×103 1.0×10-11 1.334×10-2 -44.030 4.137 -44.060 4.130
Multilevel Details of Information
n Let random variables be x Interval of z|x
and interval variables be y. response z
n Then response z=g(x, y) will
be in mixture of randomness
and intervals. pdf of z|y
n Worst case scenario - x
consider reliability at the
worst combination of the Interval of y
y
interval variables pdf of x
Numerical Example
n Objective: minimize the area L=100’’
h = wt
Y
t X
n Constraints w
600 600
g1 ( S , X , Y , w, t ) = S − ( 2 Y + 2 X )
wt wt
2 2
3
4L Y X
g 2 ( E , X , Y , w, t ) = D0 − 2 + 2
Ewt t w
n Random: X and Y; Interval: S and E
Results
n Required reliability = 0.9978
Design variables
Cycle Objective g1R g2R
(w, t)
1 (3.1458, 2.2244) 6.9976 -0.3497 -0.3103
2 (4.0724, 2.0698) 8.4290 -0.0027 -0.0364
3 (3.9852, 2.1196) 8.4471 0.0541×10-3 -0.2101×10-3
4 (3.9848, 2.1199) 8.4474 0.0158×10-7 0.4506×10-7
n Total # of function calls=358 (double loop
needs 4604)
S and E: are Interval S and E are uniformly
variables distributed with same intervals
(t, w) (3.9848, 2.119 ) (3.8064, 2.1528)
Area t×w 8.4438 8.1945
Model Validation
•Parameter
uncertainty
•Model structure REAL WORLD
uncertainty
EXPERIMENTS
COMPUTATIONAL
SIMULATION MODEL
Experimental
Numerical Validation uncertainties
errors Comparison
COMPUTATIONAL EXPERIMENTAL
RESULTS DATA
Our Approach
•Shift the working load to
simulation side 240
230 Experimental observation in Case 1
•Consider all the 220
uncertainties in 210 Simulation prediction
simulations y2 200
190
•Reduce the number of 180
Experimental observation
in Case 2
physical experiments 170
•Efficient joint 160
150
probability density 140 160 180 200 220 240
estimation (Saddlepoint y1
Approximations)
Framework
Experiment
Experiment Simulation
SimulationModel
Model
ENGINEERING
ENGINEERINGDESIGN
DESIGN
Modify a design
Modify a design
Parameter
Parameter
Experimental uncertainty
uncertainty
uncertainty Numerical
Numericalerror
error New design
Validated and updated
simulation model
ENGINEERING
ENGINEERINGANALYSIS
ANALYSIS
Validated and updated
Validated and updated
MODEL
MODELVALIDATION
VALIDATION simulation
simulationmodel
model
Model structure
Performance
uncertainty
prediction
Satisfactory?
Design results
Multidisciplinary
Complex EngineeringSystems Design
Systems under
Design underUncertainty
Uncertainty
Uncertainty
Uncertainty quantification
quantification Model
Model validation
validation –– improve
improve
.. Parameter
Parameter uncertainty
uncertainty confidence
confidence of
of model
model prediction
prediction and
and
.. Model
Model structure
structure uncertainty
uncertainty quantify
quantify model
model structure
structure uncertainty
uncertainty
.. Numerical
Numerical error
error
x s, x 1
Discipline 1 •Reliability based design
z1
•Robust design
y12 y21
xs, x2 •Integrated design
Discipline 2 y31 y13 •Decision based design
z2 under risk and uncertainty
y23 y32
xs, x3
Discipline 3
z3
Multidisciplinary
Multidisciplinary design
design under
under
Uncertainty
Uncertainty analysis
analysis –– impact
impact of
of uncertainty
uncertainty –– manage
manage and
and mitigate
mitigate
uncertainty
uncertainty on
on system
system performance
performance the
the effect
effect of
of uncertainty
uncertainty
Our focus is to deal with the following complexities: 1) multidisciplinary organizations,
2) computationally expensive simulations, 3) large uncertainties, and 4) multilevel
degree of information
Our methods: 1) efficient and accurate algorithms for uncertainty analysis, 2) efficient
strategies and algorithms for design under uncertainty, and 3) efficient model
validation.
Conclusions
n Uncertainty consideration helps engineers
make more reliable decisions.
n Higher reliability, lower risk
n Higher quality
n Design under uncertainty will get more and
more applications in industry.
n Model validation under modeling and
simulation uncertainty will get more
attention and applications.
