Fast Optimization Intevac

Fast Design Optimization Strategy for
Radiative Heat Transfer
using ANSYS and eArtius Gradient Based
Optimization Method – Pt. 2
Vladimir Kudriavtsev, Terry Bluck (Intevac)
Vladimir Sevastyanov
(eArtius, Irvine, CA)

ANSYS Regional Users Conference, “Engineering the
System”, Irvine, CA, Oct. 6, 2011
Equipment Products Division

Outline

 Part1 of this work presented at ANSYS Users Conference
2011 (Santa Clara) and devoted to hybrid genetic and
gradient based approach (HMGE)
http://www.slideshare.net/vvk0/optimization-intevac-aug23-7f

 Part2 of this work is presented here and is devoted to pure
gradient based method, which is best used when only
limited number of design evaluations is possible (due to
CPU time limitations or other reasons)

Equipment Products Division 2

Current Computational Design Process

8 threads i7
CPU

Computer DELAY Ingenious
Solutions

240 cores TESLA Graphic
Processing Unit GPU (x2)
Human Thinking
and Analysis

slowest component

fastest component (meetings, reviews,
and grows exponentially faster
alignments, cancelations)


Why Optimization by Computer?
Human can not match computer in repetitive tasks and consistency.
Assuming computational problem takes 4 hours of CPU time,
then in one day (between 8AM to 8 AM) computer is capable of
producing 6 design evaluations, with 42 designs completed in
just 7 work days.

Coupled with multi-processing capability of i7 workstation this number can
easily be multiplied by factors ranging from two to six.
Computer will work during the weekend; it will work when user is on vacation,
on sick leave or on business trip.

Personal “super computer” cost is now
inconsequential for the bottom line.

Software cost sky-rocketed, and its ROI and
utilization efficiency is now most important.

Computer needs algorithmic analogy of “human brain”
to self-guide solution steps.


New Paradigm and CPU Time

New paradigm of multi-objective computational design is now
being born.

No longer designer needs to approach it through “trial-and-error”
simulations, but can rather use “artificial intelligence” of optimization
method to automatically seek and to find best combination of input
parameters (design). Depending on problem size (CPU time) this process
can take from minutes to weeks.

However, now engineer can view tens and hundreds of possible solutions,
automatically singling first truly best designs and then evaluate design
trade-offs between conflicting objectives (Pareto Frontier).

In many instances, examining dozens and hundreds of
computational designs is simply time prohibitive. What to do then?


Intevac c-Si Technology
Substrate Motion
Direction

Si Substrates move on conveyer and heated
Motion Direction

http://www.intevac.com


Problem Formulation
Minimize thermal variation across single substrate and across a group
of substrates during radiant heating stage (TempDiff)

Operate in required process temperature window, T-dev1<Top<T+dev2

Optimization Formulation ANSYS WB Formulation:

Top=400 deg.C

min (TempDiff)
min abs(Tmax-Top) & min abs(Tmin-Top)

Constraints to determine design feasibility:

T<Tmax.constr & T>Tmin.constr, where

Tmin.constr= Top-dev1, Tmax.constr=Top+dev2
If dev1 and dev2 are small, then optimization problem is very restrictive.


Problem Analogy – Hidden Valley in the Mountains

Narrow optimum
Sub-Optimal
range
Optimal
range

xT^4 nonlinear
steep change
Narrow process window

Gradient method requires path, to enter narrow optimal range (due to nonlinearity)
it requires guidance or coincidence. Guidance comes from the previous history
(steps taken before, gradients) and coincidence from DOE or random mutations.


MGP Method-- Analogy
DOE #N

DOE #2

MGP

Initial DOE #3
Tolerance

Design Space

DOE #1

Initial Design
MGP Design Vector Calculated using


Problem Parameters – Geometry and
Temperature
<Tempdiff> =Tmax-Tmin
between 3 substrates

Si subst
minus rate T1

gap
T2
plus

Lamp Bank 1

Lamp Bank 2

lamps
height T1-radiation temperature of first
substrates lamp array;
T2-radiation temperature of
T1,T2 control heat flux from lamps. second lamp array;


Thermal Heating (Radiation) Solution

Interplay between
two lamp arrays

Substrate Motion
Lamp Bank 1 Lamp Bank 2 Direction

Multi-Step Transient History

Transient Heating Scenario: Row1 of substrates is first heated by Lamp Bank1,
then these Substrates moved to Lamp Bank2 and get heated again till desired Top=400 deg.C is
reached. Simultaneously, new substrates with T=Tambient populate Row1 and get heated.
Thus, Row1 heats from 22 to 250 deg.c and Row 2 from 250 to 400 deg.C.
at time t=3.5 sec Row1 T is reset at 22 deg.C; Row2 T is reset at 250 deg.C.
at time t=0 sec Row1 T is set at 22 deg.C; Row2 T is set at 250 deg.C.

About This Study
This Study Consists of Two Parts. In Part 1 (presented in Santa Clara) we
focused on hybrid genetic and gradient based method (HMGE). It has lots of
positive sides, but generally requires many design points, thus is less suitable
for quick improvement studies typical for computational models that require
many hours of CPU time.

In Part2 (presented here) we focus on gradient based approach (MGP) that is
generally capable to produce design improvements in just a few design
evaluations.

In our study we used modeFrontier as optimization enabling (Scheduler) and
statistical data post-processing tool and eArtius multi-objective optimization
methods plug-in tool to guide continuous process of selecting better input
variables to satisfy multiple design objectives.

This process follows “fire and forget” principle.
Gradient based computer thinking combines advantages of precise analytics
with human like decision making (selecting roads that lead to improvement,
avoiding weak links, pursuing best options, connecting dots).

Fundamental Design Optimization Issues
Study Motivation
The biggest issues of current design optimization
algorithms:
 Low computational efficiency
 Low scalability

Reasons:
 Absence of efficient algorithms for estimating gradients
 Curse of Dimensionality Phenomenon
 Searching for optimal solutions in the entire design space while the search
space can be reduced
 Approximating the entire Pareto frontier while the user only needs a small
part of it

Consequences:
 Artificially reduced task dimensions by arbitrarily excluding design
variables
 Overhead in use of global response surfaces and sensitivity analysis
 Have to rely only on use of brute-force methods such as algorithms’
parallelization

©Copyright eArtius Inc 2011 All Rights Reserved

Estimation of Gradients Issues

Estimation of gradients by the Finite Differences Method
(FDM) is resource consuming:
 FDM is performed on each step
 FDM requires N+1 model evaluations to estimate a gradient
(N—the number of design variables)

Consequences:
 Task dimension is limited by 5-10 for expensive simulation models
 Development of efficient gradient based techniques with FDM is
impossible
Also, gradient based optimization algorithms with FDM cannot be
applied to noisy simulation models

eArtius has developed DDRSM method (patent pending) of gradient
estimation which overcomes the issues:
 Spends 0-7 model evaluations to estimate gradients
 Equally efficient for any task dimension up to 5,000 design variables
 Not sensitive to noise in optimized models


Curse of Dimensionality Phenomenon
and Design Optimization
Example of uniformly distributed points:
- Unit interval—0.01 distance between points—100 points
- 10-dimensional unit hypercube, a lattice with 0.01 between
neighboring points—1020 sample points (Richard Bellman)

Adding extra dimensions to the design space requires an
exponential increase in the number of:
 Sample points necessary to build an adequate global surrogate
model
 Pareto optimal points to maintain the same distance between
neighboring optimal points in the design space

For Response Surface Methods:
- eArtius DDRSM spends just 0-7 points for local approximations—
no global approximations

For Approximation of the Entire Pareto Frontier:
- eArtius performs directed search on Pareto Frontier—no global
approximation of the entire Pareto frontier


Approximation of the Entire
Pareto Frontier
Current multi-objective optimization
algorithms are required to uniformly cover
the entire Pareto frontier

Curse of dimensionality: The increase in the
number of design variables causes the
distance between neighboring points in the
design space to be increased exponentially
Minimize f1 = x1
Minimize f 2 = 1 + x2 − x1 − 0.1 ⋅ sin( 3π ⋅ x1 )
2

0 ≤ x1 ≤ 1; − 2 ≤ x2 ≤ 2 1D – 89 Pareto points
Minimize f1 = 3 − (1 + x3 ) ⋅ cos( x1 ⋅ π / 2) ⋅ cos( x 2 ⋅ π / 2)
Minimize f 2 = 3 − (1 + x3 ) ⋅ cos( x1 ⋅ π / 2) ⋅ sin( x 2 ⋅ π / 2
0 ≤ x1 ≤ 0.65; 0 ≤ x 2 ≤ 1; 0.5 ≤ x3 ≤ 1

2D – 2225 Pareto points
2225/89=25 times more!

ND –> 10N Pareto points

Search in the Entire Design Space
Minimize f1 = x1
Minimize f 2 = 1 + x22 − x1 − 0.1 ⋅ sin(3π ⋅ x1 )
0 ≤ x1 ≤ 1; − 2 ≤ x2 ≤ 2
Monte Carlo method:
258 Pareto optimal points (3%)
out of 8192 model evaluations

HMGE method:
89 Pareto optimal points (35%)
out of 251 model evaluations

Pareto frontier is
a straight line x2=0
in the design space

Why do we need to search in the entire design space?
The search along the line x2=0 is also possible

Search in the Entire Design Space (continuation)
Minimize f1 = 3 − (1 + x3 ) ⋅ cos( x1 ⋅ π / 2) ⋅ cos( x2 ⋅ π / 2)
Minimize f 2 = 3 − (1 + x3 ) ⋅ cos( x1 ⋅ π / 2) ⋅ sin( x2 ⋅ π / 2)
Minimize f 3 = 3 − (1 + x3 ) ⋅ cos( x1 ⋅ π / 2) ⋅ sin( x1 ⋅ π / 2)
0 ≤ x1 ≤ 0.65
0 ≤ x2 ≤ 1
0.5 ≤ x3 ≤ 1

2225 Pareto
optimal
points out of 3500
model evaluations

Pareto frontier
is located
on the flat x3=1
in the design space

Why do we need to search in the entire design space?
The search on the plane x3=1 is also possible

Multi-Gradient Pathfinder (MGP) Method

 On the first half-step MGP improves F1
preferable objective (F2)—green arrows

 On the second half-step MGP improves
ALL objectives—blue arrows—to maintain
a short distance to Pareto frontier

 Then MGP starts the next step from the
newly found Pareto optimal point

F2


Directed Optimization on Pareto Frontier
MGP started optimization three times from the same start point {x1=1; x2=1; x3=1},
but with different preferable objectives.
Green trajectory:
Min f1
Min f2
Min+ f3
Red trajectory:
Min+ f1;
Min f2
Min f3
Blue trajectory:
Min+ f1
Min f2
Min+ f3

Light-green small markers visualize entire Pareto frontier, which is located
on the plane x3=1 in the design space


Searching the Entire Design Space
is Not Productive!
ZDT2 Benchmark Problem: multiple Pareto frontiers

Minimize + F1 = x1
  F 2 
Minimize F2 = g ⋅ 1 −  1  
 
 g 
 
 9 n 
g = 1 + ∑ xi 
 n − 1 i =2 
0 ≤ xi ≤ 1, i = 1,..n
n = 30

MGP—18 global Pareto optimal points out of 38 model evaluations
Pointer—5 optimal points out of 1500 evaluations
NSGA-II & AMGA—FAILED to find a single Pareto optimal point after 1500
evaluations!


Searching the Entire Design Space
is Not Productive!
g = 1 + 10 ⋅ (n − 1) + ( x2 + x3 + ... + xn ) − 10 ⋅ [cos(4πx2 ) + cos(4πx3 ) + ... + cos(4πxn )], n = 10
2 2 2

h = 1 − F1 / g − (F1 / g ) ⋅ sin(10πF1 ); [ X ] ∈ [0;1]
Minimize + F1 = x1 MGP spent 185 evaluations, and found exact solutions
Minimize F2 = g ⋅ h Pointer, NSGA-II, AMGA spent 2000 evaluations each, and failed


OPTIMIZATION RESULTS


MGP – Start from Arbitrary (Bad) design
(TempDiff+, SubMax400+)

34.6 Arbitrary
TempDiff, Initial
deg. C 1 design

Objective 1

Improvement direction

2

8
6.7

Objective 2 (secondary consideration) Design Table


MGP (TempDiff+, SubMax400+)
Temp Diff,
deg. C Objective 1 (main consideration)
34.6 1

Obj1
Rapid SubMax400,
improvement deg. C Objective 2 ( second consideration)
1

6.7
5
Design ID #

Design ID #
Start from Arbitrary (Bad) Design


MGP –Start from Good Point (Obj1)
12 Objective 1 (main consideration)
Obj1

2 Improvement

Obj1 got worse, Obj2 improved
SubMax400,
deg. C
Temp Diff, 23
deg. C Objective 2 (secondary consideration)

5.9 1

2
Start from Good Design (by Obj1)

3.2


MGP: Start from Small DOE (12 designs)
Temp Diff,
deg. C Temp Diff, Obj1
19.7 deg. C 19.7
Obj1 DOE Points (designs) Pareto

DOE
Obj2
Obj2
40

converged
improvement

5.8
Design ID # Obj1
10
#13, first MGP point (design)

0.6
improvement
Obj2
Design ID #

MGP: First Design after DOE (detail of previous slide)
19.8 Objective 2 (secondary consideration)
Objective 1 (main consideration)

6

DOE Points (designs)

6
improvement
~with best DOE

Objective 1

improvement 6

Objective 2 (secondary consideration)


MGP: Start from Several Best Points
Temp Diff,
deg. C
Temp Diff,
Obj1 deg. C
Objective 1
9.97

Best Initial Points
Best Initial Points
5.83

5.44
Improvement
4.87

Improvement Improvement


“Sequence Jumping” DOE for MGP

#3

Step2
Step3 #1
#2
Initial Point
Step2 #2 #3
#1
Step1
Step3
Step1

Initial Point #1, #2,#3 –best points on each step for
objective marked “+” (preferred objective)
Design Space

Multi-Step Fast Start


Multi-Step Fast Start MGP
TempDiff, Objective 1 (main consideration) TempDiff,
deg. C
10.9 deg. C
33.9
1

Design
Improved -Stop Design
Improved -Stop
10.9

5.9
2

Step1 with Initial Tolerance Step2: Tolerance Reduced
Steps continue as long as improvement is reached within short number
of designs
Quick Search for Good Starting Point: multi-step “short” MGP instead of initial DOE.
Advantage: multi-step approach has solution feedback, DOE does not.


Multi-Step Fast Start MGP (last step)

TempDiff, Objective 1 (main consideration)
deg. C 6

Results got worse,
No need to continue
multi-step improvement
any more

2

5.9 1

Step 3: Tolerance Reduced


Computer vs. Human

In head-to-head competitions best “human guided” (case-by-case) studies
resulted in system design with ±10-20 deg.C thermal uniformity and took
several weeks to accomplish, while FAST MGP method based computer
optimization approach allowed to quickly yield design solutions capable
of reaching ± 3 deg. C. It took only 8-20 design evaluations for CPU to
“independently” accomplish this task.

Such an approach will not allow to uniformly cover entire design space,
but will work for engineers who need to find quick improvements for their
designs and work with large computational models that take many CPU
hours to solve (i.e. hundreds of design evaluations are not an option).

We can conclude that “Optimization Equals Innovation”!


Conclusion: Optimization = Innovation

modeFrontier ANSYS eArtius
WorkBench

Acknowledgement

Authors are thankful to ESTECO engineers for developing
eArtius MGP modeFrontier plug-in; to Alberto Bassanese
(ESTECO) for introducing and helping with modeFrontier; to
ANSYS Distributor in Bay Area Ozen Engineering
www.ozeninc.com (Kaan Diviringi, Chris Cowan and Metin
Ozen) for help and dedicated support with ANSYS Workbench
model development and integration with modeFrontier .

Part 1 of this presentation is devoted to thermal optimization when CPU time budget is more flexible to
allow many computational design evaluations. It was presented at Santa Clara Aug. 2011 ANSYS Users
Conference
http://www.slideshare.net/vvk0/optimization-intevac-aug23-7f


SUPPLEMENTS


eArtius – new word in multi-objective
optimization capabilities

Used
in this
study

www.eartius.com


Thermal System Optimization Task
Formulation

Need to
carefully
consider

Minimize+ – preferable
objectives
278 feasible designs of 317 evaluations
Minimize – regular objective
18 Pareto+ designs of 35 Pareto optimal designs


“Fire And Forget” Solution Process - HMGE

deg.C
Temperature
Uniformity

21
First Wave Second Wave
17

13

9
Touchdown
DOE
5 Design ID (#)
46 92 138 184 230 276 300

Fast Optimization Intevac

More Related Content

Viewers also liked (17)

Similar to Fast Optimization Intevac (20)

Fast Optimization Intevac