Professional Documents
Culture Documents
Best Practice Guide CFD 2 PDF
Best Practice Guide CFD 2 PDF
Issue 1
ALESSIA:
EP 28189 Application of Large Eddy Simulation to the
Solution of Industrial problems:
BEST PRACTICE GUIDE: LES AND ACOUSTICS
Authors
Christiane Montavon*, Ian Jones*, Dick ten Bosch+, Stefan Szepessy , Roland Henriksson,
Hans Moberg, Roberto Tregnago , Zoubida El Hachemi #, Massimiliano Piccirillo#, Michel
Tournour#, Sylvie Dequand#, Frederic Tremblay !, Rainer Friedrich!
SUMMARY
Time dependent flow phenomena play an important role in many industrial processes,
including aerodynamic noise generation. In many cases experimental and theoretical analyses
of the problems are not feasible, either due to their high expense or to extreme operating
conditions. The primary aim of the Alessia Project has been to develop validated and
supported software tools for the prediction of industrially important fluctuating flow
problems. The tools are based around the technique of Large Eddy Simulation (LES), which
solves for the large-scale fluctuating flows, the Large Eddies, and uses sub-grid scale
turbulence models for the small-scale motion. The leading European software packages, CFX
for Computational Fluid Dynamics (CFD), and SYSNOISE for the acoustics, have been
coupled together in this project to perform the prediction of industrially relevant flows, and
the consequent noise field. The main limitation to the solution of these problem has been the
computational cost of computer predictions. HPCN has been an enabling technique that has
provided the solutions for these multi-physics and multi-scale problems.
This report is a subset of one of the deliverables of the project and summarises some of the
experiences learnt from the project, in the form of a Best Practice Guide. It also contains some
considerations for the choice of computing system for LES computations.
CONTENTS
INTRODUCTION
4.1
Boundary Layers
4.2
Inlets
4.3
Outlets
4.4
Geometrical Domain
4.5
Meshes
4.6
Numerical Considerations
4.7
Advection scheme
4.8
4.9
Accelerating Convergence
4.10
4.11
AEROACOUSTIC CALCULATIONS
5.1
SYSNOISE modules
5.2
5.3
5.4
Acoustic I-FEM
5.5
5.6
5.7
Numerical Consideration
5.7.1
Mesh coarsening
9
9
5.7.2
5.7.3
6
Free edges
Junctions
9
9
10
6.1
Operating environment
11
6.2
11
REFERENCES
12
1 Introduction
Time dependent flow phenomena play an important role in many industrial processes,
including aerodynamic noise generation. In many cases experimental and theoretical analyses
of the problems are not feasible, either due to their high expense or to extreme operating
conditions. The primary aim of the Alessia Project [ 1] has been to develop validated and
supported software tools for the prediction of industrially important fluctuating flow
problems. The tools are based around the technique of Large Eddy Simulation (LES), which
solves for the large-scale fluctuating flows, the Large Eddies, and uses sub-grid scale
turbulence models for the small-scale motion. The leading European software packages, CFX
for Computational Fluid Dynamics (CFD), and SYSNOISE for the acoustics, have been
coupled together in this project to perform the prediction of industrially relevant flows, and
the consequent noise field. The main limitation to the solution of these problems has been the
computational cost of computer predictions. High Performance Computing and Networking
(HPCN) has been an enabling technique that has provided the solutions for these multiphysics and multi-scale problems.
A summary of Validation results for this project are to be found in [ 2]. The partners in the
project have combined their experiences on these problems to produce a Best Practice Guide
and Recommendations. This focuses on:
When to consider the use of LES, instead of conventional Reynolds Averaged modelling.
How to use CFD software for large-scale transient simulations in an HPC environment.
This will include the recommendations on solver parameters, for efficiency.
Recommendations for size of cluster, and cluster configurations, speed of interconnection,
memory requirements for different size problems and applications.
Description of the parallel algorithms, and the specific issues on the running of large
dynamic transient cases
geometries. This has wasted significant memory because of the blocking off. This project has
attempted to use the latest generation of commercial CFD software, with powerful methods
for dealing with complex geometries in conjunction with LES. Because of the difficulty of
validating the methods for solving problems which are inherently chaotic, the strategy has
been to use very accurate solutions from the MGLET software, from the Technical University
of Munich, to provide benchmark quality data for testing and evaluating the industrial
software. This has inevitably required the use of the largest supercomputers available in
Europe, for the benchmarking, along with lower cost industrial strength systems, based on
commodity workstations, for the industrial computations.
4.2 INLETS
The representation of the turbulent structures at inlets can be a difficult part of the setup. As
could be seen from the NTH case, the detailed properties of the incoming flow had a strong
effect on the development of the jet in the enclosure.
For the case of the cyclone however, the turbulent conditions at the inlet seemed to have
relatively little impact on the flow in the device, which was essentially determined by very
strong anisotropy effects in the cyclone body.
Cases for which the inlet turbulence plays a significant role are developing boundary layers
and turbulent jets.
If the inlet turbulence is felt to play an important role, the user should consider
a) Going upstream, with an appreciable region of the inlet resolved to use LES for the
flow upstream.
b) Use of pre-computed LES flows, storing only the values on a plane
c) Use of SNGR model and RANS models to generate inlet flows with the right turbulent
structures.
d) New methods such as Detached Eddy Simulation (DES) show promise, but they are
not fully proven yet. In particular, there could be upstream effects from the LES
portion, or the turbulent fluctuations in the upstream boundary layer could be
necessary to trigger off the unstable shedding of vortex sheets. This is a new area, and
more experience is required. Beware the claims of the originators of new methods,
since they may not report the bad experiences, and the difficulties encountered.
4.3 OUTLETS
With the transport of turbulent eddies outside of the computational domain, some
recirculation can occur at outlets. Our experience shows that if the code is allowed to bring
some fluid back into the domain at outlets, it can destabilise the solution. We therefore
recommend in CFX-5 to use OUTLET boundary conditions rather than OPENING when
performing a LES. OPENING type of boundary conditions allow the flow to come back in,
whereas with OUTLET boundary conditions, the code builds artificial barriers at the
boundary, when the flow tries to come back in. These barriers are later taken away when the
flow goes out again.
The use of artificial walls at outlets might introduce some not very physical behaviour locally,
but it increases the robustness of the calculation.
4.5 MESHES
Remember that the mesh and time steps are an inherent part of the model. The LES models
make use of the grid scale for filtering out the turbulence. If the mesh is anisotropic to resolve
a jet, for example, the longer length scale in the flow direction may also have an undue effect
on the cross-stream turbulence. For this reason, consider the use of isotropic grids, maybe
using tetrahedral elements, rather than hexahedral meshes.
Where possible, methods should be at least second order accurate in space. Second
order CENTRAL DIFFERENCE methods have certainly given good answers within
the Alessia project, both in CFX-5 and MGLET.
Second order Higher Upwind Differencing Methods and the QUICK scheme, based
on backward difference methods appear to be a little too diffusing, and can damp out
turbulence. Similarly, almost second order accurate methods, with a flux limiter to
prevent overshoots also seem to be too diffusive. Some Japanese experience, however,
indicates that QUICK, in conjunction with no turbulence model, appears to give good
results in some cases. This is not recommended by the results from the Alessia project.
Tuned LES solvers conventionally use explicit methods, with central differencing in
space, with second or fourth order methods both in space and time, with regular
rectangular grids. Very simple, Smagorinsky based models are typically used. These
are best for highly accurate solutions, which challenge computing limits. With these
methods there is a reluctance to use methods that transport turbulence quantities, for
issues such as memory usage. This limits the method to length scales such that the
transport of sub-grid scale turbulent kinetic energy is not important.
For tuned LES explicit time-stepping programs, the dominant cost is that of solving a
Poisson equation at each time step, for the pressure field. This can still be very time
consuming.
CFX-5 has the advantages for complex geometries, where the use of a body-fitted or
unstructured grid is necessary, either to resolve the geometry, or to resolve the flow (at
steps, near separation points, in boundary layers). It will also be able to make use of
new models being developed, eg for combustion or multi-phase flows, and provide
commercial support and infrastructure.
Spectral and pseudo-spectral methods can give the most accurate results, but they are
very limited in scope, and when there is insufficient grid resolution, the results can
have large errors.
Use second order time stepping. 2nd order backward differencing seems to be
relatively robust, and gives the right behaviour, despite the fact it is not an A-stable
method. That is, it can have different stability properties to that of the underlying
differential equation.
Second order centred time stepping, like Crank-Nicolson, leapfrog, are more difficult
to run, and are prone to having different solutions on alternate time steps.
First order fully implicit methods in time usually are too diffusive, and the turbulence
is damped out. For highly unstable problems, such as the cyclone, lower order
methods may work, but the results will be very damped, unless very small time steps
are used.
For Explicit methods, keep CFL condition below 0.25. For a non-uniform grid, this
can restrict the time step to be constrained by the smallest grid size in the problem,
even if it is not in an important region of the flow. Refining the mesh also means that
much smaller time steps have to be taken, to maintain the CFL condition.
Implicit methods can have a higher CFL condition, especially in regions that are not
particularly important, and the grid size is small for geometric considerations.
However, for accuracy, CFL numbers of Order 0.5 1 should be used. Larger values
can give stable results, but the results may be damped.
For smaller time steps, with an implicit coupled solver as in CFX-5, one iteration
(Coefficient loop) can be carried out at each time step, to give a semi-implicit method.
This method may not be as stable as using two or three iterations. When the flow
properties have settled down, then the method may be better1.
A large proportion of computing time is spent getting to the stage when it is possible
to start taking statistics. If this time can be shortened, faster results can be obtained.
This is discussed in more detail later.
1,000 10,000 time steps are typically required for getting converged statistics. More
steps are required for second order quantities, (variances) than for means. Check the
convergence of the statistics. For a vortex-shedding problem, several cycles of the
vortex shedding are required. This effect has also been seen in experimental results,
where the evidence of the results of TUM indicate that measurements should have
been taken for a longer time scale.
When more complex cases are considered with e.g. mixing of multicomponent flows, it is advised to perform
more coefficient loops per time step (typically 6-8). For these cases, unless the mass fraction equations are solved in
a fully coupled way together with the momentum and continuity equation, the mass conservation at the
component level will not be guaranteed with only one coefficient loop per time step.
The Smagorinsky model contains one free parameter, c0 that has a default value of 0.1.
The standard value has been giving good results in most applications, although it
appears that for boundary layer flows and for jets, a lower value can improve
agreement with measurements (e.g. CRF test case, NTH test case). It has however
been observed that it is easy to blame the model constant, when some other
shortcomings like a bad representation of the upstream conditions can affect the flow
just as much as a change in the constant (e.g. NTH case).
To use dynamic models such as Germano, which compute the model coefficient
incorporating explicitly filtered resolved fields, the resolution needs to be sufficiently
good to give better results than Smagorinsky. Any zero order turbulence model will
have the same considerations.
It is sometimes argued that it is important for the models to allow for backscatter
(energy cascading from the smaller to the larger scales close to walls) to improve the
simulation of boundary layers. Models such as INC & DINC developed by TUM have
shown encouraging results during a priori testing. These models however when used
in a simulation still require arbitrary clipping of the viscosity in order to be stabilised.
These models still lack the required robustness to be implemented in a commercial
package.
The system should definitely have a read-write CD on one of the systems, and possibly a
large dat tape system. DVD writers are just becoming available commercially, (600 Euros)
and may also be a viable alternative.
Once it is becomes necessary to create an animation, or results for acoustic calculations, it is
necessary to dump these results out at more frequent intervals, possible every time step. To
minimise file space, the user ought to exercise care as to what and when data is required.
Ideally the system should have a very large scratch disk available for the storage of these
results.
Note that because the acoustic calculations are typically carried out in the frequency domain,
the whole time information is required, in order to perform the FFT for the frequency domain.
If the grid for the acoustic calculations is known a-priori, it is possible to create a CFX
definition file for this grid, and interpolate fine grids onto the coarser mesh, for acoustics postprocessing.
5 AEROACOUSTIC CALCULATIONS
The Alessia project has used an approach of decoupling the flow and acoustics calculations,
using the Lighthill analogy. This is good for situations where the acoustics is required in the
far field, and the flow domain can be much smaller. For near field calculations, the fluctuating
pressure field the pseudo-pressure can give a good indication of the local noise field, but not
if reflections and diffraction effects are important.
Some authors advocate the use of highly accurate methods, to couple the flow and acoustics
in a single framework, using very high order methods. This may be appropriate for some
simple flow cases, but does not make full advantage of the facilities and models available in
commercial software packages.
If the flow is strongly compressible, it may be necessary to couple the flow and acoustics
more closely, for example by carrying out a full compressible flow calculation, within the
CFD solver. This is particularly good when there are acoustic resonances, and the acoustics
affects the flow.
In the approach chosen in the project (relevant for incompressible, low Mach number flows),
the aeroacoustic calculations are performed using SYSNOISE. The next two sections give
guidelines on which module to use in SYSNOISE together with some numerical
consideration.
absorbent elements to model thick absorbent regions, such as foam or wool fill). Transfer
across gaps between element free faces can be modelled using special transfer impedance
relationships, for example to model a perforated membrane. Boundary conditions can include
surface vibrations (assumed, imported from a structural model, or from tests) surface
impedance (absorption layer) and acoustic sources. Solutions can be direct acoustic response
to the excitation, or modal analysis (resonance of the enclosed region) followed by acoustic
response by modal superposition.
Boundary conditions include surface vibration (as in acoustic FEM) surface impedance
(absorption layer) and acoustic sources. In the Indirect BE method, surface-related boundary
conditions can be applied on one side of the element only, if required. Transfer through
selected elements can be modelled using special transfer impedance relationships, for
example to model a perforated membrane. Fluid properties have to be homogeneous
throughout the domain. Acoustic results at any location can be found by using appropriate
field point grids.
A node located along a junction needs to be duplicated (eventually several times) to correctly
represent the different pressure discontinuity through the surface. A compatibility equation
needs to be defined to account for the fact that the total pressure discontinuity is equal to zero.
It is recommended to use the default automatic junction algorithm that has been tested on a
wide range of problem and proved to accurately detect junctions.
7 References
[ 1] Jones, I P
and Montavon, C A, The Alessia Project, available from
http://www.software.aeat.com/cfx/european_projects/alessia/alessia.htm
[ 2] Jones, I P and Montavon, C A, The Alessia Project, Validation Report, available from
http://www.software.aeat.com/cfx/european_projects/alessia/alessia.htm