Convergence study, mesh sensitivity on heat flux to the wall?
715 views
Skip to first unread message
chen...@gmail.com
Apr 3, 2009, 10:21:36 AM4/3/09
to FDS and Smokeview Discussions
Hi, dear All.
For fire simulation, is it necessary to refine the mesh and do the
convergence study? since it is really expensive to do the convergence
study on large scale fire.
I know when the mesh is refine enough, LES will converge to DNS.
Since fire is turbulence based, LES based on coarse mesh should be
enough to give reasonable results.
But I still have a concern what is the mesh sensitivity on heat flux
to the wall? How refined mesh is good enough?
Any information is appreciated. Thanks in advance.
For fire simulation, is it necessary to refine the mesh and do the
convergence study? since it is really expensive to do the convergence
study on large scale fire.
I know when the mesh is refine enough, LES will converge to DNS.
Since fire is turbulence based, LES based on coarse mesh should be
enough to give reasonable results.
But I still have a concern what is the mesh sensitivity on heat flux
to the wall? How refined mesh is good enough?
Any information is appreciated. Thanks in advance.
Kevin
Apr 3, 2009, 10:26:30 AM4/3/09
to FDS and Smokeview Discussions
> For fire simulation, is it necessary to refine the mesh and do the
> convergence study? since it is really expensive to do the convergence
> study on large scale fire.
Yes, it is necessary. I suggest you do a grid sensitivity study on the
> convergence study? since it is really expensive to do the convergence
> study on large scale fire.
part of the domain that contains the fire and plume, at the very
least.
> I know when the mesh is refine enough, LES will converge to DNS.
> Since fire is turbulence based, LES based on coarse mesh should be
> enough to give reasonable results.
> But I still have a concern what is the mesh sensitivity on heat flux
> to the wall? How refined mesh is good enough?
chen...@gmail.com
Apr 3, 2009, 10:48:57 AM4/3/09
to FDS and Smokeview Discussions
I notice the mesh in aLl the FDS example cases are quite coarse.
In these situations, LES is used rather than DNS?
How good the results are compared with DNS on refined mesh? Are they
good enough?
In these situations, LES is used rather than DNS?
How good the results are compared with DNS on refined mesh? Are they
good enough?
Kevin
Apr 3, 2009, 11:47:32 AM4/3/09
to FDS and Smokeview Discussions
We cannot do DNS simulations of large compartments. We use LES, and we
compare our predictions with experimental data. "Good enough" is
really up to you to decide.
> > You are the end user and you should determine this on your own.- Hide quoted text -
>
> - Show quoted text -
compare our predictions with experimental data. "Good enough" is
really up to you to decide.
>
> - Show quoted text -
daserra
Apr 3, 2009, 3:02:19 PM4/3/09
to FDS and Smokeview Discussions
In the context of turbulence you should first define convergence. In
the context of a numerical scheme convergence is quite clear but this
definition does not apply to turbulence. Just look at the Kolmogorov
curve: the finer the grid the more scales are resolved.
the context of a numerical scheme convergence is quite clear but this
definition does not apply to turbulence. Just look at the Kolmogorov
curve: the finer the grid the more scales are resolved.
Dimitrios
Apr 5, 2009, 11:18:54 AM4/5/09
to FDS and Smokeview Discussions
Whenever you solve equations for a specific quantity you can expect
convergence.
If the quantity depends from the grid itself then you solves for these
quantity for which you have choose the cell width.
convergence.
If the quantity depends from the grid itself then you solves for these
quantity for which you have choose the cell width.
Kevin
Apr 5, 2009, 1:49:40 PM4/5/09
to FDS and Smokeview Discussions
I do not understand what you mean.
Dimitrios
Apr 5, 2009, 3:42:22 PM4/5/09
to FDS and Smokeview Discussions
sorry, i mean that the question is for what a quantity stands the
degree of freedom in the discretised equations.
If the underlying differential equations stands for a specific
velocity or temperature field, then the answer of the solution to the
discretised counterparts should be exactly these specific field
regardless of the cell width.
In such a case one could speak about a certain filtered field.
If the equations stands for a velocity field which depends on the cell
width, for example the portion of the energie which is tracked with
the directly resolved eddies, then the solution is no longer grid
independent. In this case i wan't to solve for a quantity which is
predifined via the set of underlying equations but for a quantity
which is directly represented through the discretised equations and i
will get a resolved field (a portion of the field).
If i decide what a portion of the field i want to resolve, then it
makes no sense to look for convergence because for different cell
widths i ask also for different results.
degree of freedom in the discretised equations.
If the underlying differential equations stands for a specific
velocity or temperature field, then the answer of the solution to the
discretised counterparts should be exactly these specific field
regardless of the cell width.
In such a case one could speak about a certain filtered field.
If the equations stands for a velocity field which depends on the cell
width, for example the portion of the energie which is tracked with
the directly resolved eddies, then the solution is no longer grid
independent. In this case i wan't to solve for a quantity which is
predifined via the set of underlying equations but for a quantity
which is directly represented through the discretised equations and i
will get a resolved field (a portion of the field).
If i decide what a portion of the field i want to resolve, then it
makes no sense to look for convergence because for different cell
widths i ask also for different results.
rmcdermo
Apr 6, 2009, 9:29:11 AM4/6/09
to FDS and Smokeview Discussions
Demitrios,
Your point is well taken. FDS does not perform explicit filtering,
and so as we change the grid spacing h we also change the filter width
Delta. In a discussion such as this it is imperative that we make a
distinction between these quantities.
You made the following comment below, "Whenever you solve equations
words: For a given Delta (filter width) we can derive a set of PDEs
that we call the LES equations. Now, using finer and finer grids h/D -
> 0, we can expect convergence of the numerical solution to the exact
solution of the PDE. The seminal work of Ghosal in 1995 showed that
for a second order scheme you require D/h = 4 and for a fourth order
scheme you require D/h = 2 in order for the numerical errors to be
sufficiently smaller than the magnitude of the physical subgrid terms.
If this were indeed essential then we would be in some trouble because
for practical calculations D/h = 4 becomes prohibitively expensive and
development of an engineering code that is truly fourth order accurate
is harder than it sounds. Somewhat fortuitously, experience has shown
us that we can get away with D/h = 1 by using energy-conserving second-
order numerics combined with physically-based subgrid models. From an
engineering perspective what we hope to achieve is convergence of the
statistical quantity of interest. In this sense, I am echoing
daserra's point. Suppose we are interested in the mean temperature in
the upper layer of a compartment. Let us denote the actual mean temp
(that which we would obtain from an experiment) as <T(x,t)>. LES
(FDS) produces a filtered field \bar{T}(x,t). By statistical
convergence we mean that we want <\bar{T}(x,t)> -> <T(x,t)> as Delta =
h -> 0.
Similar ideas are more clearly spelled in
Pope, S.B. "Ten questions concerning the large-eddy simulation of
turbulent flows", New Journal of Physics, 2004.
To expect statistical convergence from an LES code like FDS, a guiding
principle of development should be that the subgrid terms vanish as
Delta = h -> 0. I call this LES to DNS convergence. For practical
reasons, this has not always been a priority in FDS, but we are
transitioning as best we can to subgrid models with this property. As
examples, I will cite the inclusion of the dynamic Smagorinsky model
and the Werner and Wengle wall model, which are likely to become
defaults in the near future.
Cheers,
Randy
Your point is well taken. FDS does not perform explicit filtering,
and so as we change the grid spacing h we also change the filter width
Delta. In a discussion such as this it is imperative that we make a
distinction between these quantities.
You made the following comment below, "Whenever you solve equations
for a specific quantity you can expect
convergence." I agree, but let me just say this in slightly different
words: For a given Delta (filter width) we can derive a set of PDEs
that we call the LES equations. Now, using finer and finer grids h/D -
> 0, we can expect convergence of the numerical solution to the exact
solution of the PDE. The seminal work of Ghosal in 1995 showed that
for a second order scheme you require D/h = 4 and for a fourth order
scheme you require D/h = 2 in order for the numerical errors to be
sufficiently smaller than the magnitude of the physical subgrid terms.
If this were indeed essential then we would be in some trouble because
for practical calculations D/h = 4 becomes prohibitively expensive and
development of an engineering code that is truly fourth order accurate
is harder than it sounds. Somewhat fortuitously, experience has shown
us that we can get away with D/h = 1 by using energy-conserving second-
order numerics combined with physically-based subgrid models. From an
engineering perspective what we hope to achieve is convergence of the
statistical quantity of interest. In this sense, I am echoing
daserra's point. Suppose we are interested in the mean temperature in
the upper layer of a compartment. Let us denote the actual mean temp
(that which we would obtain from an experiment) as <T(x,t)>. LES
(FDS) produces a filtered field \bar{T}(x,t). By statistical
convergence we mean that we want <\bar{T}(x,t)> -> <T(x,t)> as Delta =
h -> 0.
Similar ideas are more clearly spelled in
Pope, S.B. "Ten questions concerning the large-eddy simulation of
turbulent flows", New Journal of Physics, 2004.
To expect statistical convergence from an LES code like FDS, a guiding
principle of development should be that the subgrid terms vanish as
Delta = h -> 0. I call this LES to DNS convergence. For practical
reasons, this has not always been a priority in FDS, but we are
transitioning as best we can to subgrid models with this property. As
examples, I will cite the inclusion of the dynamic Smagorinsky model
and the Werner and Wengle wall model, which are likely to become
defaults in the near future.
Cheers,
Randy
Dimitrios
Apr 6, 2009, 11:43:09 AM4/6/09
to FDS and Smokeview Discussions
I agree with your post.
Also i agree with Kevins suggestions to do sensitivity analysis of the
results rather than convergence investigation.
At the moment the only method to verify the correspondence between the
simulation and the reality is comparison with experiments and
sensitivity analysis to examine influence parameters.
Also i agree with Kevins suggestions to do sensitivity analysis of the
results rather than convergence investigation.
At the moment the only method to verify the correspondence between the
simulation and the reality is comparison with experiments and
sensitivity analysis to examine influence parameters.
daserra
Apr 6, 2009, 3:23:49 PM4/6/09
to FDS and Smokeview Discussions
This is a quite interesting discussion but it should not be limited to
the turbulence problem. One has to look for integral parameter. E.g
the HRR curve and the other parameters in the hrr-file are good
indicators of convergence. The radiation or the wall temperature may
also be a good parameter.
In fully develloped turbulence one has the Kolmogorov curve of scales.
Then - if the grid is chosen in the inertial range AND if the
Smagorinski is an appropriate model AND the Smagorinski factor is a
good choice - a change in grid size should no have a substantial
effect on integral parameters.
I did some numerical experiments and I have found that 1/10 of the
fire scale is an excellent choice for the grid.
the turbulence problem. One has to look for integral parameter. E.g
the HRR curve and the other parameters in the hrr-file are good
indicators of convergence. The radiation or the wall temperature may
also be a good parameter.
In fully develloped turbulence one has the Kolmogorov curve of scales.
Then - if the grid is chosen in the inertial range AND if the
Smagorinski is an appropriate model AND the Smagorinski factor is a
good choice - a change in grid size should no have a substantial
effect on integral parameters.
I did some numerical experiments and I have found that 1/10 of the
fire scale is an excellent choice for the grid.
rmcdermo
Apr 7, 2009, 10:26:18 AM4/7/09
to FDS and Smokeview Discussions
daserra,
So it sounds like you agree with Kevin's h = D*/10 recommendation(?).
In the LES literature there is talk of "quality measures". The idea
is to define a suitable scalar that is a measure of local quality of
the LES solution. In principle, this scalar can then be used to
decide whether grid refinement is necessary.
In Pope's "10 questions" paper he suggests using the fraction of
resolved kinetic energy. The reasoning behind this is that for
homogeneous isotropic turbulence it can be shown that resolving 80% of
the kinetic energy puts you into the inertial subrange on the
Kolmogorov spectral curve.
In FDS we have implemented a new output QUANTITY = 'TURBULENCE
RESOLUTION', which is an estimate of the fraction of subgrid kinetic
energy. The scalar in the code is 0 < MTR < 1, where MTR stand for
'measure of turbulence resolution'. 0=fully resolved. So, if MTR<0.2
you should have a high quality LES.
Of course, 0.2 is based on a canonical flow. We need to start
amassing experience with this number to see if it is helpful for
engineering flows. I just wanted to make you aware of it. I have
looked at the time average of MTR for a helium plume and actually 0.2
seems to be a good value to shoot for in this case. That is, when
<MTR> = 0.2 I get good agreement between FDS and experiment.
Cheers,
Randy
So it sounds like you agree with Kevin's h = D*/10 recommendation(?).
In the LES literature there is talk of "quality measures". The idea
is to define a suitable scalar that is a measure of local quality of
the LES solution. In principle, this scalar can then be used to
decide whether grid refinement is necessary.
In Pope's "10 questions" paper he suggests using the fraction of
resolved kinetic energy. The reasoning behind this is that for
homogeneous isotropic turbulence it can be shown that resolving 80% of
the kinetic energy puts you into the inertial subrange on the
Kolmogorov spectral curve.
In FDS we have implemented a new output QUANTITY = 'TURBULENCE
RESOLUTION', which is an estimate of the fraction of subgrid kinetic
energy. The scalar in the code is 0 < MTR < 1, where MTR stand for
'measure of turbulence resolution'. 0=fully resolved. So, if MTR<0.2
you should have a high quality LES.
Of course, 0.2 is based on a canonical flow. We need to start
amassing experience with this number to see if it is helpful for
engineering flows. I just wanted to make you aware of it. I have
looked at the time average of MTR for a helium plume and actually 0.2
seems to be a good value to shoot for in this case. That is, when
<MTR> = 0.2 I get good agreement between FDS and experiment.
Cheers,
Randy
JWilliamson
Apr 7, 2009, 1:25:50 PM4/7/09
to FDS and Smokeview Discussions
Randy,
Is this output quantity available in v5.3?
Do you have any additional details about the definition of the
variable that you can share?
Thanks
Is this output quantity available in v5.3?
Do you have any additional details about the definition of the
variable that you can share?
Thanks
rmcdermo
Apr 9, 2009, 9:18:23 AM4/9/09
to FDS and Smokeview Discussions
Justin,
It is certainly available in 5.3.1, which I believe Kevin is releasing
today.
Something I forgot to mention is that in order to output
QUANITY='TURBULENCE RESOLUTION' you need to set CHECK_KINETIC_ENERGY=T
on MISC.
The metric is defined locally in space and time as
M(x,t) = k_sgs/(k_les + k_sgs)
where
k_les = U_i U_i (sum over i=1,3)
k_sgs = u_i u_i
and
u_i = U_i - \hat{U_i}
is an approximation for the subgrid velocity. U_i is the LES
(resolved) velocity (taken after the corrector step). The \hat{ }
operator is a test filter of width 2*dx.
When you time average this quantity in SMV you get something like what
Pope defines as M in his "10 questions" paper I keep harping about.
I'll be interested in any experience you have with this.
Cheers,
Randy
On Apr 7, 1:25 pm, JWilliamson <williamson.justin.w...@gmail.com>
wrote:
It is certainly available in 5.3.1, which I believe Kevin is releasing
today.
Something I forgot to mention is that in order to output
QUANITY='TURBULENCE RESOLUTION' you need to set CHECK_KINETIC_ENERGY=T
on MISC.
The metric is defined locally in space and time as
M(x,t) = k_sgs/(k_les + k_sgs)
where
k_les = U_i U_i (sum over i=1,3)
k_sgs = u_i u_i
and
u_i = U_i - \hat{U_i}
is an approximation for the subgrid velocity. U_i is the LES
(resolved) velocity (taken after the corrector step). The \hat{ }
operator is a test filter of width 2*dx.
When you time average this quantity in SMV you get something like what
Pope defines as M in his "10 questions" paper I keep harping about.
I'll be interested in any experience you have with this.
Cheers,
Randy
On Apr 7, 1:25 pm, JWilliamson <williamson.justin.w...@gmail.com>
wrote:
Deepak
Apr 9, 2009, 9:32:07 AM4/9/09
to fds...@googlegroups.com
Randy,
This output QUANTITY i have tried with CHECK_KINETIC_ENERGY=.TRUE. in FDS 5.3.0, but i have got this error message " forrtl: severe <157>: program exception - access violation". How can I rectify this??
BR
Deepak
This output QUANTITY i have tried with CHECK_KINETIC_ENERGY=.TRUE. in FDS 5.3.0, but i have got this error message " forrtl: severe <157>: program exception - access violation". How can I rectify this??
BR
Deepak
rmcdermo
Apr 9, 2009, 12:30:29 PM4/9/09
to FDS and Smokeview Discussions
Deepak,
Please post this issue to the Issue Tracker along with a simplified
(i.e., a case that should run quickly, if it runs) input file.
Thanks,
Randy
On Apr 9, 9:32 am, Deepak <deepak.gan...@gmail.com> wrote:
> Randy,
>
> This output QUANTITY i have tried with CHECK_KINETIC_ENERGY=.TRUE. in FDS
> 5.3.0, but i have got this error message " forrtl: severe <157>: program
> exception - access violation". How can I rectify this??
>
> BR
> Deepak
>
Please post this issue to the Issue Tracker along with a simplified
(i.e., a case that should run quickly, if it runs) input file.
Thanks,
Randy
On Apr 9, 9:32 am, Deepak <deepak.gan...@gmail.com> wrote:
> Randy,
>
> This output QUANTITY i have tried with CHECK_KINETIC_ENERGY=.TRUE. in FDS
> 5.3.0, but i have got this error message " forrtl: severe <157>: program
> exception - access violation". How can I rectify this??
>
> BR
> Deepak
>
JWilliamson
Apr 9, 2009, 2:39:39 PM4/9/09
to FDS and Smokeview Discussions
Randy,
I like this quantity for resolving turbulence. I can also see why
M=0.2 might be a good value to shoot for in turbulent plumes. I have
seen several studies where time averaged turbulent statistics say that
the vertical velocity fluctuations are very nearly
U = U_bar + u -> u_rms ~ 0.3 U_bar
Assuming that the horizontal and in plane statistics are similar, but
about 1/2 the magnitude of the vertical component
v_rms ~ w_rms ~ u_rms/2 ~ 0.15 U_bar
Calculating a characteristic value for M where U_bar >> V_bar and
W_bar you get
k_les ~ U_bar^2
k_sgs ~ u_rms^2 + v_rms^2 + w_rms^2 ~ 0.135 U_bar^2
M ~ (0.135 U_bar^2)/(U_bar^2 + 0.135 U_bar^2) ~ 0.12
This seems to agree with your choice of M=0.2 as a good characteristic
value. Of course, I just did this on a whim on the back of a scrap
piece of paper, so I would appreciate it if you notice any errors I've
made. In the mean time, I'll look into the quantity when I have
opportunities to use 5.3.1.
Justin
I like this quantity for resolving turbulence. I can also see why
M=0.2 might be a good value to shoot for in turbulent plumes. I have
seen several studies where time averaged turbulent statistics say that
the vertical velocity fluctuations are very nearly
U = U_bar + u -> u_rms ~ 0.3 U_bar
Assuming that the horizontal and in plane statistics are similar, but
about 1/2 the magnitude of the vertical component
v_rms ~ w_rms ~ u_rms/2 ~ 0.15 U_bar
Calculating a characteristic value for M where U_bar >> V_bar and
W_bar you get
k_les ~ U_bar^2
k_sgs ~ u_rms^2 + v_rms^2 + w_rms^2 ~ 0.135 U_bar^2
M ~ (0.135 U_bar^2)/(U_bar^2 + 0.135 U_bar^2) ~ 0.12
This seems to agree with your choice of M=0.2 as a good characteristic
value. Of course, I just did this on a whim on the back of a scrap
piece of paper, so I would appreciate it if you notice any errors I've
made. In the mean time, I'll look into the quantity when I have
opportunities to use 5.3.1.
Justin
Tony Bova
Apr 9, 2009, 3:16:07 PM4/9/09
to FDS and Smokeview Discussions
Deepak & Randy-
I have had a similar error (segmentation fault) when
CHECK_KINETIC_ENERGY=.TRUE. is added to the MISC section. I've posted
this as issue 696 in the issue tracker. Cheers, ~tony
I have had a similar error (segmentation fault) when
CHECK_KINETIC_ENERGY=.TRUE. is added to the MISC section. I've posted
this as issue 696 in the issue tracker. Cheers, ~tony
> Randy,
>
> This output QUANTITY i have tried with CHECK_KINETIC_ENERGY=.TRUE. in FDS
> 5.3.0, but i have got this error message " forrtl: severe <157>: program
> exception - access violation". How can I rectify this??
>
> BR
> Deepak
>
>
> This output QUANTITY i have tried with CHECK_KINETIC_ENERGY=.TRUE. in FDS
> 5.3.0, but i have got this error message " forrtl: severe <157>: program
> exception - access violation". How can I rectify this??
>
> BR
> Deepak
>
rmcdermo
Apr 9, 2009, 6:00:52 PM4/9/09
to FDS and Smokeview Discussions
All:
My apologies from piping up about a feature that I clearly had not
tested very well. Long story short, the fix to Deepak and Tony's
problem is committed with SVN 3748.
Justin:
This is interesting insight and I see nothing wrong with your
arguments.
The number M=0.2 comes from that fact that for homogeneous isotropic
turbulence resolving 80% of the kinetic energy gets you into the
inertial subrange regardless of the Reynolds number (well, provided it
is high enough to have an inertial range) -- see Pope Exercise 13.10.
This is of course useful for LES because a tacit assumption of most
subgrid models is that the smallest resolved scales of the LES are
within the inertial range. I suspect that 0.2 works well for plumes
because in a Lagrangian frame the plume looks a lot like decaying
isotropic turbulence.
Cheers,
Randy
My apologies from piping up about a feature that I clearly had not
tested very well. Long story short, the fix to Deepak and Tony's
problem is committed with SVN 3748.
Justin:
This is interesting insight and I see nothing wrong with your
arguments.
The number M=0.2 comes from that fact that for homogeneous isotropic
turbulence resolving 80% of the kinetic energy gets you into the
inertial subrange regardless of the Reynolds number (well, provided it
is high enough to have an inertial range) -- see Pope Exercise 13.10.
This is of course useful for LES because a tacit assumption of most
subgrid models is that the smallest resolved scales of the LES are
within the inertial range. I suspect that 0.2 works well for plumes
because in a Lagrangian frame the plume looks a lot like decaying
isotropic turbulence.
Cheers,
Randy
Reply all
Reply to author
Forward
0 new messages