Optimisation, design studies and analyses

Transcription

1 VTT TECHNICAL RESEARCH CENTRE OF FINLAND LTD Optimisation, design studies and analyses SIMPRO VTT subproject, task 2 Juha Virtanen, Pekka Rahkola, Timo Avikainen, Kai Katajamäki, Erin Komi, Paul Klinge, Seppo Uosukainen

2 Task 2: Main ideas and objectives Utilisation of optimisation, and design and sensitivity analyses in product development process Acceleration of design processes and avoiding unnecessary evaluation loops by using calculation power of HPC environment Design process aided by design studies, design of experiment analyses and optimization algorithms Specific properties (ie. pros and cons) of calculation environment Effective control over on-going processes, overall management and monitoring Cluster task management and queuing systems Objective is to make HPC environment as a daily calculation tool for engineers/researchers working with large scale numerical models Different types of analyses including optimisation, and design and sensitivity analyses Large FEM/MBS/CFD models, coupled problems, etc. with extensive parameter space 13/11/2015 2

3 Task 2: subtasks - tasks 2.4 and 2.6 were not activated during project Task 2.1 Task 2.2 Task 2.3 Task 2.4 Task 2.5 Task 2.6 Task 2.7 Parameter optimisation Parameter variation and optimisation with FEM analyses Parameter variation and optimisation with MBS simulation Topology optimisation Optimisation of complex geometrical shape and structural topology in HPC environment Structure optimisation with discrete element method Large scale DEM models Open source tools in engineering optimisation Optimisation of fluid flow case using open source tools Acoustic simulation in HPC and cloud environment Material parameter variation and optimisation for cabin acoustics Optimisation of welding simulation Welding simulation, optimisation in sense of fatigue life for welding parameters like deflections, residual stresses and weld pass order Final report 13/11/2015 3

4 Task 2.1 Parameter Optimisation with Multibody Simulation Models 4

5 Research Case Study of parameter optimization with multibody simulation model Vibrations features of an elevator model were optimized using different design variables of the system during elevator rise Two case studies done with optimization codes: HEEDS MDO DAKOTA Scalar valued objective functions (weighted sum of components) and inequality constraints were defined for case studies 13/11/2015 5

6 Results and Conclusions Both case studies were successful: optimized design were found Building the optimization processes were straightforward With HEEDS MDO (with GUI) Adams portal was used, usage was fluent and efficient With DAKOTA the connection with Adams needs some scripting, but after that the usage is robust and efficient No special attention spend on optimization methods; default methods and parameters were used in both case studies Objective function components as a function of objective value (case 1) Objective function components as a function of objective value (case 2) More emphasis spent to define the optimized simulation process in a way that makes sense For example, are all relevant excitations present in the model? 13/11/2015 6

7 Task 2.1 Case KONE Guiderail Optimization with FEM

8 Research Case Presented study case: KONE Oyj Guide Rail applied in the elevator frames What was done in the study: Optimization of the Guide Rail properties The optimization was tested by both the simple beam modeling and more heavier solid modeling FEM techniques Abaqus FE-code was used as the main FEM software in preprocessing and as a solver) The linear beam modeling FEM techniques Limited beam picked out from the frame Applied Heeds with Abaqus Applied Excel with Heeds The simplified solid FE model of the guide rail Solid modeling by I-sight with Abaqus Cut from the original model Symmetry conditions in the both ends Cross section dimensions as parameters

9 Parameters used in optimization Objectives of the optimization problem Minimize linemass of guide rail beam Maximize beam support span Minimize support bracket stiffness Variables Beam span length Support stiffness Variable T-profile shape Load positions (guide loads are constant) Constraints Max deflection Min buckling load Max Torsion angle Several dimensions min/max values Section dimension aspect ratios (t/b, thinness, stability)

10 Optimization related activities Phase 1. Phase 2. Testing Beam elements Low DOF models Interactions Abaqus CAE tools Global phenomena More testing Solid elements Abaqus CAE tools Abaqus CAE kinematics Time integration Implicit / Explicit Larger model dimensions Phase 3. CPU time estimation Actual models Phase 4. Planning for actual optimization model or models Investigating and matching the optimization software possibilities from HEEDS and Isight Selection of actual input parameters, geometry, spring stiffness etc. Selection of actual output parameters, acceleration, defflection, stress level at selected locations Global checks for buckling, stress levels etc. Phase 5. Running optimization computations with HEEDS and Isight with one or multiple models Gaining optimization understanding of the case KONE ELEVATOR GUIDERAIL Gaining optimization knowledge and experience of the computational process for SIMPRO FEM Gaingin experience from the cluster computation HPC Final computations Reporting internally to VTT SIMPRO people Reporting externally to KONE and SIMPRO group Conclusions and recommendations for computational processes keeping SIMPRO goals and future customer optimization work in mind Best practices, best software and tools combinations, encountered practical problems, CPU demands, Optimization related demands for the problem in general Post processing issues

11 DOE and optimization process using Simulia/Isight The mass of the guide rail reduced (roughly 15 %) by the preliminary optimization: Three cross section parameters and length of the shorter guide rails defined as the design variables Displacement defined as the constraint Guide rail volume defined as the objective Static analysis (no time involved) with the moving elevator (sweep) Pareto plot describes the importance of the parameters in terms of the mass reduction in the guide rail

12 Guide Rail Analyses with Abaqus and HEEDS Abaqus dynamic implicit model Solution is dynamic in time Model with beam elements and springs Beam element model built at VTT Five spans, six brackets, no fishplates (with first model) Selected set of five T-profiles modeled as simplified I-profile sections Abaqus user defined Fortran subroutine for moving guide shoe forces Simulation length is 6 seconds, 5 ms time step (ok up to ~30 Hz) Single Abaqus solution takes about 5 minutes with good desktop computer HEEDS (MO-SHERPA) multi objective optimization with 300 solutions, 25 hours Buckling analysis steps included (found to be problematic in same analysis model) Excel beam model Springs at end supports Single beam span Deflection calculated Buckling load calculated HEEDS (SHERPA-MO) multi objective optimization with 300 solutions, 5 minutes 300 times faster Good for investigating the optimization problem field

13 Conclusions Solid Element Model Enables more detailed design to be optimized e.g. extra holes in the guide rail web Heavier alternative in terms of both computing and update modification In the case of the complex multiparametric model some parameters can be in contradiction to each other The optimization was possible in a quite narrow parameter range Dynamic Beam Element Model It is important to keep over all DOF size low (i.e. model resolution, time resolution, optimization cycles) It is important to keep the dynamic model well defined and robust Moving elevator guide forces were applied with a Fortran user subroutine Optimization It is important to choose the right analysis tool to begin with MBS or FEM or analytic problem? Static or Dynamic? Objective was to find parameter trade-off dependencies (Pareto frontiers) as optimization results Instead, defined input constraint values were limiting Red Cedar Technologies HEEDS Optimization software was taken successfully into use as a new tool

14 Task 2.1 Case Optimisation of Dieselgenerating Set Vibration Levels with Dakota

15 Optimisation of diesel-generating set vibration levels with Dakota Need for automatization and optimisation Analysis time reduction and cost savings Reliability increases and avoidance of human errors Possibility to link different analysis types in the same calculation process Temperature, stress vibration, acoustics, Enables repeatability and modularity in analysis process Automatic model and results documentation and reporting Increases understanding related to effect on input variables, sensitivity studies and statistical analyses

16 Calculation engine in optimisation Dakota as optimisation software Calculation engine programmed with Python Based on templates for model and analysis Automatic natural mode identification Automatic results post-processing

17 Calculation engine flowchart programmed with Python

18 Optimisation results Base frame optimisation problem Minimizing mass Vibration response levels as constraints 6 continuous and 8 discrete parameters Evolutionary algorithm 4000 case evaluations in HPC environment Result: Generating set mass reduced 10 %, vibratory levels reduced 80 % Parameters Resonance removed Mass reduction ID SID COV BLH CHU BOT BOTH CH1 CH2 CH3 CH4 CH5 CH6 CH7 CH8 Mass Vel Evo OFF OFF OFF ON ON ON ON ON -9 % -78 % Dir OFF OFF OFF ON ON ON ON ON -0.1 % -78 % Par OFF OFF OFF ON ON ON ON ON 1 % -82 %

19 Task 2.2 Topology Optimization in HPC Environment

20 Research Case Study case Generating set, base frame as design space Project goal Find a more efficient way of utilizing existing software tools for topology optimization in product development Produce a large number of potential design solutions automatically Project idea HyperStudy is used to create a design of experiments that launches and controls the entire simulation process The design space is parameterized with HyperMorph and used as design variables within the DOE OptiStruct is launched and each variation of the design space is topology optimized OSSmooth is used to interpret each design (several variations of the interpretation criteria can be tested) All new designs are reanalyzed with OptiStruct Important results are labelled as responses and are automatically tabulated DOE (HyperStudy)

21 Results and Conclusions After preparing the model and DOE, it only takes one mouse-click to generate literally hundreds of possible designs which can be evaluated by the designer based on the tabulated DOE responses 48 out of 64 topology optimization runs resulted in a converged solution It was found that even small changes in the size of the design space could lead to unique optimized solutions

22 Task 2.3 Optimization of a particle damper with DEM

23 Optimisation of a particle damper with DEM A particle damper is an attractive alternative for reducing mechanical vibration, since it can attenuate a wider frequency range. Basically it is composed of a cavity that contain particles like ball bearings. Vibration energy is removed by impacts between the particles. The idea of the study was to test, if it is feasible to design a particle damper utilising the computational power of current parallel super computers, the discrete element method (DEM), and up to date optimisation tools.

24 A simple example case A spring and a mass with a rigid cavity partly filled with beads and excited by a harmonic force Three optimisation parameters: The length the cylindrical part The radius of the hemispheres at the ends The number of particles in the cavity Steel balls turned out to require a very short time step. A more suitable run time was achieved by using only 1/5 stiffness of the steel for the balls The upper limit for the number of balls was set to 3000 The effect of the length of the cylinder part to the shape of the cavity

25 Analysis and optimisation software The open source code DAKOTA was used for optimisation It contains a large and diverse library of optimisation methods The peek velocity of the damped mass after 10 s was the target result ABAQUS that included a new DEM solver was used for mechanical simulation A Python program was written that read the velocity time series from the ABAQUS database, calculated the average response amplitude of the last 2 seconds, and wrote the result into a file for DAKOTA

26 Results and Conclusions Optimisation methods based on numerical derivatives either stopped or found the optimum near the starting point Global algorithms worked better DIRECT (DIvision of RECTangles) always stopped before the optimum because the ABAQUS simulation ended with error PDS (Parallel Direct Search) worked otherwise fine, but it continued to refine the result long after finding the optimum It was shown that the basic concept works The PDS optimisation method was able to find the optimal configuration for the particle damper Optimisation can be used to design a particle damper, but to be practical A much faster DEM solver is needed Parallel versions of the optimisation method should be available A straightforward parameter study may be a better option

27 Task 2.5 Acoustic simulation in HPC and cloud environment

28 Research case Goals To perform an acoustic simulation test case with Waveller Cloud and to compare the results to those of Actran Waveller Cloud (by Kuava Ltd) Available for local and cloud computing Uses CompA algorithm (by Aalto University) o o Based on the boundary element method (BEM) for solving the acoustic wave equation in 3D Multi-level fast multipole algorithm (MLFMA) BEM is mostly used and best applicable to exterior problems In this context it was used to interior problems o o Not as efficient as the finite element method (FEM) True fluid-structure interaction and propagating waves in structures cannot be modelled with BEM V V e r n e r n A A

29 z(m) z(m) z(m) z(m) y(m) Test case study: Valtra cabin T888M Noise emission from an acoustic point source into the interior of a tractor cabin Inner roof centre part Inner roof outer part 100 Hz, three orthogonal cross sections Torso (driver s head and chest) Actran (FEM) Point source #1, f = 100 Hz, Lp in x = m plane y(m) Point source #1, f = 100 Hz, Lp in y = 0 m plane x(m) Point source #1, f = 100 Hz, Lp in z = m plane x(m) Waveller Cloud Point source #1, f = 100 Hz, Lp in x = m plane y(m) Point source #1, f = 100 Hz, Lp in y = 0 m plane x(m) Inner roof case 1 Inner roof case 2

30 Conclusions Largely different results at 100 Hz for the inner roof case 1 The structurally and acoustically coupled centre and outer parts of the inner roof treated separately and as locally reacting in BEM The centre and outer parts have quite different properties and frequency dependencies The calculation results for the other inner roof cases are tolerably in agreement Small deviations probably due to treating boundary surfaces in different ways in FEM and BEM Waveller Cloud seems to give a little higher levels because some dissipation effects and the sound transmission outwards of the cabin cannot be taken into account by BEM

31 TECHNOLOGY FOR BUSINESS