Ask the Expert:

Ask the Expert - Parametric

Optimization Using
DesignXplorer
John Higgins, PE

1 © 2016 ANSYS, Inc. January 19, 2018

 What is DesignXplorer?
• DesignXplorer is a powerful approach to explore, understand and
optimize your engineering challenges.
– Determine the key parameters influencing the design
– Explore and understand the performance at other design or operating conditions
– Find the conditions which give the best performance
– Explore the robustness of the design

?
?

Response
Single Point What If? Surface

2 © 2016 ANSYS, Inc. January 19, 2018

 DX Features
What if Study
Manual Search
Sensitivity / Correlation Analysis
Find the relevant parameters

Design of Experiments
Run a smart set of Design Points

Response Surface
Build a Mathematical model

Optimization
With or without a Response Surface

Robust Design
Six Sigma Analysis

Optimized and Robust Design

3 © 2016 ANSYS, Inc. January 19, 2018

 Step 1: Define Input & Output Parameters

4 © 2016 ANSYS, Inc. January 19, 2018

 Parameters
Parameterizing CAD dimensions

• CAD parameters are filtered with a prefix/suffix

• The default filter is “DS” (can be changed or left blank)
• Leaving the filter blank will allow all CAD parameters to be
passed to Workbench.

5 © 2016 ANSYS, Inc. January 19, 2018

 Parameters
Parameterizing WB dimensions, BCs, etc.

Parameters created in DM
and Mechanical

Parameters created in CFX, Fluent and CFD

Post

6 © 2016 ANSYS, Inc. January 19, 2018

 Parameters
Derived Parameters
• Derived Parameters:
– Parameters created using analytical expressions composed of input and output parameters.
– Derived parameters can be defined using various built in arithmetic, trigonometric and
statistical functions.
– Examples include (but are not limited to):
• cost function (i.e., the product of mass and cost per mass)
• normalized stress (i.e., stress response divided by an applied stress)
• Mesh sizing (setting a mesh parameter as a function of a geometric parameter)

7 © 2016 ANSYS, Inc. January 19, 2018

 Parameter Set
Table of Design Points (“What-If” Study)

Parameter values currently Update All Design Points

loaded in applications

Select Retain to keep all files for

this Design Point, and be able to
switch between them.

RMB > Update Selected Design Points

8 © 2016 ANSYS, Inc. January 19, 2018

 Workbench Power
Parameters coming from any source

CAD Structural CFD Electromagnetics

DesignModeler Engineering data CFX Ansoft Products:

SpaceClaim Meshing Fluent Designer
Catia V5 Mechanical CFD Post HFSS
UG NX Mechanical APDL Polyflow Maxwell
SolidWorks ExplicitSTR AQWA Q3D Extractor
Creo Parametric Vista TF, CCD… ICEM CFD Simplorer
Solid Edge Icepak
Autodesk Inventor nCode
… And also for some custom usages:
…
Excel, External Connection

9 © 2016 ANSYS, Inc. January 19, 2018

 Workbench Power

• Workbench input and output parameters

is all that DX needs

• Run Design Points simultaneously

– On a cluster
– On available machines in your office

• Take advantage of HPC Parametric Pack

Licenses
– Multiply your number of licenses to run Design
Points simultaneously

• Ability to reserve the licenses to run

10
Design Points
© 2016 ANSYS, Inc. January 19, 2018
 Step 2: Use Parameter Correlation to
Prioritize Input Parameters (Optional)

11 © 2016 ANSYS, Inc. January 19, 2018

 Parameters Correlation
Purpose

• In a DOE study, the amount of design points increases quickly as

the number of input parameters increases, which can reduce the
efficiency of the analysis process.
• It is recommended to exclude unimportant input parameters from
the DOE sampling in order to reduce unnecessary sampling points.

12 © 2016 ANSYS, Inc. January 19, 2018

 Parameters Correlation
Sampling

• The Parameters Correlation method will perform simulations based on a

random sampling of the design space (using Latin Hypercube sampling), so as to
identify the correlation between all parameters
• Latin hypercube – Points are randomly placed, but care is taken to ensure that
no two points share input parameters of the same value

Monte Carlo Latin

Sampling Hypercube
Sampling

13 © 2016 ANSYS, Inc. January 19, 2018

 DX Features
Sensitivity / Correlation Analysis

• Identify unimportant parameters

Input Parameters

These inputs have a low

influence on outputs

Outputs

14 © 2016 ANSYS, Inc. January 19, 2018

 DX Features
Sensitivity / Correlation Analysis

• Identify the degree to which the relationship is linear/quadratic

15 © 2016 ANSYS, Inc. January 19, 2018

 Parameters Correlation
Purpose

• The parameter correlation tool allows us to identify important parameters

• Correlation matrices, determination matrices, correlation scatter plots, and
sensitivity charts also help to understand the parametric relationships.
• Useful when there are many input parameters (more than 15).

16 © 2016 ANSYS, Inc. January 19, 2018

 Step 3: Choose Optimization workflow:
Response Surface vs. Direct

17 © 2016 ANSYS, Inc. January 19, 2018

 Optimization
• DesignXplorer proposes two Optimization workflows:
• Response Surface Optimization (RSO)
• Based on a DOE + Response Surface
• Computation time was spent at DOE step
• Changing Optimization criteria and re-running
is almost costless
• Optimum results are approximated and
should be verified by an actual resolution

• Direct Optimization
• Straight forward
• Based on “real” solves
• Changing Optimization criteria and re-running is expensive
• Optimum results rely on an actual resolution

18 © 2016 ANSYS, Inc. January 19, 2018

 Step 4: DOE for Response Surface
Optimization

19 © 2016 ANSYS, Inc. January 19, 2018

 Step 4: DOE for Response Surface
Optimization

CFD-Structural
analysis
Input and output parameters
are defined in the Fluid Flow
and the Structural systems

Parameter Set
List of all parameters
Design Points table (What If)

DX system
3-step workflow

20 © 2016 ANSYS, Inc. January 19, 2018

 DX Features
Design of Experiments

• Specify the DOE Type

• Specify each parameter range and type
(Continuous, Discrete, Manufacturable Values)
• Design Points are automatically chosen to explore the
parametric space efficiently

21 © 2016 ANSYS, Inc. January 19, 2018

 DOE purpose and schemes
The purpose of a Design of Experiments is to gather a representative set of data to compute a
Response Surface, and then run an Optimization (for a Response Surface Optimization).
Basically, a set of Design Points will be calculated.
The Response Surface accuracy will greatly depend on the DOE scheme, and especially the
number of Design Points that were calculated.
DesignXplorer proposes several DOE schemes. Design Points are automatically chosen to
explore the parametric space efficiently.

• Central Composite Design (CCD) [default]

• Box Behnken Design
• Optimal Space Filling Design
• Custom + Sampling
• Sparse Grid Initialization
• Latin Hypercube Sampling Design

22 © 2016 ANSYS, Inc. January 19, 2018

 Central Composite Design (CCD)

• Central Composite designs are five-level fractional factorial designs

that are suitable for calibrating the quadratic response model
• A CCD consists of:
– 1 center point
– 2*N axis points located at the –a and +a positions on each axis of the selected input
parameter
– 2^(N-f) factorial points located at the -1 and +1 positions along the diagonals of the input
parameter space [factorial number f discussed on next slide]
max

P1 diagonals

min

min max axis

P2
23 © 2016 ANSYS, Inc. January 19, 2018
 Central Composite Design (CCD)

• Factorial (f) : In order to restrict the number of design points to a

reasonable number, some diagonal points are not included based
on the factorial f
Number of
Parameters

24 © 2016 ANSYS, Inc. January 19, 2018

 Additional Settings
Manufacturable Values

– Use to represent real world manufacturing or production constraints.

• Only values that realistically represent manufacturing capabilities are included
in the post-processing analysis
• Verification points and optimization candidates will all be “manufacturable”

25 © 2016 ANSYS, Inc. January 19, 2018

 Extension (New at R16.0)
• Code your own Design of Experiments via an ACT extension
• Some examples are already available

26 © 2016 ANSYS, Inc. January 19, 2018

 Design of Experiments
• Fewest design points
– Custom/Optimal space filling

• Most design points

– Sparse Grid *
– Custom/Optimal space filling

• Use with highly non-linear response

– Sparse Grid *
– Custom/optimal space filling
– Any DOE + Kriging (with auto-refinement)

• Coverage of extremes / even distribution

– CCD
– Box-Behnken
– Custom

Good default choice:

• DOE: CCD (default)
* Few design points are created during the DOE but many refinement points are automatically generated with the response
surface

27 © 2016 ANSYS, Inc. January 19, 2018

 Step 5: Generate Response Surface

28 © 2016 ANSYS, Inc. January 19, 2018

 DX Features
Response Surface
– Generate a Surrogate Model
– Min/Max Search
– 2D/3D plots
– Local Sensitivities
– RS Quality Assessment

29 © 2016 ANSYS, Inc. January 19, 2018

 Response Surface Types
• There are five response surface types in DX
1. Standard Response Surface (2nd order polynomial) [default]
2. Kriging
3. Non-parametric Regression
4. Neural Network
5. Sparse Grid

Standard Response Surface /

Kriging Non-parametric Regression Neural Network

30 © 2016 ANSYS, Inc. January 19, 2018

 Standard Full 2nd Order Polynomials

• This is the default response surface type and a good starting point
• Based on a modified quadratic formulation
Output=f(inputs)
where f is a second order polynomial

• Will provide satisfactory results when the variation of the output parameters is
mild/smooth

f(X)

31 © 2016 ANSYS, Inc. January 19, 2018

 Kriging
• A multidimensional interpolation combining a polynomial model similar to the one of the
standard response surface, which provides a “global” model of the design space, plus local
deviations determined so that the Kriging model interpolates the DOE points.
Output=f(inputs) + Z(inputs)
where f is a second order polynomial (which dictates the “global” behavior of the model) and Z a perturbation term (which
dictates the “local” behavior of the model)

• Since Kriging fits the response surface through all design points the Goodness of fit metrics
will always be good

Z(X) : local deviations

f(X)
y(X)

32 © 2016 ANSYS, Inc. January 19, 2018

 Non-parametric Regression

• Belongs to a general class of Support Vector Method (SVM) type

techniques
• The basic idea is that the tolerance epsilon creates a narrow
envelope around the true output surface and all or most of the
sample points must/should lie inside this envelope. f(X) + 
f(X): Response surface
with a margin of
f(X) - 
tolerance

33 © 2016 ANSYS, Inc. January 19, 2018

 Summary
• Standard Response Surface 2nd-Order Polynomial (default)
– Effective when the variation of the output is smooth with regard to the input parameters.
• Kriging
– Efficient in a large number of cases.
– Suited to highly nonlinear responses.
– Do NOT use when results are noisy; Kriging is an interpolation that matches the points exactly.
– Always use verification points to check Goodness of Fit.

• Non-Parametric Regression
– Suited to nonlinear responses.
– Use when results are noisy.
– Typically slow to compute.
Good default choice:
• Neural Network Kriging with auto-refinement
– Suited to highly nonlinear responses.
– Use when results are noisy.
– Control over the algorithm is very limited.
• Sparse Grid
– Suited for studies containing discontinuities.
– Use when solve is fast.

34 © 2016 ANSYS, Inc. January 19, 2018

 Goodness of fit table
Table of metrics to assess the Response Surface quality

Metrics:
• Coefficient of Determination (R2 measure):
– Measures how well the response surface represents
output parameter variability.
– Should be as close to 1.0 as possible.

• Adjusted Coefficient of Determination:

– Takes the sample size into consideration when
computing the Coefficient of Determination.
– Usually this is more reliable than the usual coefficient
of determination when the number of samples is small
( < 30).

• Maximum Relative Residual:

– Similar measure for response surface using alternate
mathematical representation.
– Should be as close to 0.0 as possible.
Equations in Appendix
35 © 2016 ANSYS, Inc. January 19, 2018
 Response Surface Results

• Response Points
• Min-Max Search
• Charts
– 2D, 2D Slices, 3D Response
– Local Sensitivity Bars/Pies
– Local Sensitivity Curves
– Spider

36 © 2016 ANSYS, Inc. January 19, 2018

 Response Surface Results
• From the DOE results, response
surfaces are calculated to relate the
input to the output parameters
• ANSYS then uses these response
surfaces to calculate optimized sets of
input parameters
• If criteria changes – you do NOT need
to re-solve!

37 © 2016 ANSYS, Inc. January 19, 2018

 Response Surface Results

• Response chart
– 2D, 3D: standard (x, y) and (x, y, z) plots
– 2D Slices (handbook):
• X axis: Input Parameter 1
• Slices: Input Parameter 2
• Y axis: Output Parameter

38 © 2016 ANSYS, Inc. January 19, 2018

 Response Surface Results
Local Sensitivity
• Bar chart
• The change of the output based on the
change of each input independently
• Pie chart
• The relative impact of the input parameters
on the local sensitivity

Output max  Output min

Output avg

39 © 2016 ANSYS, Inc. January 19, 2018

 Response Surface Results
• From the response surfaces, spider plots can be used to dynamically investigate
changes to the input parameters and how they affect the output parameter values

40 © 2016 ANSYS, Inc. January 19, 2018

 Step 5: Optimization

41 © 2016 ANSYS, Inc. January 19, 2018

 Optimization
• Optimization algorithm behaviors are the same for Response Surface
Optimization and Direct Optimization (few exceptions, as seen in next slides)
• Default settings may be different for a RSO or a Direct Optimization system
• For RSO, output parameter results are retrieved from RS evaluations (very quick)
• For Direct Optimization, output parameter results are retrieved from real solves

42 © 2016 ANSYS, Inc. January 19, 2018

 Objectives and Constraints

• One or several Objectives can be defined

• However, some algorithms don’t support multiple objectives
• A weight can be assigned to each objective
• One or several Constraints can be defined

• Objective Types: Constraint Types:

• Minimize • Value <= Limit
• Maximize • Value >= Limit
• Seek Target • Lower Limit <= Value <= Upper Limit

43 © 2016 ANSYS, Inc. January 19, 2018

 Optimization methods
• There are six optimization methods in DX

1. Screening (Shifted Hammersley) [default]

2. MOGA (Multi-Objective Genetic Algorithm)

3. NLPQL (Non-linear Programming by Quadratic Lagrangian)

4. MISQP (Mixed-Integer Sequential Quadratic Programming Method)

5. Adaptive Single-Objective
(for Direct Optimization only)

6. Adaptive Multiple-Objective
(for Direct Optimization only)

44 © 2016 ANSYS, Inc. January 19, 2018

 Screening
• A non-iterative sampling method using a quasi-random number generator
• Generates a large collection of samples from the response surfaces and
sort them based on objectives, weighting and constraints
• Usually used for preliminary designs
• Benefit:
• Provides a global overview of the design space
• Allows you to identify global and local minima
• Provides several candidates
• Available for both continuous and discrete input parameters
• Drawbacks:
• Not fully accurate (accuracy improves with more sample points)

45 © 2016 ANSYS, Inc. January 19, 2018

 Convergence
• For all Optimization types, the history of each parameter values is reported as a
curve in the UI.
• The convergence criteria are also plotted

46 © 2016 ANSYS, Inc. January 19, 2018

 Optimization Candidates
• One or more Candidates can be returned, depending on user settings
• A number of gold stars or red crosses are displayed next to each objective-
driven parameter to indicate how well it meets the stated objective, from
three red crosses (the worst) to three gold stars (the best) *
• Custom Candidates can be created manually or by right clicking from charts

47 © 2016 ANSYS, Inc. January 19, 2018

 Post-processing
Tradeoff chart
• Input/Output parameters can be displayed on each
axis (2D or 3D)
• Help to visualize the tradeoff frontier (first Pareto
Front) between several objectives
• A Pareto front is a group of solutions such that
selecting any one of them in place of another will
always sacrifice quality for at least one objective,
while improving at least one other.
• The best set of samples (first Pareto front) is
indicated in blue
• The worst set of samples (worst Pareto front) is
indicated in red

48 © 2016 ANSYS, Inc. January 19, 2018

 Post-processing
Verify Candidates (for Response Surface Optimization)
• RSO Candidates are based on Response Surface approximation
• DesignXplorer verifies Candidate Points by creating and updating Design Points with
a "real solve" using the input parameter values of the Candidate Points. The output
parameter values from the real solve are displayed in the row below the response
surface generated output values to allow for easy comparison
• If the results are not similar, it indicates that the response surface is not accurate
enough in that area and perhaps refinement or other adjustments are necessary. It
is possible to insert the Candidate Point as a refinement point.

49 © 2016 ANSYS, Inc. January 19, 2018

 Response Surface Optimization
• Should be based on an accurate Response Surface
• Thousands of configurations can be analyzed within a few seconds
• Several algorithms are proposed to find a refined Candidate

Good default approach:

• If one objective is defined: Screening followed by NLPQL or MISQP
• If several objectives are defined: MOGA

50 © 2016 ANSYS, Inc. January 19, 2018

 Direct Optimization
• Robust and straight forward
• Several algorithms are proposed to find a refined Candidate
• “Hybrid algorithms” fully take advantage of DX technologies

Good default approach:

• If one objective is defined:
• If no discretes, no relationships, and nb inputs < 10: Adaptive Single-Objective
• If no discretes and nb inputs > 10: NLPQL
• Otherwise: MISQP

• If several objectives are defined:

• If some parameters are discrete: MOGA
• Otherwise: Adaptive Multiple-Objective

51 © 2016 ANSYS, Inc. January 19, 2018

 Step 6: Robust Design through 6s
Analysis

52 © 2016 ANSYS, Inc. January 19, 2018

 DX Features
Six Sigma Analysis
Input parameters vary!

Output parameter

• Understand how your performance will vary with your design tolerances
• Determine how many parts will likely fail
• Understand which inputs require the greatest control

53 © 2016 ANSYS, Inc. January 19, 2018

 Six Sigma Analysis
What is it?
• Typical analyses assume a fixed value for each input quantity and assigns a
safety factor to account for these assumptions (deterministic)
• Design For Six Sigma provides a mechanism to include and account for scatter
in input and provide insight into how they affect the system response
(probabilistic)
• A product has Six Sigma quality if only 3.4 parts out of every 1 million
manufactured fail

54 © 2016 ANSYS, Inc. January 19, 2018

 Six Sigma Analysis
What is it?

• Helps answer the following questions:

– How large is the scatter of the output parameters? How robust are the
output parameters?

– If the output is subject to scatter due to the variation of the input

variables, then what is the probability that a design criterion given for
the output parameters is no longer met?

– How large is the probability that an unexpected and unwanted event

takes place (i.e., what is the failure probability)?

– Which input variables contribute the most to the scatter of an output

parameter and to the failure probability? What are the sensitivities of
the output parameter with respect to the input variables?

55 © 2016 ANSYS, Inc. January 19, 2018

 Six Sigma Analysis
Example

If you are performing a thermal analysis and want to evaluate the thermal
stresses, the equation is:
σtherm = E α ΔT
because the thermal stresses are directly proportional to the Young's modulus as
well as to the thermal expansion coefficient of the material.
The table below shows the probability that the thermal stresses will be higher
than expected, taking uncertainty variables into account.

56 © 2016 ANSYS, Inc. January 19, 2018

 Six Sigma Analysis
Procedure
1. Specify input parameter distribution

Uniform Triangular

Truncated
Normal Normal

Exponential
Lognormal

Beta Weibull

57 © 2016 ANSYS, Inc. January 19, 2018