HyperStudy 10 Training Manual

HyperStudy®

Altair Engineering
September 4th, 2009

Altair Proprietary and Confidential Information Copyright © 2008 Altair Engineering, Inc. All rights reserved.

Agenda

Morning Session (8:30 AM - 12:30 PM)
1. Introduction to HyperStudy
2. Exercises 1.1 and 1.2
3. BREAK (10 Minutes)
4. Design of Experiments (DOE)
5. Exercises 2.1, 2.2 and 2.3
6. Approximations

Afternoon Session (1:30 - 4:30 PM)
8. Optimization
9. Exercises 4.1, 4.2 and (one of these 4.3, 4.4 or 4.5)
10. BREAK (10 Minutes)
11. Stochastic


Altair’s Vision

Today…

Altair…

Optimization is Driver in CAE Driven Design Process


Optimization with Altair HyperWorks

Topology/Topography Shape/Size Optimization
optimization OptiStruct OptiStruct/HyperStudy

Design Preliminary
Product Design Validation
Prototype
Concept Design CAD Testing
CAE

Parameter Studies Robust/Reliable Design
HyperStudy HyperStudy
HyperMorph


Chapter 1: Introduction to HyperStudy


Altair HyperStudy

• is a software to perform Design of Experiments (DOE),
approximations, optimization, and stochastic studies.
• is applicable to study the different aspects of a design under various
conditions, including non-linear behaviors.
• can be applied in the multi-disciplinary optimization of a design
combining different analysis types.
• is integrated with HyperWorks through direct links to the models in
HyperMesh (.hm), HyperForm (.hf), and MotionView (.mdl).
• models can be easily parameterized. Aside from the typical
definition of solver input data as design variables, the shape of a
finite element model can also be parameterized with ease through
HyperMorph.


Study Definition

DOE Approximation Optimization Stochastic

Parameters Screening
System Performance Study
Response Surface Evaluation
Optimum Design
Variation Analysis
Robust Design
Reliability Design


DOE Studies

Design Space Evaluate
Run DOE results
Sampling

• Invoke solvers
• Batch submission
• Parallel job
submission

• Factorial designs
• Plackett-Burman
• Effects plots
• Box-Behnken
• ANOVA
• Central-Composite
• Latin HyperCube
• Hammersley
• User defined
• External Matrix


Approximations

Responses
Approximation

• Least square
• Moving least square
• HyperKriging


Optimization Studies

Multi-disciplinary Define Run Evaluate
Optimization Design Variables Optimization results

• Invoke solver directly
• From approximation
NVH Analysis
• Genetic Algorithm
• Sequential Quadratic
Programming
• Adaptive Response Surface
Size variables
• Method of Feasible Directions
• External optimizers

• Objective function
• Design constraints
• Design variables
• Responses

Crash Analysis Shape variables


Stochastic Studies

Define Select Run Evaluate
input distributions sampling method stochastics results

• Call solver directly
• Using response
surface

• Normal • Simple Random • Probability Distribution
• Uniform Function (PDF)
• Latin HyperCube
• Triangular • Cumulative Distribution
• Hammersley Function (CDF)
• Exponential
• Histogram Distribution
• Weibull
• Ant-Hill plots
• Statistical moments


HyperStudy Process Flow

Study Setup

DOE Setup

Approximation

Optimization Setup

Stochastic Setup


Preference file


Old Editor


New Editor


Job Management


Solver Script


Parametrization of a FEA model

The process of selecting design variables / factors / random variables for
Optimization / DOE / Stochastic studies can be summarized as:

Models in HyperMesh, Parameterized Model using
MotionView, or HyperForm Create Template GUI
Input file for HyperStudy

DOE Approximation Optimization Stochastics

Direct Response Selection for
Automated Post-Processing


Parametrization

Direct linking to Hypermesh-MotionView-HyperForm provides HyperStudy direct
access to simulation models and to the features such as thickness, concentrated
masses, shape changes which are used as the design variables in DOE,
optimization or stochastic studies .

When HyperStudy is launched in a standalone mode, the design variables need to
be identified using a common method as HyperStudy interacts with solvers of
different physics.

HyperStudy uses TEMPLEX to parameterize the solver input deck. The solver
input deck is taken and “parameter” statements are added, which in turn is linked
to the property such as thickness, concentrated masses. This file with “parameter”
statements is called as “parameterized input deck”. This deck forms the input file
for HyperStudy.


HyperStudy/HM-MV-HF link

Input file for HyperStudy (.hm, .mdl, .hf)




HyperMesh-HyperStudy Link

Shapes in HyperMesh

Model Parameterization

HyperMesh-HyperStudy Link


Exercise 1.1:
Create Design Variables Through HyperMesh


Parametrization Using Create Template GUI

Parameterized File for HyperStudy (.tpl)




Parametrization using Create Template GUI

A parameterized solver input deck is required when HyperStudy is launched in standalone mode.


What is Templex?

Templex is a general-purpose text and numeric processor. Templex
generates output text based on guidelines defined in a template file.
The input files can be created using any text editor or word processor that
can save a file in ASCII format.
Templex solver can be used to process the template file and the output can
be sent to a monitor, to another file, HyperGraph or to other output streams.
Templex is built into HyperStudy as an expression builder for creating
responses for Objectives and Constraints.
Templex can also be used as a standalone solver on mathematical
expressions to perform Design of Experiments (DOE) / Optimization /
Stochastic studies.


Important Templex Rules

• Statements, expressions, and variables must be placed between braces, { }, so that
Templex will evaluate them. Otherwise they are treated as plain text and not
processed.
• After Templex processes a parameterized template file, the plain text is copied to the
specified output stream verbatim.
• When Templex encounters an expression within braces, it evaluates the expression
and sends the resulting value to the output stream. If the item between the braces is a
Templex statement, Templex carries out the statement instructions. If the Templex
statement generates output, the resulting value is sent to the output stream too.
• Variable names can be up to eighty characters long. They consist only of letters,
numbers and underscores; no other characters are allowed. Variable names must
always begin with a letter and are case sensitive.
• In addition to scalars Templex supports one- and two- dimensional arrays. In
HyperStudy to define the responses the results from the analysis runs are
represented as one-dimensional arrays “VECTORS”. The index starts with 0 as in
C language.
• The template file must have at least one parameter statement to be able to read
into the HyperStudy.


Create Template GUI

The parameterized input deck is a Templex template, *.tpl, file.

To create the *.tpl file, the following are added to the analysis input file :

1. Templex parameter statements,
2. Field variable calls, and
3. External-file-include statements


Create Template GUI

1. Templex parameter statements are used for individual field changes
and shape vector (DESVAR) nodal perturbations. All parameter
statements are placed at the top of the *.tpl file.
The templex parameter function has the following format:
{parameter(varname, label, nominal, minimum, maximum)}

Example:
{parameter(TH1, ”Thickness 1”, 1.6, 0.5, 3.5)}


Create Template GUI

2. Field variable statements must be inserted into the analysis deck for
each field to be optimized, such as component thickness, density, etc.
The formats for field variable statements is as follows:
{variable,%field_width.decimal_places}
‘Variable’ is the same as in the corresponding parameter statement.
The ‘field width’ value depends on the solver being used so that the proper
spacing is maintained in the analysis deck. Each solver will have it’s
own filed width: LS-DYNA is 10, OptiStruct is 8, and HyperForm is 8.
The ‘decimal_places’ value is a user preference.
Example
{TH1, %8.5f}


Example: LS_Dyna Input Deck

Original LS-DYNA Deck

$HMNAME PROPS 1beam
1 71.0 20.0 0.0 0
1.6 1.6 1.6 1.6 0.0

Parameterized LS-DYNA Deck (.tpl)
$HMNAME PROPS 1beam
1 71.0 20.0 0.0 0
{TH1,%10.5f}{TH1,%10.5f}{TH1,%10.5f}{TH1,%10.5f}0.0

3. If the design variables in optimization are shape vectors, further editing of the
deck is necessary such as external-file-include statements.


Example: Nastran Input Deck
HyperStudy offers a simple GUI to parameterize input decks.


Example: Nastran Input Deck


Exercise 1.2:
Create Design Variables using the Create
Template GUI


Direct Solvers & Results Access

• Use of HyperWorks result readers: no configuration needed
• Supported Solvers • HyperWorks Solvers
• ABAQUS • OptiStruct
• ADAMS • MotionSolve
• ANSYS • HyperForm
• DADS • General Result Formats
• LS-DYNA • MS Excel (.csv file)
• MADYMO • Any XY data
• MARC
• File Parser
• NASTRAN
• Additional interfaces
• PAM-CRASH
• Import Reader Language
• RADIOSS
• External Readers Programmed in C
• SIMPACK


Interfacing with Solvers
By default, HyperStudy has 4 solvers: OptiStruct, Templex, HyperForm and
HyperXtrude. Additional, solvers may be added through the
Register Solver Script GUI or
Preference file located in <install_dir>/hw/prefinc directory.


Interfacing with Solvers through the GUI


Interfacing with Solvers through the preference file


Integration with Excel


HyperStudy/HyperMorph Coupling

• Morphing: automatic mesh parameterization
• Morphing shapes used as design variables in
HyperStudy and OptiStruct . Modifies the
variable values
• Mesh based, no CAD data needed
FEM
HyperMesh
HyperMorph

Updates

Optimizer
Solver
HyperStudy
Runs

Responses:
stress,
volume, etc. Extracts the
responses


Shape Variable Generation using HyperMorph

Shape design variables necessary for DOE/Optimization/Stochastic studies can be easily
handled using HyperMorph module within HyperMesh.

The finite element model is morphed to a given shape and then saved as “shape”. Multiple
“shapes” can be saved in HyperMorph. These saved “shapes” are then used by
HyperStudy as design variables.
Define Domains/Volumes
Domains
and handles

Morph handles or
alter dimension

Morph Volumes
Morph to Geometry Save “Shape”
Handles
• Generated automatically
• Can be added by hand
Undo morph
• Global
• Local
• Biasing allows for C1 continuity


Shape Variable Generation using HyperMorph


Example

Design Variable:
Boss Radius

Old Radius
New Radius


Exporting Shape Variables

Select
Nodes


Exporting Shape Variables


Chapter 2: Design of Experiments (DOE)


Design of Experiments (DOE)





Design of Experiments (DOE)

What is DOE?
Design of Experiments (DOE) can be defined as a series of tests in which purposeful changes are
made to the input variables of a process or system so that the reasons for changes in the output
responses can be identified and observed.

Objectives of DOE Study
• To determine which factors are most influential on the responses.
• To determine where to set the influential controlled input variables so that:
The response is close to the desired nominal value.
Variability in output response is small.
The effects of the uncontrolled variables are minimized.

• To construct an approximate model that can be used as a surrogate model for the actual
computationally intensive solver.


DOE Definitions

• A Factor is a input parameter (or design variable) of the system.
Factors can be controlled or uncontrolled

• A Level is a discrete (or continuous) value of the factor.

• A factor can be either Discrete i.e., slow (-) or fast (+) (Ex. a variety of
seed, type of paint etc.) or Continuous (Ex. temperature, a sinusoidal
input for frequency response analysis etc.)

• Controlled factors are design variables that can be realistically
controlled in the production (real world) environment. Examples
include gauge thickness of sheet steel, shape of a support bracket, and
mold temperature.

• Uncontrolled factors (Noise) are variables that cannot be realistically
controlled in the production (real world) environment, but can be
controlled in the lab. Examples include ambient temperature and
occupant seating positioning.


DOE Definitions

• Main Effects are the influence of the factors on the responses

• Interaction Effect is the interdependence among the factors. Due to
this interdependence a difference in the response occurs when the
factors are changed simultaneously as against what would have
occurred when the factors were changed individually
Main Effects

Interaction Effects


DOE Example

Key DOE concepts are illustrated by considering an airbag
design. Here we want to study how the Speed of Bag Inflation
(S), Size of Vent (V) and the Size of Bag (B), influence the
HIC of an air bag and also find out how these factors interact
with each other.

FACTOR LABEL LEVEL
(Design Variable) (-) (+)
Speed of Bag Inflation S Slow Fast

Size of Vent V Small Large
Size of Bag B Small Large


DOE Example

Initially, consider the simplified problem where the vent size (V) is constant
and consequently the are only two factors, S and B. For this, a simple full
factorial DOE matrix is given as:

ROW S B S X B

1 - - +

2 + - -

3 - + -

4 + + +

D iv is o r 2 2 2

The following DOE study will result in a polynomial expression that relates HIC to the two factors:
Speed of the air bag and Size of the Bag.

Example: HIC = a 0 + a1* S + a 2 * B + a3 * S * B


DOE Study Types - Applications

• DOE Types Available in HyperStudy
• Full Factorial
• Fractional Factorial
• Central Composite
• BoxBehnken
• Plackett Burman
• Latin Hypercube
• Hammersley
• Run Matrix
• User Defined
• DOE for Screening
• DOE for Factorial Studies
• DOE for Response Surface (RSM) evaluation


DOE for Screening

Objective
• A simple DOE study (ex. two level design with no interactions) will
provide a global understanding of the complete system i.e give the
magnitude and direction of effects
• This initial screening exercise will allow parameters which do not
influence the system to be discarded thus reduce the number of factors
and runs.
• Lower precision
Types
• Fractional Factorial
• Plackett-Burman
• D-Optimal


DOE for Factorial Study

Objective
• Fewer factors
• Main effects and some factor interaction effects
• Linear model
Types
• Full Factorial
• Fractional factorial
• D-Optimal


DOE for Response Surface

Objective
• Fewer factors
• Model of relationships
• Accurate prediction
• Optimization
Types
• Box-Behnken
• Central composites
• D-Optimal


DOE Process

1. Assign factors (dv’s) from .tpl, .hm,.mdl, .hf file.
2. Perform nominal run to create responses for DOE study.
3. Select the DOE type for controlled and/or uncontrolled factors.
4. Divide the factors into controlled and uncontrolled if needed.
5. Export the solver input files for the specified runs as required by the
DOE matrix.
6. Solve the above exported files.
7. Extract the responses for the above solved files.
8. Study the main effects, interaction effects, sensitivity index.


Exercise 2.1: DOE Study of a Rail Joint using OptiStruct


Exercise 2.2: DOE Study of a Cantilever Beam using
Templex

Responses:
Design Variables

Width: 20 < b < 40 σ max (L, b, h ) = Mc
I

Height: 30 < h < 90
Lenght: 50 < L < 100 U max (L, b, h ) = PL3
3EI

Design Space: All beam elements

h
Vol (L, b, h ) = Lbh

L b


Exercise 2.3: DOE with MotionSolve Data Mining


Chapter 3: Approximations


Approximation





Approximation

• Approximations are surrogate models that represent the actual responses.

• Why the need for approximations?
- Some simulations are computationally expensive which makes it impractical
to rely on them exclusively for design studies. Use of approximations in such
cases lead to substantial savings of computational resources.
- Optimization can fall into local minimum or maximum when the responses
are nonlinear. Using approximate responses, the user can avoid this issue.

• The challenge:
- When using approximations, the issue of a trade off between accuracy and
efficiency is ever present.
- The question is how approximate the representation of the design space
can be while remaining accurate enough.


Approximations Definitions

• Regression is the polynomial expression that relates the response of
interest to the factors that were varied. It is only as good as the levels
used when performing the study. For example, a two-level parameter
only has a linear relationship in the regression. Higher order
polynomials can be introduced by using more levels. Note that using
more levels results in more runs.

• Linear Regression model
• F(X) = a0 + a1X1 + a2X2 + (error)

• Interaction Regression Model
• F(X) = a0 + a1X1 + a2X2 + a3X1X2 + (error)

• Quadratic Regression Model (2nd order)
• F(X) = a0 + a1X1 + a2X2 + a3X1X2 + a4X12 + a5X22 + (error)


Approximations Definitions

Anova (Analysis of Variance) is a representation of the contributing
percent of each design variable for the selected response.


Approximation Types

• Least Squares Regression
• Moving Least Squares Method
• HyperKriging


Approximation Types

Least Squares Regression (LSR)
LSR attempts to minimize the sum of the squares of the differences
(residuals) between responses generated by the approximation and
the corresponding simulation results.
HyperStudy allows the creation of least squares regressions for any
polynomial order.

Moving Least Squares Method (MLSM)
MLSM is a generalization of a conventional weighted least squares
model building. The main difference is that the weights, associated
with the individual DOE sampling points, do not remain constant but
are functions of the normalized distance from a DOE sampling point
to a point x where the approximation model is evaluated.
HyperStudy provides the choice of first, second, and third order
functions for the Moving Least Square fit.

Approximation Types

HyperKriging
An approximation function, built using Kriging, has the property that it agrees with the original
response function at the DOE points, i.e. this approach produces an interpolation model. This
makes the technique suitable for modeling highly nonlinear response data that does not contain
numerical noise. HyperKriging is computationally more expensive that the LSR and MLSM
method.


Approximations Process

1. Select approximation types for responses
2. Import the run matrix
3. Build approximation models for each response
4. Check the residuals
5. Do trade-off studies and view ANOVA results


Exercise 3.1: Approximation Study of a Rail Joint using
OptiStruct


Exercise 3.2: Approximation Study of a Cantilever Beam
using Templex


Chapter 4: Optimization


Optimization





Optimization Problem Formulation

Objective: min f(x)
min cost ($)

Constraints: g(x) ≤ 0.0
σ < σallowable

Design Space: lower xi xi upper xi
≤ ≤
2.5 mm < thickness < 5.0 mm
number of bolts ∈ (20, 22, 24, 26, 28, 30)


Optimization Definitions

Design Variables: System parameters that can be changed to
improve the system performance.
beam dimensions, material properties, diameter, number of bolts
Objective Function: System responses that are required to be
minimized (maximized). These responses are functions of the
design variables.
mass, stress, displacement, frequency, pressure drop
Constraint Functions: System requirements that need to be
satisfied for the design to be acceptable. These functions are
also functions of the design variables.
displacement, frequency, pressure drop, cost


Optimization Definitions

Feasible Design: Design that satisfies all the constraints.
Infeasible Design: Design that violates one or more constraints.
Optimum Design: Set of design variables along with the
minimized (maximized) objective function that satisfy all the
constraints.

DV2
optimum design
upper

objective function
Feasible
Region

constraint
lower
DV1
lower upper


Optimization Methods

Optimization methods can be classified in three categories:
1. Gradient Based: are effective when the sensitivities (derivatives) of the system
responses w.r.t design variables can be computed easily and inexpensively.

2. Response Surface Based: are very general in that they can be used with any
analysis code including non-linear analysis codes. Global optimization
methods use higher order polynomials to approximate the original structural
optimization problem over a wide range of design variables.

3. Exploratory Methods: are suitable for discrete problems such as finding the
optimum number of cars to manufacture. These methods do not show the typical
convergence of other optimization algorithms. Users typically select a maximum
number of simulations to be evaluated. These algorithms are good in search on
nonlinear domains however they are computationally expensive.


Gradient Based Methods

Simulation

Conv ? Optimum • Design sensitivity analysis (DSA)
must be available
• Applied for linear static and dynamic
Design Sensitivity problems
Analysis (DSA) • Mostly integrated with FEA Solvers
• Not feasible for Non-Linear solvers

Design Update


Surface Approximation Methods

Simulation

• Sequential Response Surface update
• Linear step
Response Surface
Computation • Quadratic response surface
• Non-linear physics
• Experimental Analysis
Design Update • Wrap-Around Software
• HyperOpt is solver neutral


Exploratory Methods

Create a number
of designs (population) • Gradients are not needed
• Suitable to nonlinear simulations
Simulate all and pick
• Computationally expensive as it may
the best one
require large number of analysis
• Convergence not guaranteed

Conv ? Or
# of simulations > Optimum
Max # of sim?

Population Update


Optimization Methods in HyperStudy

HyperStudy offers four optimization engines: Adaptive Response Surface,
Method of Feasible Directions, Sequential Quadratic Programming, Genetic
Algorithm. It also allows for User-defined Method.

With user-defined method, user’s may interface their optimization code with
HyperStudy to perform optimization.

Also, the approximations can be used as surrogate model to perform
optimization.


Adaptive Response Surface Method (HyperOpt)

In this approach, the objective and constraint functions are approximated in
terms of design variables x using a second order polynomial.
The polynomial coefficients are determined using a least squares fit of the
functions on to the previous design points (actual nonlinear analysis
results).
In general, more designs are available than are required for an exact least
squares fit, making the system over-determined. HyperOpt uses a very
efficient algorithm to estimate a response surface to be closer to certain
designs of interest. HyperOpt also uses move limits to make the
optimization algorithm robust.


Method of Feasible Directions

The method of feasible directions is one of the earliest methods for
solving constrained optimization problems.

The fundamental principle behind this method is to move from one
feasible design to an improved feasible design. Hence, the
objective function must be reduced and the constraints at the
new design point should not be violated.


Sequential Quadratic Programming

Sequential quadratic programming (SQP) is a method for solving
constrained optimization problems.

The fundamental principle behind this method is to create a
quadratic approximation of the Lagrangian and to solve that
quadratic problem to define the search direction s.

The constraints are linearized during the search.

This quadratic problem can be solved by a variety of methods. The
solution of the problem yields the search direction along which
the next design that improves the objective function and does not
violate the constraint can be found.


Genetic Algorithm

A genetic algorithm is a machine learning technique modeled after the evolutionary
process theory.

Genetic algorithms differ from conventional optimization techniques in that they work on
a population of designs. These designs are then evaluated for their fitness, which is a
measure of how good a particular design is. Following Darwin’s principle of survival
of the fittest, designs with higher fitness values have a higher probability of being
selected for mating purposes to produce the next generation of candidate solutions.

In addition to reproduction, selected individual designs go through crossover and
mutation. The designs that result from this process (the children) become members
of the next generation of candidate solutions.

This process is repeated for many generations in order to artificially force the evolution of
a population of designs that yield a solution to a given problem.

Genetic algorithms do not show the typical convergence of other optimization algorithms.
Users typically select a maximum number of iterations (generations) to be evaluated.
A number of solver runs is executed in each generation, with each run representing a
member of the population.


Using Approximations as a Surrogate Model
Ordinary Response Surface Method
Ordinary response surface methods are based on DOE studies
Response surface is fitted (quadratic polynomial) with a fixed number of designs analyzed,
consequently it is not sufficient to approximate and optimize highly non-linear functions


Optimization Process

1. Assign factors (dv’s) from .tpl, .hm,.mdl, .hf file
2. Perform nominal run to create responses.
3. Select the Optimization Algorithm
4. Select the Design Variables for optimization study
5. Define Objective function and optimization constraints
6. Launch Optimization
7. Postprocess the optimization results


Optimization Problem Example
• A cantilever beam is modeled with 1D beam elements and loaded with
force F=2400 N. Width and height of cross-section are optimized to
minimize weight such that stresses do not exceed yield. Further the
height h should not be larger than twice the width b.



• Objective
• Weight: min m(b,h)

• Design Variables
• Width: bL < b < bU, 20 < b < 40
• Height: hL < h < hU, 30 < h < 90

• Design Region: All beam elements

• Design Constraints:
σ (b,h) ≤ σmax, with σmax = 160 MPa
τ (b,h) ≤ τmax, with τmax = 60 MPa
h ≤ 2*b


Mathematical Design Space
Beam height, h (mm)

Beam width, b (mm)


HyperStudy Case Studies and Examples


Case Studies Examples

• Optimization studies are increasing in popularity. It provides added
value to the baseline study, it helps the engineer in understanding
the physics of the problem, it allows exploration in the design space.
• The technology has been applied to full body crash. An example
consists of an objective to minimize mass using footwell intrusion
displacements as constraints and several material thickness values
as design variables
• The technique has been applied to headform assessments. Initially
to identify the worst attitude of impact. Once this is established the
rib thickness of the crashbox is determined
• Four case studies are presented. Two relating to structural
crashworthiness, two relating to occupant safety .
• In addition fours examples are given. Two of these examples relate
to computational fluid dynamics and two of them relate to fatigue.


Case Study 1: Rail Optimization

• A box rail of length lx=800mm is clamped at the right side and is
impacted by a moving stone wall of mass 1,000kg and initial velocity
2m/s from the left side
• The rail is required to absorb the maximum internal energy within an
analysis time of ts=30ms. The displacement of the stone wall (x-
direction) should be less than 50mm. The problem is typical of the
designs encountered in car bumpers and crash boxes
• A size optimization is performed with six design variables consisting
of variable thickness patches. The initial thickness of all design
variables was 2mm and could vary between 0.2mm and 3mm



• The HyperOpt optimization (Sequential Optimization) is compared
against local approximation and DOE approaches. Although
structural behaviour is similar, HyperOpt maximizes the internal
energy
• Design iteration history is presented showing the variation of the
objective with the constraint



Problem Setup

max Internal Energy
∆lx = 50mm

Moving stone wall

Fixed end

V0 = 2m/s
m = 1000kg



Size Optimization

Initial Ord. RS Seq. RS Local app.
DV1 [mm] 2.00 1.70 2.02 2.25
DV2 [mm] 2.00 1.92 1.57 2.95
DV3 [mm] 2.00 0.20 1.56 2.11
DV4 [mm] 2.00 1.95 1.65 1.94
DV5 [mm] 2.00 2.63 3.00 2.28
DV6 [mm] 2.00 3.00 2.48 2.28
Emax [Nm] 1010 1177 1254 1170
∆lx [mm] 52.08 49.95 49.88 49.85


Case Study 2: Roof Crush

• The US standard FMVSS216 requires a certification which simulates
the roll over of a vehicle
• The certification test consists of moving an undeformable plate into
an A-pillar at a constant velocity and the load / displacement
response monitored
• The objective is to minimize the mass of the design and still ensure
that the applied load meets the certification requirement. The design
variables consist of five material thicknesses in the A-pillar and the
roof frame



Objective : minimize mass
Constraint : max F(d) > Flimit
Design variables: Thicknesses of A-Pillar
A-
F

plate



Variation of the Normalized Constraint with Time


Case Study 3: Airbag

• Airbag impactor test requires the HIC not be exceed 1000. The impactor
should not displace more than 260mm. The design variables are vent
orifice area, fabric permeability, ignition time and mass flow rate of the
gas
• The test set-up consists of allocating an initial velocity to the impactor
• After twenty-two iterations the HIC has been considerably reduced



d0 Objective : minimize head injury criterion (HIC)

2 .5
 1 
(t − t1 )
t2
HIC = max 
t 2 − t1 ≤ 36 ms t − t
2
∫t
1 1
&&
dimpactor dt 
 2

Constraint : Deflection dmax < 260mm

Design Variables: Ignition time
Fabric permeability
Mass flow rate
Vent orifice area



Objective with Time



Variation of the Objective / Constraint with Design Iteration History


Case Study 4: Knee Bolster Design

• This study considers the occupant behavior to determine the initial
package space for a knee bolster. The model consists of the lower
extremities of the dummy.
• The dummy is loaded by an acceleration pulse a x (t ) resulting from
the crash test in accordance with FMVSS208
• The objective is to minimize the femur load with a limit on the knee
penetration into the knee bolster. The design variables consist of
geometric and stiffness parameters



Substitute Model



Optimization Problem

Objective : Minimize Femur Load:
Ffemur ( ∆θ , ∆k x , ∆lx , ∆lz , ∆α ) ⇒ min
Constraint : Knee Penetration: δ (∆ θ , ∆ k x , ∆ l x , ∆ l z , ∆ α ) ≤ 0 .07 m
Design Variables : Geometric and Stiffness Parameters (∆ θ , ∆ k x , ∆ l x , ∆ l z , ∆ α )



Results



Design Variable Results

DESIGN LOWER UPPER OPTIMAL
VARIABLE BOUND BOUND VALUE
∆ϑ ° -10.00 10.00 -10.00
∆k x -0.10 0.07 -0.10
∆l x -0.05 0.03 0.03
∆l z -0.05 0.05 0.05

∆α ° -5.00 5.00 -4.65


Case Study 5: Exhaust System Design

• Objective: Uniform flow velocity in
inflow section S
- Average velocity
- Local deviation:
Cross section
outlow - Uniformity-Index:
Catalyst
• Constraint: Min. pressure drop
between inflow and outflow

S1
S3 S6
6 Shape Variables S2
S4
S5


Case Study 5 : Exhaust System Design

Results

γ +12.0%

∆p -16.0%

Initial

Flow Velocity in Section S
Optimized Section S


Case Study 6: Notch Design

Static load

Dynamic

Dynamic load
Static

Loads & Boundary
Stresses Damage
Conditions


Case Study 6 : Notch Design

Design Variables for Shape Optimization



• Optimization Problem Design Variable
History
• Minimize maximum stress at
notch edge
• Both load cases considered
• Shapes scaled +/- 1 Stress
• Damage deteriorated! Based
Optimization

Damage

Stresses



• HyperStudy identified dynamic load case to Damage
be sensitive to damage
• New contour different to stress optimized
Based
contour Optimization
• Damage significantly reduced


GENERAL DISCUSSION

• Optimization commences once a clean baseline analysis has been
performed (eg. negligible hourglassing and sliding energies)
• Care with shape variables: large perturbations can result in a
reduced time step size
• Mass scaling used with shape optimization should be used with
caution. Particularly if the objective is to minimize mass
• For size optimization, numerical contact can be effected by changing
the thickness. Workaround by keeping the contact thickness
constant
• Technology can be simultaneously applied to multiple loadcases
(eg. central / oblique impact attitudes)


GENERAL DISCUSSION

• Changes in design variables can result in a completely different
system response (eg. components not in contact may now contact)
• Original analysis duration may extend to capture the systems
response optimization
• Computational time can be a barrier with each iteration requiring a
complete explicit analysis
• Convergence is quickly achieved
• Optimized design can be conveniently restarted with new design
variables or further optimization starting at a new position in the
response surface


Exercise 4.1: Shape and Size Optimization of a Rail Joint


Exercise 4.2: Size Optimization Study using RADIOSS


Exercise 4.3: Optimization Study Using an Excel
Spreadsheet


Exercise 4.4: Shape Optimization Study using
HyperForm


Exercise 4.5: Shape Optimization Study using ABAQUS


Chapter 5: Stochastic


Stochastic





Uncertainty in Design
Physical uncertainty Loads
• Loads • Direction
• Boundary and Initial condition
• Magnitude
• Material properties
• Random noise (uncontrolled)
• Geometry
variables

Numerical simulation
uncertainty Material data
• Conceptual modeling • Elastic properties
• Mathematical modeling
• Failure
• Random noise or design
Manufacturing variables
• Sheet metal thickness
• Welds
• Random design (controlled)
variables

Robust Design – Stochastic Study

Minimizing variations in performance caused by variations
in design variables
Performance
g

Noise (Uncontrolled)
Variables
Robust
Solution
Design Optimum
Solution
System Input Design Performance
Response

Design (Controlled) µ x
µ opt µrobust
Variables
Design Variable


Robustness Definition

• The design should be invariant against small changes of the
parameters which are inevitable in the normal design process .
• Scattering of design parameters due to real world conditions in
manufacturing and boundary conditions should not reduce the
performance of the product.
• Robustness is the degree to which a system is insensitive to effects
that are not considered in the design.
• Robust statistical procedures has to be designed to reduce the
sensitivity of the parameter estimates to unreliability in the assumptions
of the model.


Robust Design


Robustness Definition


Deterministic Optimization Approach

Find acceptable solution to the design
problem
Design variables are continuous or
discrete quantities FEA
Environmental conditions are given as
deterministic load cases
f ( x ) ⇒ min Approximation
Objective
g j (x ) ≤ 0
Optimization
Constraints x ≤x≤x
L U

Design space


Deterministic Optimization Approach

Deterministic optimization problem formulation does not
incorporate uncertainty of design variables

Optimal solutions for systems exhibiting highly non-linear
responses can be misleading

Surrogate models (response surfaces) built using certain
approximation techniques tend to smooth the noisy behavior

Most optimization algorithms push the constraints to the bounds
in search of optimal solution


Robustness Optimization Approach

Since the objective of any robust design problem is to bring the mean on
target and minimize the variance, robust design problems should be modeled
as bi-objective problem

The two objectives are :
Minimize the variance of performance
Bring the mean on target or optimizing the “mean of performance”
Probability
distribution

Bias
Quality distribution
Tolerance Quality
(System Range) distribution
distribution (System range)
(Design range)

Quality within
specifications
(Common range)

0 l
Target σ Mean µ σ Performance


Robust Design Optimization Formulation

Objective σ (f ( x,r )) ⇒ min

µ ( x,r ) − fTarget ≤ 0
Constraints ?
µ j ( x,r ) − g jTarget ≤ 0

x L ≤ x ≤ xU

Design variables
r = [ri (µi ,σ i )]

Random variables
Objective: Deviation

Constraint: Mean gj


Exercise 5.1: Stochastic Study of a Rail Joint


Exercise 5.2: Stochastic Study of a Plate Model


REFERENCES

‘Structural Optimization in Occupant Safety and Crash Analysis’, U Schramm, D Schneider and
H Thomas, OptiCON’98, California, Oct. 1998.
‘HyperOpt User’s Manual’, Version 6.0, Altair Engineering, Jan. 1999
‘AutoDv User’s Manual’, Version 3.0, Altair Engineering, June 1998
‘Validation of the Hybrid III Head and Neck Models for the Pole Test of FMVSS 201 Using
Optimization Techniques’, L Fredriksson and A Hirth, CAD-FEM User’s Meeting, Oct. 1998
‘Crashworthiness optimization Study of a Bumper System’, K Simpson and G Harte,
HyperWorks Conf. Proceedings, Jan 1999
‘Structural optimization - An Efficient Tool in Automotive Design’, U Schramm, ATZ WorldWide,
pg 17-20, 100 (1998) 6.
‘Shape Optimization using Star-CD and HyperStudy’, M Ratzel and C E Baumann, STAR
American Conference, Detroit, May 2006.


HyperStudy 10 Training Manual

More Related Content

What's hot (20)

Viewers also liked (6)

Similar to HyperStudy 10 Training Manual (20)

More from AltairKorea (20)

Recently uploaded (20)

HyperStudy 10 Training Manual