17
AUTOMATED POSTPROZESSING OF MULTIMODEL OPTIMIZATION DATA Markus Schemat, BMW Group ALTAIR TECHNOLOGY CONFERENCE MUNICH 2014

Automated Post-processing of Multi-Model Optimization Data

Embed Size (px)

DESCRIPTION

Due to more compressed timelines in the product development process there is an increasing demand regarding CAE simulations. The increase of HPC resources in combination with a massive parallelization more complex simulations could be performed in a proper timeline. Numerical optimization (in combination with CAE simulation) is such a complex process because of the resource requirements in combination with an iterative solution scheme. In order to handle different simulation disciplines so called Multi Model Optimization (MMO) and Multi Disciplinary Optimization (MDO) optimizations are getting more important. Unfortunately these calculations will lead to large result data and demanding hard- and software requirements. Taking this into account there are two major aspects which need to be addressed in the future. First of all the analyst needs to get as much information out of the simulation (optimization) in order to physically understand the CAE model. For this task optimization is a well suited utility as the numerical processes could handle a large amount of design variables and different constraints efficiently. But what needs to be taken into account is a proper visualization of this data. Secondly it will be still important to keep the timelines in the development process even with a complex optimization task. To reach this target it is necessary to force the optimizer to converge in a few iterations or to react on changing circumstances in the development process. In this paper an automated approach BMW is currently applying will be described which is addressing the two above points for linear CAE simulations.

Citation preview

Page 1: Automated Post-processing of Multi-Model Optimization Data

AUTOMATED POSTPROZESSING OF MULTIMODEL OPTIMIZATION DATA

Markus Schemat, BMW Group

ALTAIR TECHNOLOGY CONFERENCE MUNICH 2014

Page 2: Automated Post-processing of Multi-Model Optimization Data

POSTPROCESSING OF MMO DATA ABSTRACT

Title Automated Postprozessing of MultiModel Optimization Data

Author Markus Schemat, BMW AG

Daniel Heiserer, BMW AG

Company

BMW Group

Knorrstrasse 147

80937 München, Germany

Abstract

Due to more compressed timelines in the product development process there is an increasing

demand regarding CAE simulations. The increase of HPC resources in combination with a massive

parallelization more complex simulations could be performed in a proper timeline. Numerical

optimization (in combination with CAE simulation) is such a complex process because of the

resource requirements in combination with an iterative solution scheme. In order to handle different

simulation disciplines so called Multi Model Optimization (MMO) and Multi Disciplinary Optimization

(MDO) optimizations are getting more important. Unfortunately these calculations will lead to large

result data and demanding hard- and software requirements. Taking this into account there are two

major aspects which need to be addressed in the future.

First of all the analyst needs to get as much information out of the simulation (optimization) in order

to physically understand the CAE model. For this task optimization is a well suited utility as the

numerical processes could handle a large amount of design variables and different constraints

efficiently. But what needs to be taken into account is a proper visualization of this data.

Secondly it will be still important to keep the timelines in the development process even with a

complex optimization task. To reach this target it is necessary to force the optimizer to converge in a

few iterations or to react on changing circumstances in the development process.

In this paper an automated approach BMW is currently applying will be described which is

addressing the two above points for linear CAE simulations.

7th European ATC, 25.06.2014 Page 2

Page 3: Automated Post-processing of Multi-Model Optimization Data

7th European ATC, 25.06.2014 Page 3

POSTPROCESSING OF MMO DATA OPTIMIZATION IN THE PRODUCT DESIGN PHASE

Developme

nt

Change

Cost

Product

Knowledge

Product Development

Timeline

Architecture

Phase

Concept

Phase

Series

Phase

SOP

Design

freedom

What is an optimal solution for an

architecture?

What is an optimal solution for multiple CAE

models?

How to approach an optimal solution in best

time?

Is there a better (global) optimum?

Optimization Potential

To answer this questions the presentation is covering a method which is combining two

major approaches (Multi Model Optimizations and Solution Spaces). Both methods are

applied to gradient based optimizations.

Page 4: Automated Post-processing of Multi-Model Optimization Data

7th European ATC, 25.06.2014 Page 4

POSTPROCESSING OF MMO DATA RESPONSE DATA QUERY INFORMATION GAIN

Method Time User

Activity

Informatio

n Density

Optimum

Global/Loca

l

Discrete ++ -- - --

Stochastic -- + +

Optimization (Gradient

based)

+ ++ +

Solution Space + ++ ++

Distributio

n

Correlatio

n Postprocessi

ng

Datamining

Page 5: Automated Post-processing of Multi-Model Optimization Data

7th European ATC, 25.06.2014 Page 5

POSTPROCESSING OF MMO DATA DEFINITION OF SOLUTION SPACE

The solution space is representing the feasible region of the current design.

It is based on the performed CAE simulations.

The real physical solution space is continuous vs. the linear ones derived from the

simulation.

The solution space is limited by design boundaries and response constraints.

Change of

Solution Space

Number of

CAE Models

Constraint 1

Design variable 1

Architecture 2…N

MDO 2…N

Multimodel (MMO)

Feasible

Solution

Space

Page 6: Automated Post-processing of Multi-Model Optimization Data

7th European ATC, 25.06.2014 Page 6

POSTPROCESSING OF MMO METHODOLOGY TO OBTAIN THE SOLUTION SPACE

Method Pro Con

Stochastic Equal sample distribution

Parallel job submission

No local optima

Large # samples

No optimal solution

Optimization (from

Baseline)

Small # of samples

Parallel job submission

Minor chance of local

optima

Unequal sample

distribution

Optimization (Sequential) Minimal # of samples Only local samples

Sequential job submission

Potential for local optima

Constraint

0.28 Constraint

0.0 Constrain

t

-0.04

Constraint

-0.13

Constraint

-0.185

Constraint

-0.28

Page 7: Automated Post-processing of Multi-Model Optimization Data

Model #n Approximat

e Model

F06

DESVAR

#n

PCH

OP2

External

Parser

Post

processin

g

Structural

Analysis

Sensitivity

Analysis

Approximat

e

Optimizatio

n

Improve

Design

Nastran - SOL200

Model #1 Approximat

e Model

F06

DESVAR

#1

PCH

OP2

External

Parser

Post

processin

g

Structural

Analysis

Sensitivity

Analysis

Approximat

e

Optimizatio

n

Improve

Design

Nastran - SOL200

7th European ATC, 25.06.2014 Page 7

POSTPROCESSING OF MMO OPTIMIZATION SETUP – SINGLE MODEL

DESVAR

#1

DESVAR

#n ≠

Independent solution for each model

Page 8: Automated Post-processing of Multi-Model Optimization Data

7th European ATC, 25.06.2014 Page 8

POSTPROCESSING OF MMO OPTIMIZATION SETUP – MULTIMODEL

Model #1

F06

DESVAR

#1

PCH

OP2

External

Parser

Post

processin

g

Structural

Analysis

Sensitivity

Analysis

Approximat

e

Optimizatio

n

Improve

Design

Nastran - SOL200

Approximat

e Model

Model #n

F06 DESVAR

#n

PCH

OP2

External

Parser

Post

processin

g

Structural

Analysis

Sensitivity

Analysis

DESVAR

#1

DESVAR

#n =

Linked solution for all models

Model #1

DESVAR

#1

Outp

ut

External

Parser

Post

processin

g

Structural

Analysis

Sensitivity

Analysis

Approximat

e

Optimizatio

n

Improve

Design

Nastran - SOL200 / Optistruct (V13)

Approximat

e Model

Model #n

DESVAR

#n Outp

ut

External

Parser

Post

processin

g

Structural

Analysis

Sensitivity

Analysis

Page 9: Automated Post-processing of Multi-Model Optimization Data

Component

Target

Global Stiffness

Eigen Modes

Energy Absorption

Weight

Eigen Modes

Dynamic Stiffness

Vibration (FR)

Stiffness

Eigen Modes

Cost

Page 9

POSTPROCESSING OF MMO DATA EXAMPLE- REQUIREMENT MANAGEMENT

Requirement management

Car Positioning

Dis

cre

tisation

of D

eve

lop

me

nt

Ta

rge

ts

Customer related

Functions and

Properties

Car A

rchite

ctu

re

BIS, TOP3 …

Responsibl

e

Department

Function

EG Active Safety

Ergonmics

Driver

Assistance

System

Fatigue Strength

EK Comfort Interior

BIW Functions

Corrosion

Protection

TI Manufacturing

Technology

Producibility

EF Driving

Dynamics

… … Ta

rge

t T

ran

sla

tion into

En

gin

ee

ring

7th European ATC, 25.06.2014

Question:

Do the component targets (eigen frequency of the steering column on

its own)

correlate with the global target (vibration at the steering column in the

BIW)?

Page 10: Automated Post-processing of Multi-Model Optimization Data

7th European ATC, 25.06.2014 Page 10

POSTPROCESSING OF MMO DATA GUI - HYPERVIEW CUSTOMIZATION

The MMO postprocessing is implemented as a user customization via the preferences file in the

Hyperview GUI.

Page 11: Automated Post-processing of Multi-Model Optimization Data

7th European ATC, 25.06.2014 Page 11

POSTPROCESSING OF MMO DATA GUI – DATA IMPORT AND CONVERSION

The import of data is defined via a directory

selection

Folders could be selected individually or by

scanning of subfolders.

All supported output files will be recognized

automatic and loaded into a database

(Conversion)

Supported Formats

Optistruct (*.out)

Nastran SOL200 (*.f06)

Page 12: Automated Post-processing of Multi-Model Optimization Data

7th European ATC, 25.06.2014 Page 12

POSTPROCESSING OF MMO DATA GUI - CONFIGURATION OF PLOT DATA

Response selection is performed via the curve labels

Plotting could be done for:

Response versus response

Response versus desvar/desprop

Responses/desvars of different MMO models

A filter to preselect the responses/desvars could be applied

An internal constraint is available to reduce the database values and predict influences of design

changes

model2_LS+TR:

model1_LS :

model3_LS+TR+RK:

Found MMO

Results

Page 13: Automated Post-processing of Multi-Model Optimization Data

7th European ATC, 25.06.2014 Page 13

POSTPROCESSING OF MMO DATA GUI – DATA ANALYSIS PLOTTING

The prior configured responses are plotted in a 2D plotting client and shown in a tabular format

The solution space is created automatically for the plotted data

For additional information notes with the response values could be attached to desired points of

interest

Multiple notes could be generated

Filtering of notes is implemented

Intermediate Points could be interpolated based on the neighboring points and exported as new

designvariables

Page 14: Automated Post-processing of Multi-Model Optimization Data

7th European ATC, 25.06.2014 Page 14

POSTPROCESSING OF MMO DATA GUI – DATA ANALYSIS ANIMATION

Each design iteration could be visualized into the animation client

The postprocessing is available for

Individual optimizations

Solution space

To review the design changes an animation mode is available

Final or in between results could be exported to an PPT report.

Page 15: Automated Post-processing of Multi-Model Optimization Data

Non Feasible

Designspace

7th European ATC, 25.06.2014 Page 15

POSTPROCESSING OF MMO DATA EXAMPLE- REQUIREMENT MANAGEMENT

Example Summary

a. Even without fulfilling local

constraints (eigen frequency

targets) the overall vibration target

could be achieved.

b. The baseline is not weight optimal

c. 3mm/s is the physical limit of the

Response

d. Fixing the cross member is limiting

the weight potential drastically

(blue region)

Maxval = 3

mm/s

Maxval = 12

mm/s

Base

Thickness

Page 16: Automated Post-processing of Multi-Model Optimization Data

7th European ATC, 25.06.2014 Page 16

POSTPROCESSING OF MMO DATA DATA HANDLING AND PROZESS FLOW

Externa

l Parser

(Conver

t)

Plotting

Animatio

n

Export

Model #1

DESVAR

Outp

ut Structural Optimizer

Model #n

Outp

ut

MMO #1

DCONST

#1

Model #1

DESVAR

Outp

ut Structural Optimizer

Model #n

Outp

ut

MMO #n

DCONST

#2

Scripted

Plotting

DB

Animati

on

DB

Preprocessin

g

Solving Postprocessi

ng

Solution Space GUI in Hyperview

Filter

Data

Constrain

Data

Add new design

loops

Page 17: Automated Post-processing of Multi-Model Optimization Data

7th European ATC, 25.06.2014 Page 17

POSTPROCESSING OF MMO DATA CONCLUSION / Q&A

Conclusion

The described method is showing great potential in combining optimization techniques.

Multi model optimizations are extremely helpful for architectural decisions

The data handling is working well with the Hyperview customization

Outlook

Currently only gauge is supported in the postprocessing

Integration of further CAE disciplines

Export and import of Excel datasheet

Interpolation with gradient based information instead of interpolations