Computer-aided optimization software has existed for many years, and many vendors delight in showing rather obvious examples of how it works - thinning the spokes of a wheel to reduce weight while maintaining static strength being quite common. While such programs do have their place, it is important to realize the complexity of the problem faced when optimizing NVH performance - and even more so for multi-disciplinary attribute optimization, such as NVH and durability. The main challenges: the sheer size of the model(s) which leads to extremely long compute times; the computer resources; and the scheduling of the tasks involved in exploring the design space (“the virtual test”) that allows the automated searching for an optimum point.
However, the base technologies do exist. LMS OPTIMUS has proven itself as a tool for general design space exploration - it can automatically generate, analyze, explore, and track design alternatives. As part of the international research project ESPRIT/HPC-VAO (High Performance Computing - Vibro-Acoustic Optimization) engineers from LMS and Rover, and computer scientists from the Parallel Computing Centre (PAC) at the University of Southampton worked to further develop the technologies that would be needed to make multidisciplinary optimization not only possible, but practical as well.
The central aim of the HPC-VAO was the implementation and demonstration of a state-of-the art CAE Environment to support design optimization in the field of vehicle NVH engineering. Specific objectives covered both the essential aspects of optimization technology, and the implementation of this methodology in a coherent HPCN (high performance computer networks) framework. The involvement of Rover was to assess the suitability of such an environment to solve industrial problems - this involvement had mainly been concentrated on the technical aspects of developing tools (such as LMS Gateway, Link, and SYSNOISE) which had the numerical accuracy and analytic functionality required by engineers. The final stage of the project was to integrate these analytic tools with the simulation management software LMS OPTIMUS linked to an intelligent resource manager, “INTREPID”, developed during the project by PAC.
Two test cases were deployed for the evaluation tasks; a simple box model in the form of a rectangular enclosure coupled to a surrounding structure, and a vehicle system model prepared for the project.
The simplicity of the box model was desired to obtain a quick response for the sequential solve and parallel solve investigation for a Design of Experiments (DOE). In this DOE application, the experimentation phase consisted of a number of designs which corresponded to a variation of material thickness of the sides of the box ranging from 0.1mm to 1.1mm. The simulation phase comprised a sequence of vibro-acoustic analysis steps; NASTRAN was used to calculate the vibration response to a mechanical force input, the output was then passed over to SYSNOISE for a coupled acoustic analysis using the FEM solver option. The sound pressure level over a specified range, defined as the optimization target, was calculated in SYSNOISE. The target value being known for all designs, an approximate response surface was calculated in the design space and investigated in OPTIMUS. The minima of the response surface, the SPL response at an interior point was then obtained.
A more realistic case was provided by a complete vehicle system model of a Land Rover 4x4 vehicle with the ability to calculate the interior noise up to 140Hz in response to real-life engine excitations. A set of design variables relating to dynamic stiffness of the engine and gearbox mounts was identified as critical vehicle parameters, and the optimization challenge was to minimize the SPL over a broad frequency range at the driver’s ear.
The biggest component is the body structure with over 600,000 DOF. However, it was not possible nor even desirable to carry out experimentation/optimization with a full scale vehicle model - earlier work in the project by BMW had demonstrated the benefits of applying modal formulation for vibro-acoustic optimization, especially to large problems such as this. A similar approach was used by Rover, where components bigger than 10,000 DOF in a vehicle can be reduced to a NASTRAN “superelement”, or characterized by an experimental modal model. Other components, such as subframe, exhaust, and engine can also be reduced in the same way, provided that they are not design variables to be used in the DOE. In fact, only components which should be retained for the analysis as “physical” are those required for force excitation, response definition and the definition of optimization parameters.
The figure shows the optimization sequence for one complete iteration for the box. The four variables represent the material thickness of the box sides and were initially set to 1mm. The NASTRAN model consisted of 80 elements and 100 nodes. Each of the box sides was associated with a distinct material property, which was in turn associated with each of the design variables. OPTIMUS controlled the DOE, processing the events required to perform the optimization sequence and the target number of iterations. It also controls the files needed to interrogate the desired response and where it is recorded. The OPTIMUS sequence file examines the NASTRAN template file, copies and creates a new NASTRAN file, replacing the design variable names with physical material thickness values within the bounds of the DOE. Once the NASTRAN file has been created it is submitted for analysis, which returns the response weight of the box model and the forced displacements of the structure. These are used by SYSNOISE to compute the first 10 acoustic modes for the cavity, leading to an integrated pressure response between 10Hz and 1000Hz at the receiver location. This completes one iteration run.
A distinction is then made for sequential solve (where the subsequent iterations required for the optimization are simply repeated one after another) and where they are controlled by the HPC-VAO PAC development INTREPID in a parallel way. The rationale for the latter is that several iterations are usually required for a DOE and if they can be solved simultaneously using a parallel computing approach significant time saving can be achieved.
For the box model the required time to sequentially complete the DOE was 24 minutes; using a parallel approach with 2 NASTRAN and 2 SYSNOISE licenses was 12 minutes. When the model complexity was increased to 550 elements and 552 nodes the times were 118 and 40 minutes respectively - already showing that the benefits of parallelization become more apparent the larger the model. For the full vehicle, using superelements for the body and chassis the sequential method required 27:40 hours. Using 9 NASTRAN and 4 SYSNOISE licenses, and running overnight, the DOE required a little over 5 hours - a time saving of 80%.1 The integrated OPTIMUS/INTREPID system is still a prototype, however, its benefit in the vehicle development cycle has been proven.
The HPC-VAO project was a three year research project funded partially by the ESPRIT-HPCN program of the European Union.1 The consortium comprised leading CAE suppliers; LMS International for Acoustic, Noise, Vibration, Durability and Optimization and Design Space Exploration Techniques; MSC for Structural FEM; major car manufacturers Rover and Renault; and internationally recognized research centers IMEC and PAC.
This article was condensed from the ESPRIT HPC-VAO Report/Deliverable 7.2. LMS thanks Faruk Turgay, Rover Group and Tim Cooper, PAC Southampton University for their assistance.
