Material property parameters.
142 views
Skip to first unread message
bryan...@nist.gov
Jan 14, 2009, 2:34:41 PM1/14/09
to FDS and Smokeview Discussions
This is the new thread for the discussion of methods of determining
material properties for FDS Input. This thread was started in off
topic discussions in this thread:
http://groups.google.com/group/fds-smv/browse_thread/thread/594202d954a62a56
Please continue the discussion here... starting with the following
threads:
Chris Lautenberger: http://groups.google.com/group/fds-smv/msg/ddfa81358a85340b
"One difficulty associated with material property estimation is that
not much can be directly measured (density, that's about it). Instead,
properties are INFERRED from experimental measurements (Cone, TGA,
DSC, etc.). Take the simplest example: thermal inertia and ignition
temperature are inferred from Cone Calorimeter ignition time plots via
the classical thermally thick ignition model.
We basically do the same thing with generalized pyrolysis models (such
as that in FDS), but niow there's 10 or more adjustable parameters
(material properties) to determine because the physics are treated in
greater detail. For a simple charring material, there'as at least 11:
kv, cv, kc, cc, rhoc, ev, ec, A, E, n, DHvol.
Simultaneous optimization of 10+ parameters is a major computer
science challenge. To date, people have tried stochastic optimization
(genetic algorithms), but tens of thousands or even hundreds of
thousands of trial solutions are usually necessary due to the large
number of degrees of freedom, so you have two choices: wait a long
time for your answer, or use parallel processing (even then you may
still wait a long time). Rarely does one get "good" results on the
first optimization run as these algorithms are not magic, they're just
a tool (and a tweaky tool at that).
There are much more sophisticated algorithms out there than have been
applied in the fire community to date--genetic algorithms, simulated
annealing, and hybrid methods. A badly needed contribution (and a
great project for an MS or PhD student) is development of a search/
optimization algorithm designed specifically for material property
estimation. "
Jason Floyd: http://groups.google.com/group/fds-smv/msg/e853326c76014312
"Such an approach is discussed in thes publication that Simo provided
a
link in the current thread on PU properties:
http://www.sal.hut.fi/Publications/pdf-files/TMAT08.pdf
(it was also presented at the last IAFSS)"
Simo Hostikka: http://groups.google.com/group/fds-smv/msg/d9678ac54163e903
"I would like to continue the discussion about material property
estimation in a new thread. In threads "GUI for FDS, state and
developments" and "Polyurethane Foam Reaction properties..", several
comments have been made about using reaction schemes, finding the
values, role of NIST and mathematical methods. They were getting
off-topic.
One point that I have found very challenging, is the fact that for
practical materials, knowing the material properties (or rather model
parameters) is not enough because the actual fire performance of the
_product_ is determined by 1) the material properties, and 2)
construction. Even accurate knowledge of the material properties may
not
yield good prediction of cone calorimeter test.
Let's take electrical cables as an example. Even though the physical
construction of cables may be quite far from the assumptions behind
the
FDS condenced phase solver, that's what we constantly end up
simulating.
When doing this, I have found that the actual construction of the
cable
is quite important. In addition to the material layers (typically
sheath, filler, insulator, conductor), there may be layers of thin
plastic or metal sheets, metal bands enhancing electrical properties
and
strength, etc. These layers may or may not be permeable by the
pyrolysis
products. Also, depending on the chemical compounds used, the char
residue of the materials may be impermeable to some extent - an
additional mechamism of fire retardancy.
How to handle these 'construction effects' when doing the parameter
estimation? So far, we have not described the permeability issues at
all, but rather tweaked the thermal and reaction parameters to obtain
as
good cone calorimeter performance as possible. But is this the right
way?
Yet another question is, how many different tests (TGA, cone, lateral
flame spread, SBI, ...) one has to use for estimation and validation
to
be confident that the model gives a good prediction of the actual
application."
Chris Lautenberger: http://groups.google.com/group/fds-smv/msg/3ebe16d392e90511
"Anna's MS thesis and her IAFSS paper applied a genetic algorithm.
Although genetic algorithms were proposed for property estimation in
some early work (e.g., http://dx.doi.org/10.1016/j.firesaf.2005.12.004,
among others) and have been used by several research groups since
then, there are more efficient search/optimization algorithms than
straight GA that have not yet been explored for this purpose. GA might
be a good starting point, but it is very tweaky, not always efficient,
and sometimes doesn't even give a good answer. A more efficient
search/
optimization algorithm (simulated annealing, hybrid methods, etc.)
would save frustration and headaches. Point is that property
estimation is not "solved" from a computer science or algorithmic
standpoint."
Thank you,
-Bryan
material properties for FDS Input. This thread was started in off
topic discussions in this thread:
http://groups.google.com/group/fds-smv/browse_thread/thread/594202d954a62a56
Please continue the discussion here... starting with the following
threads:
Chris Lautenberger: http://groups.google.com/group/fds-smv/msg/ddfa81358a85340b
"One difficulty associated with material property estimation is that
not much can be directly measured (density, that's about it). Instead,
properties are INFERRED from experimental measurements (Cone, TGA,
DSC, etc.). Take the simplest example: thermal inertia and ignition
temperature are inferred from Cone Calorimeter ignition time plots via
the classical thermally thick ignition model.
We basically do the same thing with generalized pyrolysis models (such
as that in FDS), but niow there's 10 or more adjustable parameters
(material properties) to determine because the physics are treated in
greater detail. For a simple charring material, there'as at least 11:
kv, cv, kc, cc, rhoc, ev, ec, A, E, n, DHvol.
Simultaneous optimization of 10+ parameters is a major computer
science challenge. To date, people have tried stochastic optimization
(genetic algorithms), but tens of thousands or even hundreds of
thousands of trial solutions are usually necessary due to the large
number of degrees of freedom, so you have two choices: wait a long
time for your answer, or use parallel processing (even then you may
still wait a long time). Rarely does one get "good" results on the
first optimization run as these algorithms are not magic, they're just
a tool (and a tweaky tool at that).
There are much more sophisticated algorithms out there than have been
applied in the fire community to date--genetic algorithms, simulated
annealing, and hybrid methods. A badly needed contribution (and a
great project for an MS or PhD student) is development of a search/
optimization algorithm designed specifically for material property
estimation. "
Jason Floyd: http://groups.google.com/group/fds-smv/msg/e853326c76014312
"Such an approach is discussed in thes publication that Simo provided
a
link in the current thread on PU properties:
http://www.sal.hut.fi/Publications/pdf-files/TMAT08.pdf
(it was also presented at the last IAFSS)"
Simo Hostikka: http://groups.google.com/group/fds-smv/msg/d9678ac54163e903
"I would like to continue the discussion about material property
estimation in a new thread. In threads "GUI for FDS, state and
developments" and "Polyurethane Foam Reaction properties..", several
comments have been made about using reaction schemes, finding the
values, role of NIST and mathematical methods. They were getting
off-topic.
One point that I have found very challenging, is the fact that for
practical materials, knowing the material properties (or rather model
parameters) is not enough because the actual fire performance of the
_product_ is determined by 1) the material properties, and 2)
construction. Even accurate knowledge of the material properties may
not
yield good prediction of cone calorimeter test.
Let's take electrical cables as an example. Even though the physical
construction of cables may be quite far from the assumptions behind
the
FDS condenced phase solver, that's what we constantly end up
simulating.
When doing this, I have found that the actual construction of the
cable
is quite important. In addition to the material layers (typically
sheath, filler, insulator, conductor), there may be layers of thin
plastic or metal sheets, metal bands enhancing electrical properties
and
strength, etc. These layers may or may not be permeable by the
pyrolysis
products. Also, depending on the chemical compounds used, the char
residue of the materials may be impermeable to some extent - an
additional mechamism of fire retardancy.
How to handle these 'construction effects' when doing the parameter
estimation? So far, we have not described the permeability issues at
all, but rather tweaked the thermal and reaction parameters to obtain
as
good cone calorimeter performance as possible. But is this the right
way?
Yet another question is, how many different tests (TGA, cone, lateral
flame spread, SBI, ...) one has to use for estimation and validation
to
be confident that the model gives a good prediction of the actual
application."
Chris Lautenberger: http://groups.google.com/group/fds-smv/msg/3ebe16d392e90511
"Anna's MS thesis and her IAFSS paper applied a genetic algorithm.
Although genetic algorithms were proposed for property estimation in
some early work (e.g., http://dx.doi.org/10.1016/j.firesaf.2005.12.004,
among others) and have been used by several research groups since
then, there are more efficient search/optimization algorithms than
straight GA that have not yet been explored for this purpose. GA might
be a good starting point, but it is very tweaky, not always efficient,
and sometimes doesn't even give a good answer. A more efficient
search/
optimization algorithm (simulated annealing, hybrid methods, etc.)
would save frustration and headaches. Point is that property
estimation is not "solved" from a computer science or algorithmic
standpoint."
Thank you,
-Bryan
Rein
Jan 15, 2009, 7:52:25 AM1/15/09
to FDS and Smokeview Discussions
Regarding the discussion on different search methods to optimize the
kinetics parameters, this topic has also been explored outside fire/
combustion science and there are plenty of similar works.
Indeed, an recent and interesting paper from a Chemical Engineering
stand point is:
* http://dx.doi.org/10.1016/j.cej.2007.05.024
"Hybrid genetic algorithm to find the best model and the globally
optimized overall kinetics parameters for thermal decomposition of
plastics" (Chemical Engineering Journal 138, 2008)
This work uses Hybrid Genetic Algorithm (HGA) to improve the fine
tuning capabilities as compared to Genetic Algorithm (GA) alone. As
the authors say, "GA employs a probabilistic approach and has better
global optimizing properties but shows poor convergence to optimality.
Whereas, HGA using a typical basic GA with elitist strategy to reach
near gradient or/direct-based search method shows faster convergence
to global optima". The observation that GA fails in the fine tuning
near the optimum was already mentioned in the pioneering work of
Alander et al in 1994 (ftp.uwasa.fi/cs/report94-2/Alander3.ps.Z).
Another paper, this time applied to a flame-retarded epoxy resin:
* http://dx.doi.org/10.1016/j.polymdegradstab.2007.07.011
"Global kinetics of thermal degradation of flame-retarded epoxy resin
formulations" (Polymer Degradation and Stability 92, 2007)
They used Powell minimisation (a gradient based method) as the
optimization technique. The results are promising but the model fails
to predict TGA data a different heating rates, which is a step
backwards compared to previous work that do capture different heating
rates.
G.
On Jan 14, 7:34 pm, bryan.kl...@nist.gov wrote:
> This is the new thread for the discussion of methods of determining
> material properties for FDS Input. This thread was started in off
Reply all
Reply to author
Forward
0 new messages