Introduction to modeFRONTIER 4.5.

0 New Features

Summary

1. Workflow Enhancement

2. New Run Analysis and File System

3. New Design Space and RSM Enhancements

4. Algorithm Improvements and New Tools (ISF, Lipschitz, MOGT, MCDM, RELIABILITY)

Workflow Enhancements

Workflow Enhancement
1. New Layout

2. New Parameter Chooser

3. Subprocess & Scheduling Project

4. Star-CCM+ Node

5. Node Preferences Update

1. Workflow Layout Enhancement

Transparent Overlook and Logic Log panels

Properties panel now under Workflow Tree by default

Node Palette
Bookmark favorite nodes Node search text box

New Parameter Chooser File Management

Once file name is defined new options are available: Is relative: a file with the given name is expected to be connected to this node Embedded: if selected, the given template file is saved inside the prj Select Parameter Chooser

New Parameter Chooser from scratch

Select variables from the model and drag to right (or use + button above) If workflow is empty, new variables will be created with the same name

New Parameter Chooser from existing workflow

If workflow already contains some variables with same names, a prompt panel appears: You may need to link the existing ones You may need to create new parameters with different names

New Parameter Chooser from existing workflow

If workflow already contains some variables with same names, a prompt panel appears: You may need to link the existing ones You may need to create new parameters with different names

Scheduling Project: New mF Batch Node

In many cases, a multi-objective optimization problem can be transformed into a nested single-objective optimization (Hierarchical Games) Advantage: one of the two objectives may be obtained by a fast solver (this becomes the internal optimization/follower which can be repeated in loop in little time)

Fast solver

Heavy solver

Traditional (cooperative) approach: 320 simulations by MOGAII

Scheduling Project: New mF Batch Node

Nested (hierarchical) approach: 13 heavy solver simulations!

Scheduling Project: Data Nodes

Scheduling Project: Data Nodes (to transfer scalar, vector, string, files, DB)

Note: only files of the last design can be transferred (in this case it practically coincides with the best design; generally, the file of the best design may need to be reproduced in the main .prj)

In nested .prj, variables (scalar, vector, matrix, string) which are exchanged as data (constant) with external project can be defined by corresponding variables Files can be transferred in the same way (File Attachment Nodes)

Scheduling Project: Introspection

Scheduling Project: Multiple Optimization Steps Use Case

Selecting DesignDB option in the generic Buffer Node, users can transfer a complete database from one project to another In this way, users can first run a global search optimization, then apply a refinement

Scheduling Project: Multiple optimization steps use case

First .prj: extract Pareto Design> Entire Pareto Design DB

Second .prj: enter Buffer DB in Design Table and DOE table so that the NBI algorithm will run refinement starting from those points

Subprocess
Put a subprocess node inside the workflow Open it and select Edit Subprocess

Subprocess: Edit modeler

Define any workflow and save it as .prc file (can be reused by any other project) There are no scheduler, I/O and objs/cons: ONLY processes /data chains to be executed

Input parameter node (select data type)

Subprocess: Introspection to main .prj

After saving, exit and select this (or any other) prc file from Subprocess node Use Parameter Chooser to assign workflow I/O variables to Subprocess variables

Subprocess: Define Loops

If needed, select Enable Loops: the subprocess will be repeated under the specified logic loop until the prescribed condition is satisfied (any post-process expression may be defined if needed)

Subprocess: use case

A workflow containing different modules (CAD, structural analysis, etc.) Part of this workflow is common to other projects (the analysis is defined by the same node, here an EasyDriver), so we may use a subprocess for it In addition, we want to repeat the analysis until a result is obtained (to save a design in case of random errors, license server loss, etc..)

Subprocess: external prj

The external prj will only contain the I/O variables to be used in the optimzation and the part of workflow to be executed directly The subprocess node will instead contain the part of workflow to be executed by the subprocess

Subprocess: Use of Modeler

Open Modeler; import prj (the one normally used to run the solver, i.e. the second half of the complete prj)

I/O Parameters of Data node type are used here (their values should come from the external project) In addition, the file to be obtained by the external project is defined here via Input File Attachment Node The process is saved as a .prc file

Subprocess: Subprocess Node Proprieties

Select the .prc file created in Modeler

Use Parameter Chooser to link internal subprocess parameters to external .prj optimization variables

Subprocess: Loop Condition

Define as loop condition(After type) : exitPort =Fail if the exit condition of the subprocess is Fail, the loop cycle will continue This prevents design failure due to random problems; however a max number of loops can be specified (here 10)

Star-CCM+ Optimate MYNode

Drag and drop Optimate Node (MyNode tool from ESTECO-NA)

Star-CCM+ Optimate MYNode

Input Variable Introspection

Only internal parameters can be introspected: internal CAD, B.C. ...not external CAD

Star-CCM+ Optimate MYNode

Output Variable Introspection

Star-CCM+ Optimate MYNode

Autobuilt Workflow

Direct interface with STAR-CCM+ and external CAD

Optimization setup with external CAD and Optimate (STAR-CCM+)

Nodes Preferences

As of mF 4.5.0, Node preferences become PC settings and not User settings as before: they are saved in a folder visible and modifiiable by every user of that PC. This folder can be set with the variable all.users.home.dir from : C:\Programs\modeFRONTIER450\etc\jobagent\jobagent.properties Otherwise, mF uses the defaults : Windows = C:\ProgramData Linux = /usr/local/modeFRONTIER NOTE : Keep Alive option of every node has been moved in the Preferences menu

Notes about JVM

Keep Alive means that the Java Virtual Machine (JVM) that hosts the Integration software is the same between multiple executions. "Fast Launch by Process Fork" for almost every script node if you need fast execution: it does not isolate the integration software in a separate JVM so, if something goes wrong, you may also encounter some problems in mF.

GRID Improvements

1. port specification for firewall setups (to be set on every pc node): etc\jobagent\jobagent.properties 2. more than 1 prj can run on the same grid (and so on the same integrations) 3. Nodes can be added / modified on the fly during the Run

New Run Analysis and File System

New Run Analysis Main Components

Summary and Overall Designs Browser

Dashboard Management: Single Design Detailed View

Layout (columns) Add/Remove Tabs

Dashboard
Gadget Palette Common Legend

New Run Analysis Browse/Filter

Summary and Overall Design Browser Optimization progress
Overall design view - Design status (e.g. error, unfeasible, feasible) Design Filtering action (selection area) Summary Information: Total designs, Percentages per Design Type

Single Design Detailed View

Filter Panel: based on ID interval or last generated designs One-click design exploration (update Info Designs and related gadget, selection on charts) Zoomable/Browsable Design List Star/Stop playing function Error/out log Direct connection to application files (open folder and/or open shell inside application working directory)

New Run Analysis Log Gadgets

Files or Images

Info Project

Designs

Scheduler

Process Table

Design Data

New Run Analysis Chart Gadgets

History Chart for one or more variables

Scatter Chart for 2, 3, 4 Variables

Broken Designs Chart

Summary Pie Chart

New File System Option

The New File System option offers compact easy navigation that complements the classic view. The dimension tree has been reduced in terms of number of directories: one main log and one main proc category instead of repeated log and proc directories for each job related folder

CLASSIC

NEW
Run Options: new feature to select which files to keep after the job execution (never, always, or not on failure conditions)

New Design Space and RSM Enhancements

New Design Space Layout

Tool Launcher

Enhanced Explorer Tree

Chart Palette Bookmark Favorite Charts Chart Search Textbox

Design Space Explorer Tree

Different Tree Views Hierarchical By Family Preview

Filtering options

Fully Comprehensive Tree

Design Space - Clustering

- Hierarchical Clustering an additional step introduced after the Run Algorithm step: Dendogram Chart allows you to select the number of clusters to be applied to the Clusters Table without having to return to the Browse menu (cancelled): more user-friendly, reduction of clicks - number of clusters visible under the relevant hierarchical clustering function in the Category Tree

- Partitive Clustering - an additional step introduced after the Run Algorithm step: DB Index chart K-Means function can be directly applied to the Clusters Table without having to return to the Browse menu (cancelled): more user-friendly, reduction of clicks

Design Summary chart

The 2 Pie charts have been joined together

Correlation Matrix Chart

RSM Enhancements RSM node new parameter chooser RSM Wizard enhancements RSM Validation

RSM Node
Before

After

- Parameter Chooser; new variable nodes are created if the desired inputs/outputs are not present in the workflow; variable nodes are linked if the desired inputs/outputs exist (linked or not to the node)

RSM Wizard Improvements

One Tool for All (merge Single RSM and Multiple RSM Tools) User-friendly enhancements (less clicks and steps) Automatic removal of outliers Repeated measurement handling (arithmetic mean considered)

RSM Wizard Improvements

New Function Tree in Design Space
- RSM summary report - possibility to choose either a report with only algorithm settings and log, or with everything (table, charts, algorithm settings and log); you can also choose the relevant table - RSM in the Category Tree: divided by output within each output you can find the functions related to it, with the selected inputs next to function name (e.g. {x,y}) - By right-clicking on a function you can create a new function of the same category

User-friendly enhancements (less clicks and steps)

- RSM Function creation on the same page as algorithms and input/output selection (inputs are pre-selected) -the Plugin Description panel shows the settings for any function (only those of a new function may be modified)

Response Surface Validation

Directly select Enable RSM validation tool: first train RSM on a table then select another table for validation At the end, look at validation table

Alternatively, select RSM validation tool from the design space, select RSM to validate, and look at table

Response Surface Validation

For each RSM, the following information is reported: Mean absolute error Mean relative error Mean normalized error R-Squared error AIC

New RSM: DACE-Kriging and SS-ANOVA

- New RSM algorithm DACE-Kriging introduced at the specific request of Honda - At present DACE-Kriging will be distributed to version 4.4.2 users as a plugin, together with the necessary documentation, benchmarking and instructions for its integration into the above version; regularly available as of version 4.5 - SS-ANOVA is now also available as a stand-alone RSM

Variable Screening Tool SS-ANOVA

What is SS-ANOVA? Smoothing Spline ANOVA is a statistical modeling algorithm based on a function decomposition similar to the classic analysis of variance (ANOVA) decomposition and the associated notions of MAIN effect and INTERACTION. Each term main effects and interactions reveals a measure of its contribution to the global variance. Given a dataset, SS-ANOVA detects important variables. When to use it? To better understand the model. To reduce the input variables of the problem to train RSM and to run an optimization algorithm with less effort. To improve inaccurate RSM.

Cubic Smoothing Spline ANOVA case univariate

The Smoothing Spline ANOVA is the solution to the problem:

1 1 n 2 2 ( min f i f ( xi )) + [ f ( x )] dx 0 f n i =1

It corresponds to the usual natural cubic spline.

-The left term guarantees a good fit to the data. - The right term represents a penalty on roughness.

General Outline for Screening Usage (1)

Perform a variable screening.

If possible, perform interaction screening. Verify Collinearity Indices. If at least one is much greater than 1, the screening analysis is bad. Stop here (sampling is not adequate). Set a filter. Perform another Screening Analysis, if necessary. The set of important variables are plotted in the cumulative chart or are printed in bold in the RSM Functions Creation.

SS-ANOVA as RSM stand-alone (2)

The internal optimization routine finds the lambda and theta values and also minimizes the GCV score (collinearity index)

Algorithm Improvements and New Tools

Algorithm Improvements
New scheduler arrangements ISF Lipschitz MOGT MCDM New MORDO distributions RELIABILITY

Scheduler Arrangements
The list of scheduler algorithms has been re-arranged by algorithm type: the new categorization is much more user-friendly, maintaining numeric consistency

Incremental Space Filler

Incremental Space Filler

existing points in the database (previously generated designs)

new points are added in order to fill the space uniformly

Periodic Boundary Conditions (PBC)

In the new ISF, the periodic boundary conditions (PBC) have been introduced in ISF GA variant. Distances of points closed to boundaries are computed as if boundaries were periodic In this away, it avoids placing too many points close to boundaries

Lipschitz
Local Lipschitz constant could be used as a complexity indicator Large Lipschitz constant Small Lipschitz constant The design space is tessellated in zones and Lipschitz constants are estimated locally (higher where gradients are higher: more points needed)

Lipschitz
Number of designs that will be evaluated If activated, points will be added only inside spheres centered on marked points (radius is fraction of input range) Definition of which variables are to be used for analysis

New option: A fraction of the designs can be produced by ISF (to avoid excessive accumulation in highest Lipschitz constant areas)

SSE as alternative to generational for the internal optimization

Lipschitz: Exploration Fraction Option

MOGT: Multi-Objective Game Theory Algorithm

SIMPLEX1 is run (Player 1) Obj.: min. f1 Var: X , Y0 fixed SIMPLEX2 is run (Player 2) Obj.: min f2 Var: Y, X0 fixed

Calculation of f1 for every configuration

n1 iteratons

Calculation of f2 for every configuration

n2 iterations

Best X= X1 is found

Best Y= Y1 is found

SIMPLEX1 is run (Player 1) again Optimise X with Y fixed to Y1

SIMPLEX2 is run (Player 2) again Optimise Y with X fixed to X1

A converged optimized solution (XN , YN)=(XN-1 , YN-1 ) is found

Algorithm Improvements: New MOGT

The adaptive decomposition of the variables can also be made through SS-ANOVA as an alternative to t-Student SS-ANOVA is efficient for identifying the significant variables even when the database is small and when Full Factorial isnt used (as required by t-Student) Another t-Student limitation was that when the database grew, the significance of any parameters tended to reach 100%

New MOGT

Why MCDM ?
Ranking between alternatives is a common and difficult task. Multiple Criteria Decision Making (MCDM) is a process finalized to solve decision problems involving multiple and conflicting goals During this process there are different actors:
1. The Decision Maker (DM) chooses one reasonable alternative from among a limited set of available ones; 2. Alternatives are the possible solutions; 3. Attributes are parameters that the DM uses to make a decision.

Ranking between available alternatives

modeFRONTIER MCDM: Attribute Selection

Select attributes and goals

Select designs

modeFRONTIER MCDM: Algorithms

Five different MCDM algorithms are available:

Linear MCDM GA MCDM Hurwicz MADM Savage MADM AHP

List of available algorithms Algorithm parameters

modeFRONTIER MCDM: Utility Function

modeFRONTIER MCDM allows the correct grouping of outputs through the definition of a single utility function. The utility function is coherent with the preferences expressed by the user (weights on attributes are created) A ranking is defined based on utility function values

Preference Indifference Margin

You can set the Preference and Indifference thresholds. Preference threshold identifies designs considered dominated (worst ranked) Indifference threshold refines the non-dominated design ranking by highlighting the best ranked designs (green)
Preference occurs when difference of rank value is higher than 0.15 (designs in red)

Tolerance occurs when difference of rank value is higher than 0.09 (designs in yellow)

Raising preference (0.5) red preferred designs are reduced; tolerance is still 0.09

73

New Distributions
New probability distribution functions are available for Robust Design analysis (including their link to Polynomial Chaos) Available built-in distribution types: Uniform Normal Logistic Chi-square Exponential Log-Uniform Student

MORDO: Resample Repeated Designs Option

When this option is checked and a repeated design is found, the sample is generated again and statistical properties are computed again (e.g., test on same design changing PC order, etc.) During an optimization, option can be unchecked to avoid repeating performance analysis for repeated designs

Need for Reliability-Based Optimization: Automotive Example

In a design problem under uncertainties, there is a clear need to define objs/cons on a given percentage of the performance distribution (Reliability Analysis)

30 crash test parameters subject to uncertainties

Goal: 90% of tests are below NCAP limitc

90% best NCAP limit
RIB MID
30 25 20 15 10 5 0 0.005 0.007 0.009 0.011 0.014 0.016 0.018 0.021 0.023 original best

90% original

The optimization which aims to find an optimal design and minimizes the failure probability is called Reliability-Based Design Optimization We introduce a RBDO methodology based on Polynomial Chaos Expansion

Generalized Polynomial Chaos Theory

Any output function F of deterministic variables (x,t) and uncertain variables x can be expressed as spectral expansion of polynomials orthogonal w.r.t probability function of the uncertain variables (Hermite for Gaussian distr.)

F (x, t , ) = i (x, t ) i ( )
i =0

Statistical moments may be found easily in function of unknown weights:

E PC (F ) = F = 0 (x, t ) VarPC (F ) = F = i (x, t ) i
2 2 i =1 N 2

To find the unknown i coefficients and finally express the moments we must sample F in N points to minimize:

Hermite Polynomials

F ( ) ( )
N k j =1 j i =0 i i j

Efficiency of PC Sampling: Exponential Convergence

The advantage of PCE methodology is that the convergence to exact distribution moments follows an exponential rate: accurate and fast

Reliability-Based RDO with PC Sampling

SMALL SAMPLING
(for each design proposed by the external RDO optimization)

F (x, t , ) = i (x, t ) i ( )
i =0

Polynomial Chaos interpolation of the performance function

The % relative to the given failure region is extracted accurately

G(u)<0 G(u)>0

FULL MONTECARLO
(large sampling obtained analitically)

Application: Reliability Optimization of a Boomerang Throw

Boomerang trajectory is computed by a Matlab script, solving motion equations by Runge-Kutta Aerodymical forces are provided by a Response Surface (meta-model) trained by a series of different CFD analysis (changing velocity and angle of attack) All the simulations have been executed by modeFRONTIER*, including the CAD parameterisation and the optimization to find optimal geometry

*R. Russo, A. Clarich, C. Poloni, E, Nobile, Optimization of a Boomerang shape using modeFRONTIER, AIAA Proceedings, Indianapolis, September 2012

Application: Problem Definition

Launch parameters are uncertain We want to optimize their nominal value to minimize the 99-ile of the returning distance At the same time, a second objective is given to the maximization of range (average)

Input variables Velocity (V) Spin Aim angle Tilt Angle Objectives Returning distance RD 99-ile Range

Range of variation [5-30]m/s [0-10] Hz [0-30] [0-50]

Uncertainty (standard deviation) 2m/s 1Hz 2 2 Goal

Minimize 99-ile of RD Maximize average value

Reliability Settings in mF

Set percentile(s)

Use when defining objectives

Application: Optimization Results

Optimal range

Optimal return

Input variables Velocity (V) Spin Aim angle Tilt Angle Objectives Returning distance RD 99-ile Range

Optimal range 21.6 m/s 4.98 Hz 4.2 20.1 8.5m 31.4m

Optimal Return 21.7m/s 4.92Hz 4.2 7.2 2.9m 21.7m

Application: Optimization Results

Optimal range Optimal return

Average range: 33.4m 99-ile of return: 8.5m

Average range: 21.7m 99-ile of return: 2.9m