Webinar Machine Tool Optimization with ANSYS optislang 1
Outline Introduction Process Integration Design of Experiments & Sensitivity Analysis Multi-objective Optimization Single-objective Optimization Summary Thomas Most Dynardo GmbH 2
Introduction to optislang 3
Real world CAE-Applications CAE-based virtual prototyping needs significant computational resources Physical phenomena may be coupled, non-linear and high dimensional Need to deal with failed designs (design creation, meshing or simulation may fail, license failure) Need to deal with many parameters (at least in the uncertainty domain) 4
optislang is an general purpose tool for variation analysis using CAE-based design sets (and/or data sets) for the purpose of sensitivity analysis design/data exploration calibration of virtual models to tests optimization of product performance quantification of product robustness and product reliability Robust Design Optimization (RDO) and Design for Six Sigma (DFSS) serves arbitrary CAX tools with support of process integration, process automation and workflow generation 5
Excellence of optislang optislang is an integration toolbox for Process automation, Design variation, Sensitivity analysis, Optimization, Robustness evaluation, Reliability analysis and Robust design optimization (RDO) Functionality of stochastic analysis to run real world CAE-based industrial applications Easy and safe to use: Predefined workflows, Algorithmic wizards and Robust default settings 6
Example: Machine Tool Device Robotic system for drilling large structural components Rating of robot kinematics and machining quality Test device is used as a dummy for large structural components High stiffness and light weight (because of manual adjustment) are required Fraunhofer IPA Stuttgart Germany 7
Example: Optimization Task Objective functions Minimization of mass Minimization of deformation of the beam structure for a positioning in 0, 90 and 180 Initial Design Mass: 207,2 kg Deformations: 0 -position: 0,12 mm 90 - position : 0,10 mm 180 - position : 0,07 mm 8
Example: Simulation Model Three load cases in ANSYS Workbench Deformations of 3 load cases as outputs in parameter set 9
Example: Optimization Parameters Thickness of upper plate (initial: 10 mm) Width, height and thickness of upper beams (50 x 50 x 3 mm³) Width, height and thickness of middle beams (40 x 40 x 3 mm³) Width, height and thickness of lower beams (70 x 70 x 4 mm³) Steel beam structure and upper plate is Aluminium 10
Process Integration 11
Process Integration Parametric model as base for User-defined optimization (design) space Naturally given robustness (random) space Design variables Entities that define the design space Scattering variables Entities that define the robustness space The CAE process Generates the results according to the inputs Response variables Outputs from the system 12
optislang Integrations & Interfaces Direct integrations ANSYS Workbench MATLAB Excel Python SimulationX Supported connections ANSYS APDL Abaqus Adams AMESim LS-Dyna Arbitary connection of text-based solvers 13
ANSYS Workbench optislang Plugin optislang modules connect directly to ANSYS Workbench parameter set 14
CAX-Interfaces the ANSYS Workbench Node optislang Integrations provides the flexibility to extend the process chain ANSYS Workbench can be coupled with different other solvers like MATLAB, SimulationX or Abaqus External geometry or mesh generators can work together with the ANSYS Workbench node 15
Sensitivity Analysis 16
Sensitivity Analysis Understand the most important input variables! 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 17
Sensitivity Analysis Flowchart Design of Experiments Deterministic (Quasi-)Random Solver Regression Methods 1D regression nd polynomials Sophisticated metamodels Sensitivity Evaluation Correlations Variance-based quantification 1. Design of Experiments generates a specific number of designs which are all evaluated by the solver 2. Regression methods approximate the solver responses to understand and to assess its behavior 3. The variable influence is quantified using the approximation functions 18
How to Generate a Design of Experiments Deterministic DoE Complex scheme required to detect multivariate dependencies Exponential growth with dimension Full factorial: Koshal linear: Advanced Latin Hypercube Sampling Reduced sample size for statistical estimates compared to plain Monte Carlo Reduces unwanted input correlation 19
Response Surface Method Approximation of response variables as explicit function of all input variables Approximation function can be used for sensitivity analysis and/or optimization Global methods (Polynomial regression, Neural Networks, ) Local methods (Spline interpolation, Moving Least Squares, Radial Basis Functions, Kriging, ) Approximation quality decreases with increasing input dimension Successful application requires objective measures of the prognosis quality 20
Metamodel of Optimal Prognosis (MOP) Objective measure of prognosis quality Determination of relevant parameter subspace Determination of optimal approximation model Approximation of solver output by fast surrogate model without over-fitting Evaluation of variable sensitivities 21
The Sensitivity Wizard Drop the sensitivity wizard on the final solver chain Define the lower and upper bounds of the input variables The sampling method and sample number is recommended depending on the chosen solver runtime and number of parameters 22
Example: Definition of Parameter Bounds 23
Example: Design of Experiments 200 Latin Hypercube Samples 6 failed designs due to conflicting geometry parameters (red, violet) 16 designs with implausible model behavior (orange) 24
Example: MOP Results All responses can be explained very well Mass is almost linear, deformations are nonlinear but monotonic w.r.t. the input parameters 25
Example: Influence of Parameters Thickness of upper plate is most important for the mass Parameters of lower beam sections have highest influence on deformations 26
Multi-Objective Optimization 27
Optimization Optimize your product design! Optimization using MOP Start Work with the reduced subset of only important parameters Pre-optimization on meta-model (one additional solver run) Optimization with cutting edge algorithms Decision tree for optimization algorithms 28
optislang Optimization Algorithms: concepts Gradient-Based Methods Go downhill In the context of blackbox solver gradient needs to be measured via DoE schemes Adaptive Response Surface Method DoE and surface fit Scanned area shifts and shrinks Nature-Inspired Optimization Inspiration sources: evolution, swarm motion, thermodynamics, state insects Mutation, recombination, selection, propagation, go beyond nature by tthinking in operators Start 29
optislang Optimization Algorithms: applications Gradient-based Methods Clear favorite if objective function smooth and without local minima Repeated local search as strategy when there are few local minima Adaptive Response Surface Method ARSM is the default choice Balance between robustness and efficiency Best if dimension<20 Nature inspired Optimization Recommended for global search Widest applicability (binary or discrete parameters, ) Realize optimization potential also in challenging situations Start 30
The Multi-Objective Optimization Task Several optimization criteria are formulated in terms of the input variables x Uncountable set of solutions, if criteria are contradicting Balanced compromise is wanted Scan of the Pareto-front as decision base Constraint restrictions are possible 31
Decision Tree for Optimizer Selection Recommendation of most suitable optimizer depending on Number and type of input parameters (automatic) Number of constraints and objectives (automatic) Analysis status (user setting) Amount of failed designs (user setting) Solver noise (user setting) 32
Example: Definition of Optimization Criteria Goal 1: Minimization of the mass (initial 207 kg) Goal 2: Minimization of maximum deformation in 0, 90 und 180 position (initial 0.12 mm) 33
Example: Multi-Objective Optimization using MOP 34
Example: Multi-Objective Optimization using MOP As compromise solution a maximum deformation of 0.1 mm is chosen 35
Single-Objective Optimization 36
Example: Optimization using the MOP Optimization goal: minimization of the mass Constraints: Deformations (0, 90 und 180 ) smaller as 0,1mm 37
Example: Optimization using the MOP Validated best design violates the constraints slightly Validated mass is larger as approximated 38
Example: Optimization using the MOP Initial Design Mass: 207,2 kg Deformations: 0 -position: 0,120 mm 90 - position : 0,097 mm 180 - position : 0,067 mm Optimized Design Mass: 184,7 kg (-11%) Deformations: 0 - position : 0,106 mm (-12%) 90 - position : 0,087 mm 180 - position : 0,060 mm 39
Example: Optimization using the FEM model ARSM optimizer using the previous optimum as start design Deformation constraints are fulfilled Mass is increased 40
Example: Optimization using the FEM model Initial Design Mass: 207,2 kg Deformations: 0 -position: 0,120 mm 90 - position : 0,097 mm 180 - position : 0,067 mm Optimized Design Mass: 188,3 kg (-9%) Deformations: 0 - position : 0,100 mm (-17%) 90 - position : 0,084 mm 180 - position : 0,055 mm 41
Summary Sensitivity analysis helps to better understand the physical phenomena and to check or validate the simulation model Indification of important parameters helps to significantly simplify and accelerate the optimization task MOP-approximation can be used for fast pre-optimization step or multi-objective case studies 42
Need more information? Support & trial license: support@dynardo.com Training: training@dynardo.de www.dynardo.de 43