Numerische Simulation in der Automobilindustrie: Big Data

© Fraunhofer SCAI © SIDACT GmbH

Numerische Simulation in der Automobilindustrie: Big Data in der industriellen Praxis

Clemens-August Thole


SIDACT

Fraunhofer Spin-Off starting January 1st, 2013

FEMZIP – Compression of simulation results

DIFFCRASH – Robust Design: Identification of areas in geometry

causing scatter of simulation results

SDMZIP – Compression of sets of simulation results

Marketing, support, new application areas

5 full time, 4 students


HPC Systeme und Archive

DWD 34.000 Cores 3 PByte

ECMWF 160.000 Cores 40 PByte

DKFZ 36.000 Cores 20 PByte

Automobilhersteller 10-80.000 Cores ~1 (6) PByte

• Viele keine unabhängige Systeme

• 32 (Crash) – 512 (CFD) cores pro Berechnung, max. 1 Nacht pro Simulation

• Die meiste Rechenzeit für Crash und CFD

• 4 – 40 GByte pro Simulationsergebnis (Akustik auch mehrere Tbyte)

• Ergebnisse werden meist nach ein paar Monaten gelöscht


Anwendungsbereiche numerischer Simulation

Strukturmechanik

Steifigkeit, Vibrationen, Wärmeübertragung

Fahrdynamik

Innenströmungen

Komfort, Kühlung, Verbrennung

Außenströmungen

Kühlung im Motorraum, Luftwiderstand, Verschmutzung von Heckscheibe, Funktion Scheibenwaschanlage

Crashsimulation

Simulation von Produktionsprozessen

Umformen, Walzen, Beschneiden, Gießen…

Prozessketten

Elektrische Schaltkreis Simulation

Elektromagnetische Kompatibilität

Kopplung verschiedener Disziplinen


Marktbedeutung von Numerischer Simulation http://www.cimdata.com/de/

Based on CIMdata estimates, the Mainstream PLM market grew

12.6% in 2012 to $21.1 billion. Once again, simulation and analysis

(S&A) was the fastest growing segment, showing the increasing

importance of S&A across the product development life cycle.

Zum Vergleich

IBM: $19,600

Mitarbeiter: 380.000


Softwarehersteller

• CAD

• Gittergenerierung

• Numerische Software

• Postprocessing

Dassault Systems (F) CATIA

Siemens Pro-E (D, USA)

Beta (GR) – Ansa

ANSYS (USA, GB), MSC (USA),

ESI (F), CD-Adapco (USA, GB),

Simulia (F, USA), LSTC (USA),

MAGMASOFT(D), Intes (D)

GNS (D), Altair (USA), Beta (GR)

11000

300

2000

1000

200

400


FEMZIP - mp3 for simulation results

CRASH FEMZIP-L for LS-DYNA™

FEMZIP-P for PAM-CRASH ™

FEMZIP-R for Radioss™ Explicit

FEMZIP-ERFH5 for PAM-CRASH™ new format (beta)

FEMZIP-a4db for GNS Animator4 database

CFD (FEMZIP-ENSG for EnSight™ Case Gold)

(FEMZIP-CCM for StarCD™ (.ccm data format))

FEMZIP-OF for OPENFOAM (beta, incremental and distributed data handling)

NVH FEMZIP-N for OP2 (NASTRAN™ / Radioss™ Implicit)


Data Compression

• Floating-point data cannot be efficiently compressed losslessly:

» A compression factor of only 1.2 is

obtained

» The solution is FEMZIP

Airflow simulation around a car:

6 variables, 43 million elements,

21 time steps


Lossy Data Compression

• Quantization Floating point data is rounded to a given precision and mapped into the integer domain.

• Prediction A bijective transformation is performed to prepare those integers for encoding purposes (reduce entropy).

• Encoding Afterwards an entropy encoding method is used which removes redundant bits.

» FEMZIP achieves compression factors of about 10!

13534,11289527977912437

32/64-Bit float

011010110110


Result of ensemble forecast

13.11.1993

Simulation

Analysis

10% probability

bounds


Initial position of singular vectors in winter 92/93


Buckling

Knickstab

Tubes

Buckling mode as result

of linear structural analysis


DIFFCRASH : Stability analysis

• Postprocessing tool

• Statistical Analysis of full simulation models

• Basic analysis methods

• Correlation based methods

• PCA based methods


Difference PCA (Patent beantragt)

State 0

State 8

State 34

../../../../BACKUP_TERAN/Users/rteran/DIFFU/DIFF_ANALYSIS/VW/Bauteil_Disp.exe





Thank you for your attention!

Clemens-August Thole

Documents

Numerische Simulation in der Automobilindustrie: Big Data