UFR 2-12 Best Practice Advice: Difference between revisions

From KBwiki
Jump to navigation Jump to search
 
(22 intermediate revisions by the same user not shown)
Line 22: Line 22:
*Confirming a mixed tonal and broadband  nature  of  the  pressure signals on  the  surface  of  the  cylinders  by  their  spectral analysis ([[UFR_2-12_Evaluation#figure16|Figure 16]]).
*Confirming a mixed tonal and broadband  nature  of  the  pressure signals on  the  surface  of  the  cylinders  by  their  spectral analysis ([[UFR_2-12_Evaluation#figure16|Figure 16]]).


== Numerical Modelling ==
== Numerical Issues ==
{{Demo_UFR_BPA2}}
In terms of ''numerics'', based on experience accumulated in the course  of the ATAAC and related projects, the following advice can be given:
== Physical Modelling ==
 
{{Demo_UFR_BPA3}}
*In the Focus Region of simulation (see [[UFR_2-12_Test_Case#figure3|Figure 3]]), use numerical schemes with ''as low numerical dissipation as possible'',  particularly,  pure  or close to pure CDS for convective fluxes with the order of accuracy  not less than 2. Acceptability of the level of numerical dissipation may be assessed by examining snapshots of e.g. vorticity in the focus  region: the size of the smallest  resolved  eddies  should  not  be  noticeably larger than the local grid spacing.
*In the Euler and Departure Regions (see [[UFR_2-12_Test_Case#figure3|Figure 3]])  use  a  scheme  with sufficient  numerical  dissipation  to  prevent  grid  oscillations  or "wiggles" in these regions.
*Use a minimum second order accurate temporal integration scheme.
*Use a time step sufficiently fine to capture the  motion  of  the turbulent eddies resolved by the grid in the Focus  Region.  This corresponds to the approximate guideline CFL<sub>''max''</sub>&nbsp;&asymp;&nbsp;1.
 
In terms of ''grids'', although no systematic grid-sensitivity studies  for the considered UFR have been carried out, indirect evidence
allows  the following recommendations:
 
*In the Focus Region, use nearly isotropic grids with sizes  not  larger than around 0.02''D'', although even smaller values are desirable.
*Outside the Focus Region, expand the  grid  cell  size  gradually towards the inflow/outflow boundaries, avoiding sudden jumps.
 
''Size of computational domain'' should not be less than about &plusmn;20''D'' in  the streamwise and about &plusmn;10''D'' in the lateral direction.
For  the  spanwise direction, the domain size should not be  less  than&nbsp;3''D'',  but  larger domains are  strongly  recommended  provided  that  available  computer resources allow this.
 
== Physical Modelling Issues ==
=== Turbulence modelling ===
*Use  hybrid  turbulence-resolving  approaches,  e.g.  DDES  or  similar methods capable of treating  the  entire  turbulent  cylinder  boundary layer with  RANS.  Steady  and  unsteady  RANS  with  the  conventional turbulence models (either linear or non-linear eddy  viscosity  or  RSM models) should not be used because of their insufficient  accuracy  and wall-resolved LES because of its too high computational  cost  for  the considered high Re number flow. IDDES  can,  in  principle,  give  more accurate predictions than DDES but is shown to  be  more  sensitive  to numerics and span-size of the domain and so should be used with special care.
*The choice of  background  RANS  model  has  not  been  fully  studied, although in the current study SA-based DDES has  demonstrated  somewhat more "stable" (less code-dependent) behaviour and, in general, a better agreement with experiments than e.g. SST based DDES.
=== Transition modelling ===
*For the supercritical flow conditions or if the boundary layers on both cylinders  are  tripped  in  the  experiment,  no  specific  transition modelling or prescription is required and fully turbulent simulation is quite justified.
 
=== Near-wall modelling ===
All the simulations have been  carried  out  with  the  use  of  low-Re versions of background RANS models and so no recommendations
on the use of wall-functions are available.
 
== Application Uncertainties ==
== Application Uncertainties ==
{{Demo_UFR_BPA4}}
=== Time sample / statistical processing ===
#Ensure that the  simulation  has  bridged  the  initial  transient  and entered a statistically steady state before commencing  the  collection of statistics. This must either be checked by visual  inspection  of  a monitor signal (e.g. drag and lift coefficients) or  using  a  suitable statistical algorithm (see e.g. [[UFR_2-12_References#16|[16]]]<ref>A software implementing this algorithm for the estimation of statistical error and the detection of initial transient is commercially available: Contact info@cfd-berlin.com for details</ref>). As a guideline,  roughly  50&ndash;100 convective time units of initial transient can be expected.
#Compute  sufficiently  long  time  samples  for  reliable  statistical quantities. As a general  guideline,  a  minimum  time  sample  of  150 convective units is recommended. An objective evaluation of sufficiency of time sample  can  be  done  with  the  use  a  suitable  statistical algorithm (e.g. [[UFR_2-12_References#16|[16]]]<ref>Some participants of the BANC Workshops have  used  pure  LES  rather than DES-like approaches but these computations had the most  difficulty simulating the high Reynolds aspects of the flow [[UFR_2-12_References#5|[5]]].</ref>)  or  by  a  direct  checking  the  effect  of  a considerable increase of the time sample.
 
=== Comparability of CFD predictions and experiment ===
*The spanwise domain represents a key source of discrepancy between  the CFD and experimental configurations. Ideally,  inclusion  of  the  full experimental geometry (i.e. including the "floor" and "ceiling") in the CFD domain would be desirable, however  this  is  currently  associated with very high computational cost.
*Use the experimental data with transition trips on both  cylinders  [[UFR_2-12_References#4|[4]]] for best comparability with "fully-turbulent" CFD simulations.
 
== Recommendations for Future Work ==
== Recommendations for Future Work ==
{{Demo_UFR_BPA5}}
The following recommendations are made to improve further the understanding of this UFR:
*The most unsatisfactory aspect of the simulations is a wide  scatter of predictions of the rate of the initial roll-up of the shear layer separated  from  the  upstream  cylinder.  Efforts  to  improve  the situation by addressing a  "grey  area"  issue  in  hybrid  RANS-LES approaches would be most welcome.
<br/>
<br/>
----
----
<references />
{{ACContribs
{{ACContribs
|authors=A. Garbaruk, M. Shur and M. Strelets
|authors=A. Garbaruk, M. Shur and M. Strelets

Latest revision as of 12:07, 12 February 2017

Turbulent Flow Past Two-Body Configurations

Front Page

Description

Test Case Studies

Evaluation

Best Practice Advice

References

Flows Around Bodies

Underlying Flow Regime 2-12

Best Practice Advice

NOTE: the BPA formulated below are overall well in line with conclusions based on the outcome of BANC-I [5] and not yet published results of BANC-II.

Key Physics

The key physical features of this UFR are separation of the turbulent shear layer from the upstream cylinder, free shear layer roll-up and chaotization, interaction of the essentially unsteady wake of the upstream cylinder with the downstream one, and massively separated wake of the downstream cylinder. It is found that it is necessary to capture these challenging features in a simulation claiming a reliable prediction of all the UFR characteristics. Whether this is reached or not in a simulation should be checked by:

  • Obtaining a visual impression of the unsteadiness of the shear layer separated from the upstream cylinder and of a range of spatial scales present in its wake and in the wake of the downstream cylinder using e.g. a snapshot of isosurface of λ2 ("swirl") or Q-criterion (see Figure 4 for an example of the former) and vorticity contours (Figure 11).
  • Confirming a mixed tonal and broadband nature of the pressure signals on the surface of the cylinders by their spectral analysis (Figure 16).

Numerical Issues

In terms of numerics, based on experience accumulated in the course of the ATAAC and related projects, the following advice can be given:

  • In the Focus Region of simulation (see Figure 3), use numerical schemes with as low numerical dissipation as possible, particularly, pure or close to pure CDS for convective fluxes with the order of accuracy not less than 2. Acceptability of the level of numerical dissipation may be assessed by examining snapshots of e.g. vorticity in the focus region: the size of the smallest resolved eddies should not be noticeably larger than the local grid spacing.
  • In the Euler and Departure Regions (see Figure 3) use a scheme with sufficient numerical dissipation to prevent grid oscillations or "wiggles" in these regions.
  • Use a minimum second order accurate temporal integration scheme.
  • Use a time step sufficiently fine to capture the motion of the turbulent eddies resolved by the grid in the Focus Region. This corresponds to the approximate guideline CFLmax ≈ 1.

In terms of grids, although no systematic grid-sensitivity studies for the considered UFR have been carried out, indirect evidence allows the following recommendations:

  • In the Focus Region, use nearly isotropic grids with sizes not larger than around 0.02D, although even smaller values are desirable.
  • Outside the Focus Region, expand the grid cell size gradually towards the inflow/outflow boundaries, avoiding sudden jumps.

Size of computational domain should not be less than about ±20D in the streamwise and about ±10D in the lateral direction. For the spanwise direction, the domain size should not be less than 3D, but larger domains are strongly recommended provided that available computer resources allow this.

Physical Modelling Issues

Turbulence modelling

  • Use hybrid turbulence-resolving approaches, e.g. DDES or similar methods capable of treating the entire turbulent cylinder boundary layer with RANS. Steady and unsteady RANS with the conventional turbulence models (either linear or non-linear eddy viscosity or RSM models) should not be used because of their insufficient accuracy and wall-resolved LES because of its too high computational cost for the considered high Re number flow. IDDES can, in principle, give more accurate predictions than DDES but is shown to be more sensitive to numerics and span-size of the domain and so should be used with special care.
  • The choice of background RANS model has not been fully studied, although in the current study SA-based DDES has demonstrated somewhat more "stable" (less code-dependent) behaviour and, in general, a better agreement with experiments than e.g. SST based DDES.

Transition modelling

  • For the supercritical flow conditions or if the boundary layers on both cylinders are tripped in the experiment, no specific transition modelling or prescription is required and fully turbulent simulation is quite justified.

Near-wall modelling

All the simulations have been carried out with the use of low-Re versions of background RANS models and so no recommendations on the use of wall-functions are available.

Application Uncertainties

Time sample / statistical processing

  1. Ensure that the simulation has bridged the initial transient and entered a statistically steady state before commencing the collection of statistics. This must either be checked by visual inspection of a monitor signal (e.g. drag and lift coefficients) or using a suitable statistical algorithm (see e.g. [16][1]). As a guideline, roughly 50–100 convective time units of initial transient can be expected.
  2. Compute sufficiently long time samples for reliable statistical quantities. As a general guideline, a minimum time sample of 150 convective units is recommended. An objective evaluation of sufficiency of time sample can be done with the use a suitable statistical algorithm (e.g. [16][2]) or by a direct checking the effect of a considerable increase of the time sample.

Comparability of CFD predictions and experiment

  • The spanwise domain represents a key source of discrepancy between the CFD and experimental configurations. Ideally, inclusion of the full experimental geometry (i.e. including the "floor" and "ceiling") in the CFD domain would be desirable, however this is currently associated with very high computational cost.
  • Use the experimental data with transition trips on both cylinders [4] for best comparability with "fully-turbulent" CFD simulations.

Recommendations for Future Work

The following recommendations are made to improve further the understanding of this UFR:

  • The most unsatisfactory aspect of the simulations is a wide scatter of predictions of the rate of the initial roll-up of the shear layer separated from the upstream cylinder. Efforts to improve the situation by addressing a "grey area" issue in hybrid RANS-LES approaches would be most welcome.



  1. A software implementing this algorithm for the estimation of statistical error and the detection of initial transient is commercially available: Contact info@cfd-berlin.com for details
  2. Some participants of the BANC Workshops have used pure LES rather than DES-like approaches but these computations had the most difficulty simulating the high Reynolds aspects of the flow [5].


Contributed by: A. Garbaruk, M. Shur and M. Strelets — New Technologies and Services LLC (NTS) and St.-Petersburg State Polytechnic University

Front Page

Description

Test Case Studies

Evaluation

Best Practice Advice

References


© copyright ERCOFTAC 2024