Kinetic Model Completeness

5.68J/10.
652J Spring 2003 Lecture Notes Tuesday April 15, 2003
Kinetic Model Completeness

We say a chemical kinetic model is complete for a particular reaction condition when it
contains all the species and reactions needed to describe the chemical processes at that
reaction condition to some specified level of accuracy. In other words, the kinetic model
is complete when all the reactions whose rate constants it sets equal to zero really are
negligible for the specified reaction conditions and error tolerance.
Note that this definition of complete does NOT assume that the rate constants or any
other parameters in the model are correct, just that the reactions excluded from the model
are indeed negligible.
If you believe you can list all possible reactions of the species in the model, a necessary
condition for a model to be complete is that the rates of these reactions to form species
not included in the model must be smaller than some error tolerance; this approach was
introduced by R.G. Susnow et al. (1997). However, this condition is not sufficient: e.g.
one of the minor neglected species could be a catalyst or catalyst poison with a very big
influence on the kinetics at very low concentrations. If the kinetics are strictly linear you
can show that this necessary condition is sufficient (Matheu et al. 2002, 2003); an
important example of strictly linear kinetics are the ordinary master equations used to
compute pressure-dependent reaction rates.
Of course, you usually do not know all the possible reactions. But for purposes of kinetic
model reduction one usually assumes that the initial “full” model is complete; you are
satisfied if the reduced model reproduces the full model to within some error tolerances
under some reaction conditions.
Sensitivity Analyses
A. Definitions of Sensitivities
Suppose you have a model you think is complete at some reaction condition, and you
have some estimates for k and Yo so you can numerically solve
dY/dt = F(Y,k) Y(to) = Yo

to get your best prediction for the trajectory Y(t). (Y is normally made up of many mass
fractions yi(t) and a few other time-varying state variables e.g. T(t).) You never know the
rate (and other) parameters k and the initial concentrations Yo exactly, so you are always
interested in how much the predicted trajectory would vary if these values were a little
different than the values you assumed, i.e. how sensitive is your prediction to the values
of these parameters? A common way to express this is to write Y(t) as a Taylor expansion
in the parameters k and Yo:
yi(t) = yi(t; k,Yo) + Σ(∂yi(t)/∂kn)(kn’– kn)
+ Σ(∂yi(t)/∂yom)( yom’- yom) + …
The rate constants k usually depend on T and P, so if T and/or P vary with time the first
term gets to be a mess. The most popular approach (used in the latest versions of
CHEMKIN, but not consistently in earlier versions) is to imagine that all of the rate
constants kn(T,P) are multiplied by scaling factors Dn. Our predicted trajectory Y(t)
corresponds to all Dn=1. So the preferred expansion (for T,P varying) is:
yi(t) = yi(t; D=1,Yo) + Σ(∂yi(t)/∂Dn)(Dn’– 1)

+ Σ(∂yi(t)/∂yom)( yom’- yom) + …
The derivatives in this expansion are sometimes called the sensitivities, though – watch
out! - in practice what are usually reported are the corresponding unitless normalized
sensitivities:
“Normalized sensitivity of species i to reaction n”

= (1/yi(t))(∂yi(t)/∂Dn)= ∂(ln yi(t))/∂(ln Dn)
“Normalized sensitivity of species i to initial concentration of species m”
= (yom/yi(t))(∂yi(t)/∂yom) = ∂(ln yi(t))/∂(ln yom)
It must be emphasized that these sensitivities depend crucially on the assumed values of
the rate constants k and all the other parameters, and on (t-to) and Yo. If you change the
simulation in any way you will change all the computed sensitivities. The (first-order)
sensitivities are also very “local” in the sense that they are only useful if the perturbed
trajectory is very close to the original trajectory Y(t), i.e. if the perturbation in k or Yo is
very small. If you care about a big variation (e.g. if one of your rate constants is
extremely uncertain) you may not be able to get away with the first-order Taylor
expansion. If you need to go to higher orders or otherwise handle big
variations/uncertainties in the model parameters, see papers by H. Rabitz, G.J. McRae,
and Nancy Brown.
One would often like to know the “Green’s function sensitivity” ∂yi(t)/∂ym(t’) i.e. how
much species i at time t would have changed if a little bit of species m was added at time
t’. This can be computed as an ‘initial concentration sensitivity’ by setting to=t’ and
Yo=Y(t’). However, it is usually impractical to do this for many values of t’.
Finally, in addition to “concentration sensitivities”, one can also compute “rate

sensitivities”, which are how much the instantaneous slope dyi/dt would change if k or Yo
were perturbed. In addition to the ordinary integrated rate sensitivities (which depend on
t-to, and capture the fact that the whole trajectory is perturbed starting at to) one can also
compute “instantaneous rate sensitivities” which are how much the rate at time t would
change if the concentrations or rate constants were suddenly changed at time t. The
instantaneous rate sensitivities are much simpler to compute than the other sensitivities,
just a little algebra is required:
“Instantaneous species i rate sensitivity to reaction n” = ∂Fi(Y,k)/∂kn

“Instantaneous species i rate sensitivity to species m” = ∂Fi(Y,k)/∂ym
The instantaneous sensitivities do not depend on the history of the system, they will be
the same numbers regardless of how the system got to its present state Y.
Finally, sensitivities are partial derivatives, so it is important to know what other
variables are being held constant – you can get very different numerical values depending
on what you hold constant. In particular, it is important to know whether the
thermochemistry is being held constant, or whether all the forward and reverse rates are
individually being held constant. The former is generally preferred since you are usually
more confident in the thermochemical parameter values than the kinetic parameters, and
also there are usually fewer thermochemical parameters than reverse rate constants.
(Though if you are highly uncertain about the thermochemistry of some of your species,
you will also be interested in the sensitivity to those thermochemical parameters!). If the
forward and reverse rates are treated separately, it is very important to look at
sensitivities with respect to both the forward and the corresponding reverse rate
simultaneously: if they are opposite in sign and nearly equal in magnitude, it means you
are really sensitive to the thermochemistry, not the kinetics. (This usually happens when
that reaction is fast and nearly equilibrated.)
B. Computing Sensitivities
The other sensitivities are pretty hard to compute, they all require solving rather large
systems of differential equations. For example:
d(dyi(t)/dDn)/dt=d(dyi(t)/dt)/dDn=d(Fi(Y,D.*k)/dDn= Σ(∂Fi/∂ym)(dym/dDn) + ∂Fi/∂Dn
defining Sin(t) = dyi(t)/dDn evaluated on the reference trajectory Y(t) we can rewrite this:
dSin/dt = Σ(∂Fi/∂ym)Smn + ∂Fi/∂Dn Sin(to)=0

For the sensitivities to initial concentrations Ζil(t)=∂yi(t)/∂yol) the equation is very similar:
dZil/dt = Σ(∂Fi/∂ym)Zml Zil(to)=δil

In both cases, the right hand side depends on the Jacobian ∂F/∂y which depends on Y
(and so varies with time) so it is best to solve this simultaneously with the original system
of equations dY/dt = F, to avoid having to evaluate the Jacobian repeatedly. There are
several tricks which take advantage of the special simple form of the sensitivity equations
(e.g. linear in S), and good programs which use these tricks for computing the first-order
sensitivities, including DASAC, DASPK, and DAEPACK. For details, see for example
recent papers by P.I. Barton.
Note that there are a huge number of sensitivities O(Nspecies*Nreactions*Ntimesteps).

To speed the computations and reduce the output, one can only ask for the sensitivities to
a few parameters (rather than all the rate constants), or for the sensitivities of only a few
species. Exactly how much this saves you will depend on the details of the numerical
algorithm. The algorithms in CHEMKIN are not very efficient, and the post-processor is
worse, so you generally do much better if you only ask for the output you really need. If
you need to solve large sensitivity problems that CHEMKIN cannot handle, you might
want to ask Prof. Barton for advice.
C. Some Uses for Sensitivities

1) Estimating effects of Varying Initial Conditions
Very often you are interested in the behavior of your system as you make (relatively
small) variations in the initial conditions. For example, if you are doing an ignition
problem, what happens if the initial temperature is 10 K higher? In a chemical process,
what happens if you change the initial concentration of the one of the reagents by a factor
of two? Are we sensitive to small variations in any of the initial concentrations?
Y(t; Yo+∆Yo) = Y(t; Yo) + Z(t)*∆Yo
Are any of the Z’s for minor species variations greater than ε/Ymax, where Ymax is the
expected upper bound on contaminants in the input stream, and ε is the acceptable upper
bound on contaminants in the output stream?
2) Putting Error Bars on the Model Predictions

A first estimate of the uncertainty in the model prediction Y(t) is given by
|∆ln yi(t)| ∼ Σ|∂(ln yi(t))/∂(ln Dn)|*|∆ln kn|
+ Σ |∂(ln yi(t))/∂(ln yom)|*|∆ln yom|
The first term frequently dominates. Unfortunately, the uncertainties estimated in this
way are usually pretty large, accurately reflecting our inability to precisely predict
chemical kinetics. The errors in the predicted selectivities (i.e. relative yields) are usually
smaller than the errors in the predicted absolute concentration profiles.
|∆ln(yi(t)/yi(t))| ∼ Σ|Sin-Sjn|*|∆ln kn|
3) Identification of the Important Uncertainties
The sums above that estimate the total uncertainties are usually dominated by a few
terms. The uncertain parameters that correspond to these big terms are those that matter
the most for the simulation at hand; if you want a more accurate simulation you need to
tighten these error bars.
4) Design of Kinetic Experiments

In many kinetic experiments you measure a signal S(t), and what you want is for S(t) to
be very sensitive to the rate parameter k you are trying to determine, and insensitive to
the uncertainties in all the other parameters. Then you can have reasonable confidence
when you say that your observation of S(t) implies a certain value for k. When you
design an experiment, you are well-advised to immediately build a simulation model, and
check that your observable(s) really are sensitive to the rate constant you are trying to
determine.
5) Valid Range Analyses

Each kinetic model is only valid over a finite range of reaction conditions. Unfortunately,
this valid range is not usually known. For our model to be complete, we require that all
the ignored side reactions have negligible rates. By determining the sensitivity of these
side reactions to changes in the reaction conditions we can estimate the valid range, i.e.
the range over which the model will be complete. (This presumes that the rate parameters
in the model are not too far off.) See J. Song et al. Chem. Eng. Sci. 2002.
Flux Analyses
It is often very helpful to examine the rates at which each of the reactions in the system
are proceeding. In particular, you will want to know which reactions have high net rates.
By considering the forward and reverse rates separately, you can also identify the
reactions which are in quasi-equilibrium. Often only the reactions which have high net
rates, high sensitivities, or which are quasi-equilibrated really matter for the kinetics.
Almost certainly you cannot model the system correctly if you leave out a reaction with a
high net rate or a high sensitivity.
Very often it is interesting to know which reactions contribute the most to the
instantaneous formation or destruction of a key species. Often a species will be involved
in dozens of reactions, but only three or four have significant net rates. CHEMKIN will
make these net rate plots for you.
A problem with net rates is that all the species in quasi-steady-state, and all the fast
reactions which are nearly equilibrated, will have net rates near zero. It is therefore often
interesting to look at the formation and destruction rates or the forward and reverse rates
separately.

Kinetic Model Completeness

Uploaded by

Copyright:

Available Formats

Kinetic Model Completeness

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Kinetic Model Completeness

Uploaded by

Copyright:

Available Formats

5.68J/10.

652J Spring 2003 Lecture Notes Tuesday April 15, 2003

Kinetic Model Completeness

dY/dt = F(Y,k) Y(to) = Yo

yi(t) = yi(t; D=1,Yo) + Σ(∂yi(t)/∂Dn)(Dn’– 1)

“Normalized sensitivity of species i to reaction n”

Finally, in addition to “concentration sensitivities”, one can also compute “rate

“Instantaneous species i rate sensitivity to reaction n” = ∂Fi(Y,k)/∂kn

d(dyi(t)/dDn)/dt=d(dyi(t)/dt)/dDn=d(Fi(Y,D.*k)/dDn= Σ(∂Fi/∂ym)(dym/dDn) + ∂Fi/∂Dn

dSin/dt = Σ(∂Fi/∂ym)Smn + ∂Fi/∂Dn Sin(to)=0

dZil/dt = Σ(∂Fi/∂ym)Zml Zil(to)=δil

Note that there are a huge number of sensitivities O(NspeciesNreactionsNtimesteps).

C. Some Uses for Sensitivities

2) Putting Error Bars on the Model Predictions

4) Design of Kinetic Experiments

5) Valid Range Analyses

You might also like