Post on 21-Mar-2020
1
Webinar
Parameter Identification
with optiSLang
Dynardo GmbH
2Webinar Parameter Identification with optiSLang
© Dynardo GmbH
Technical Notes
Audio: • If you cannot hear the speech, please check menu Communicate –
Audio Test and Broadcast(there is no audio conference)
• In the invitation email, you find phone numbers to call in
Privacy: • The Webinar may be recorded, but only our presentation
(Participants cannot be heard)• Participants cannot see each other
Questions?• If you have any questions or remarks, please type them into the
window Questions & Answers at any time• The presenter will read them out and answers for all on occasion
3Webinar Parameter Identification with optiSLang
© Dynardo GmbH
Outline
Theoretical background
Process Integration
Sensitivity analysis
Least squares minimization
Examples:
4Webinar Parameter Identification with optiSLang
Theoretical Background
© Dynardo GmbH
5Webinar Parameter Identification with optiSLang
© Dynardo GmbH
Inverse Identification of Model Parameters
• Identification of unknown model parameters by the calibration of the model with respect to given measurements
• Direct relation between measurements and model parameters is known only inversely as forward simulation model
6Webinar Parameter Identification with optiSLang
© Dynardo GmbH
The Forward Simulation
• For given set of model parameters p the model responses y can be calculated with a given simulation model
• Deviation of model responses and measurements y* can be evaluated
• For which parameter set popt model responses and measurement agree sufficiently well?
Model parameters Simulation model Model responses
Measurements
7Webinar Parameter Identification with optiSLang
© Dynardo GmbH
Least Squares Minimization
• The likelihood of the parameters is proportional to the conditional
probability of measurements y* from a given parameter set p
• For correct model (y* - y) is caused only by measurement errors
• Assuming normally distributed measurement errors:
• If the errors are independent we obtain
• With constant standard deviation the objective simplifies
8Webinar Parameter Identification with optiSLang
© Dynardo GmbH
Requirements to the Identification Procedure
• The simulation model needs to represent the main physical behavior
(systematic model errors are not considered)
• Since the least squares minimization may lead to a local optimum a
global optimization strategy is necessary
• Only sensitive parameters can be identified
• Different parameter combinations may lead to a similar objective
Uniqueness of identified parameters has to be assessed
9Webinar Parameter Identification with optiSLang
© Dynardo GmbH
• use scalar values or signals inside ANSYS Workbench
• identify which parameters have influence and
can be calibrated
• match experimental data with simulation
Model CalibrationModel update to increase your simulation quality!
Question?
10Webinar Parameter Identification with optiSLang
Process Integration
© Dynardo GmbH
11Webinar Parameter Identification with optiSLang
Process Integration
Parametric model as base for
• User defined optimization (design) space
• Naturally given robustness (random) space
Design variablesEntities that define the design space
Response variablesOutputs from the system
The CAE processGenerates the results according to the inputs
Scattering variablesEntities that define the robustness space
© Dynardo GmbH
12Webinar Parameter Identification with optiSLang
optiSLang Integrations & Interfaces
Direct integrations ANSYS Workbench MATLAB Excel Python AMESim SimulationX
Supported connections ANSYS APDL Abaqus Adams AMESim …
Arbitary connection ofASCII file based solvers
Signals can be directly imported from MATLAB, Excel, Python, AMESim, SimulationX & ASCII
© Dynardo GmbH
13Webinar Parameter Identification with optiSLang
Signals in optiSLang
• Signals are vector outputs having an abscissa (e.g. time axis)
and several output channels (e.g. displacements, velocities)
• Signal functions enables the user to extract local and statistical
quantities and to analyze differences between several signals
• Match signal data (curves) with Signal Processing
© Dynardo GmbH
14Webinar Parameter Identification with optiSLang
Signal Processing – Definition of Signals
• The ETK node enables the
definition of several solver
and reference signals
• Reads many CAE binary output
formats and text files
• Can read signals, vectors
and matrices
• Instant visualization of
vectors and signals
© Dynardo GmbH
15Webinar Parameter Identification with optiSLang
© Dynardo GmbH
Sensitivity Analysis
16Webinar Parameter Identification with optiSLang
© Dynardo GmbH
Automatic workflow
with a minimum of solver runs to:
• identify the important parameters for each response
• Generate best possible metamodel (MOP) for each response
• understand and reduce the optimization task
• check solver and extraction noise
Understand the most important input variables!
Sensitivity Analysis
17Webinar Parameter Identification with optiSLang
© Dynardo GmbH
Least Squares Minimization
18Webinar Parameter Identification with optiSLang
© Dynardo GmbH
Definition of objective
• Monotonous increasing of abscissa
19Webinar Parameter Identification with optiSLang
© Dynardo GmbH
• Hysteresis Curve
• Decomposition in load and unload
• Additional terms for max force and intersection with x-axis
Definition of objective
20Webinar Parameter Identification with optiSLang
© Dynardo GmbH
• 2 load-curves for 1 material behavior
• Weighting of different experiments in one objective function by
normalizing the RMSE by the response ranges (or standard deviations)
Definition of objective
21Webinar Parameter Identification with optiSLang
© Dynardo GmbH
optiSLang Optimization Algorithms
Gradient-based Methods
• Most efficient method if gradients are accurate enough
• Consider its restrictions like local optima, only continuous variablesand noise
Adaptive Response Surface Method
• Attractive method for a small set of continuous variables (<20)
• Adaptive RSM with default settings is the method of choice
Nature inspired Optimization
• GA/EA/PSO imitate mechanisms of nature to improve individuals
• Method of choice if gradient or ARSM fails
• Very robust against numerical noise, non-linearity, number of variables,…
Start
22Webinar Parameter Identification with optiSLang
Decision Tree for Optimizer Selection
• optiSLang automatically suggests an optimizer depending on the
parameter properties, the defined criteria and user specified settings
© Dynardo GmbH
Question?
23Webinar Parameter Identification with optiSLang
Examples
© Dynardo GmbH
24Webinar Parameter Identification with optiSLang
© Dynardo GmbH
Examples
Identification of:
1. the geometry parameters of a press contact
2. material parameters of spring steel
3. material parameters of sandstone
4. fracture parameters of concrete
5. hyperelasticity parameters of an OGDEN law
6. the geometry parameters of a cantilever beam
25Webinar Parameter Identification with optiSLang
© Dynardo GmbH
1st example: Press fit contact
• Finite element model in ANSYS Workbench
• Variation of geometry parameters
• Reaction forces as Insertion Force and
Pull out Force
26Webinar Parameter Identification with optiSLang
© Dynardo GmbH
1st example: Problem Definition
-40
-30
-20
-10
0
10
20
30
40
50
60
70
0 0,0002 0,0004 0,0006 0,0008 0,001 0,0012 0,0014 0,0016
Forc
e [
N]
Time
Insertion Force
Pull out Force
desired behavior
initial simulation
• Simulation with initial geometry parameters vs. reference (desired behavior)
27Webinar Parameter Identification with optiSLang
• Finite element model in ANSYS Workbench
• Nonlinear material behavior
• Tensile bar is deformed by a
predefined displacement
• Reaction forces at deformed tensile bar end (1)
are monitored depending on deformation
between named selection u1 (2) and u2 (3)
and saved into the result file file.rst
1.3.2.
© Dynardo GmbH
2nd example: Tension Test of Spring Steel
28Webinar Parameter Identification with optiSLang
2nd example: Problem Definition
• Simulation with initial materials parameters vs. reference (measurements)
© Dynardo GmbH
29Webinar Parameter Identification with optiSLang
2nd example: Problem Definition
• Identification of the material parameters to optimally fit the
force-displacement curve to the measurements
• Unknown material parameters for
nonlinear isotropic hardening (nliso):
• Young´s modulus
• Yield stress σ0
• Linear hardening coefficient R0
• Exponential hardening coefficient R∞
• Exponential saturation parameter b
• Objective function is the sum of squared errors
between the reference and the calculated
force-displacement function values
σ = σ0 + R0εpl + R∞ (1-e-b εpl)
© Dynardo GmbH
30Webinar Parameter Identification with optiSLang
2nd example: Task Description
• Generation of a solver chain using ANSYS Workbench
and Signal Processing
• Definition of the input parameters
• Definition of output and
reference signals
• Sensitivity analysis of signal
extraction terms using
the given parameter bounds
• Single objective, unconstrained
optimization by minimizing
the sum of squared errors
© Dynardo GmbH
31Webinar Parameter Identification with optiSLang
• Unknown parameters defined in ASCII input file
2nd example: Tension Test of Spring Steel
© Dynardo GmbH
32Webinar Parameter Identification with optiSLang
• Displacements and forces
of measurements are
parameterized as signal
2nd example: Definition of the Reference Signal
© Dynardo GmbH
33Webinar Parameter Identification with optiSLang
2nd example: Definition of the Output Signal
• Displacements and forces
of simulation are
parameterized as signal
from a binary format
(file.rst)
© Dynardo GmbH
34Webinar Parameter Identification with optiSLang
• With Instant Visualization (1) it is possible to compare both signals
Both signals do not have the same discretization (2) and length (3)
To get the same length and discretization it is necessary to extract the abscissa from the Signal_Ref and than interpolate the Signal_raw to this abscissa
2nd example: Definition of the Output Signal
2.
3.
1.
© Dynardo GmbH
35Webinar Parameter Identification with optiSLang
2nd example: Definition of Signal Functions
• The displacement is divided in 15 equally spaced steps (1-15) to get
more detailed information about the influence of the 5 material
parameters
• At these steps the forces will be extracted
© Dynardo GmbH
15.…1.
36Webinar Parameter Identification with optiSLang
2nd example: Definition of the Design Variables
1. Adjust lower and upper bounds for all parameters
© Dynardo GmbH
37Webinar Parameter Identification with optiSLang
2nd example: Results of the Sensitivity Analysis
• The reference is covered sufficiently by the simulations
• Parameter bounds seem to be adequate for the calibration
© Dynardo GmbH
38Webinar Parameter Identification with optiSLang
2nd example: Results of the Sensitivity Analysis
• The CoP value of the signal difference indicates a good explainability
of this function
• Linear hardening coefficient R0 are not detected as important
Check also single force_steps_sim values
© Dynardo GmbH
39Webinar Parameter Identification with optiSLang
2nd example: Results of the Sensitivity Analysis
© Dynardo GmbH
force_steps[0]
force_steps[14]…
• Single values can be explained much
better as global difference
• Only Linear hardening coefficient R0 is
unimportant in all force values
• The influence of the Young´s modulus
decreases meanwhile the influence of
the exponential hardening coefficient
R∞ increases with increasing
displacement
40Webinar Parameter Identification with optiSLang
2nd example: Optimization using the MOP
© Dynardo GmbH
• Linear hardening coefficient R0 is not sensitive to any of the force values
It can not be identified and is not considered in the optimization
• Single force value are approximated by MOP and the criteria (sum of
squared errors) is formulated based on their approximation
41Webinar Parameter Identification with optiSLang
2nd example: Results of the Optimization on MOP
© Dynardo GmbH
• The optimizer converges in a few iterations
• The best design is validated
The agreement between reference and simulation is already very good
42Webinar Parameter Identification with optiSLang
© Dynardo GmbH
Summary
• The single force values could by approximated by the MOP
much better as the global difference value
The objective function was formulated directly with the force values
The optimization on the MOP obtained a very good agreement of
simulation and measurement curve
Excellent agreement could by finally achieved with the Simplex optimizer
Initial: difference = 3864N MOP: difference = 268N Simplex: difference = 205N
43Webinar Parameter Identification with optiSLang
optiSLang Training Program
© Dynardo GmbH
44Webinar Parameter Identification with optiSLang
© Dynardo GmbH
Need More information? Training?
Upcoming Events
Webinars (approx. 1hr.)Introduction to optiSLang Sep. 16, 2016optiSLang and ANSYS Workbench Sep. 19, 2016optiSLang and ANSYS Maxwell Sep. 20, 2016Data Analysis with optiSLang Sep. 21, 2016Parameter Identification with optiSLang Sep. 22, 2016optiSLang and Simulation X Sep. 27, 2016Customization and Automation in optiSLang Sep. 27, 2016Introduction to Statistics on Structures Sep. 28, 2016
optiSLang Basic training (3days in Weimar) Oct. 24-26, 2016
Info and registration athttp://dynardo.de/en/training/training/seminaroverview.html
45Webinar Parameter Identification with optiSLang
© Dynardo GmbH
Need more information?
support@dynardo.de
03643/9008-32
Trainings…
Any request or recommendation…
kontakt@dynardo.de
Need a helping hand?
Pilot projects
CAE-Consulting
46Webinar Parameter Identification with optiSLang
Thank you
For more information please
visit our homepage:
www.dynardo.com
© Dynardo GmbH