Re: Fitting coupled differential equations to experimental data
- To: mathgroup at smc.vnet.net
- Subject: [mg83852] Re: Fitting coupled differential equations to experimental data
- From: Szabolcs Horvát <szhorvat at gmail.com>
- Date: Sun, 2 Dec 2007 04:14:03 -0500 (EST)
- References: <firedg$rq6$1@smc.vnet.net>
Ausman, Kevin wrote:
> I have been beating my head against a problem for several weeks now, so
> I'm at the point of begging for help. Please help!
I don't have your data, so I will use a simplified problem with this data:
data = Table[{x, Exp[2 x]}, {x, 0, 1, .1}]
>
> I have a set of coupled differential rate equations with several
> parameters (this is a chemical system, and these variable parameters are
> rate constants and initial concentrations):
>
> {DHR'[t] == -kadD*DHR[t]*C60[t] + kdeD*DHR1[t],
> DHR1'[t] == kadD*DHR[t]*C60[t] - kdeD*DHR1[t] - kox*DHR1[t],
> RH1'[t] == kox*DHR1[t] - kdeR*RH1[t] + kadR*RH[t]*C60[t],
> RH'[t] == kdeR*RH1[t] - kadR*RH[t]*C60[t],
> C60'[t] == -kadD*DHR[t]*C60[t] + kdeD*DHR1[t] + kdeR*RH1[t] -
> kadR*RH[t]*C60[t], Fluorescence[t] == ScaleFactor*(RH[t] + =
> RH1[t]),
> DHR[0] == initDHR, DHR1[0] == 0, RH[0] == initRH, RH1[0] =
> == 0,
> C60[0] == initC60}
>
> My variables in this case are:
>
> variables={DHR, DHR1, RH, RH1, C60, Fluorescence}
>
> And my parameters are:
>
> {initC60,initRH,initDHR,ScaleFactor,kadD,kdeD,kox,kadR,kdeR}
Let the diff eq be {f'[x] == a f[x], f[0] == 1}
>
> I can successfully create a model function at a given set of values for
> these parameters using NDSolve:
>
> model[kadD_?NumberQ, kdeD_?NumberQ, kox_?NumberQ, kadR_?NumberQ,
> kdeR_?NumberQ, initDHR_?NumberQ, initRH_?NumberQ, initC60_?NumberQ,
> ScaleFactor_?
> NumberQ] := (model[kadD, kdeD, kox, kadR, kdeR, initDHR, initRH,
> initC60, ScaleFactor] =
> First[NDSolve[{DHR'[t] == -kadD*DHR[t]*C60[t] + kdeD*DHR1[t],
> DHR1'[t] == kadD*DHR[t]*C60[t] - kdeD*DHR1[t] - kox*DHR1[t],
> RH1'[t] == kox*DHR1[t] - kdeR*RH1[t] + kadR*RH[t]*C60[t],
> RH'[t] == kdeR*RH1[t] - kadR*RH[t]*C60[t],
> C60'[t] == -kadD*DHR[t]*C60[t] + kdeD*DHR1[t] + kdeR*RH1[t] -
> kadR*RH[t]*C60[t],
> Fluorescence[t] == ScaleFactor*(RH[t] + RH1[t]),
> DHR[0] == initDHR, DHR1[0] == 0, RH[0] == initRH, =
> RH1[0] == 0,
> C60[0] == initC60}, variables, {t, 0, 7200}]])
Clear[model]
model[a_?NumericQ] :=
model[a] = First@NDSolve[{f'[x] == a f[x], f[0] == 1}, f, {x, 0, 1}]
It is better to use NumericQ instead of NumberQ, because
NumberQ[Sqrt[2]] === False but NumericQ[Sqrt[2]] === True. It is also
good to avoid symbol names starting with capital letters to prevent
collisions with built-in symbols. E.g. ScaleFactor is already used by
the VectorFieldPlots package.
>
> solution =
> model[0.018, 0.024, 0.394, 0.024, 0.018, 0.78, 0, 0.018, 90700];
> Plot[Fluorescence[t] /. solution, {t, 0, 7200}]
>
> <Graphic>
solution = model[1]
Plot[f[t] /. solution, {t, 0, 1}]
Yes, this works as expected.
>
> I can even treat the parameters as adjustible using the Manipulate
> operation, plotting it with the experimental data that I want to fit:
>
> XPData :=
> Import["C:\Documents and Settings\xxx\My \
> Documents\Briefcase\Research\Kinetic Fitting with \
> Mathematica\DHR123_7.PRN", "Table"]
> XPTime = XPData[[All, 1]];
> XPFluor = XPData[[All, 2]];
>
> Manipulate[(solution =
> model[kadD, kdeD, kox, kadR, kdeR, initDHR, initRH, initC60,
> ScaleFactor];
> Show[ListPlot[XPData],
> Plot[Fluorescence[t] /. solution, {t, 0, 7200}]]), {{initC60, 0.1},
> 0, 1}, {{initDHR, 0.1}, 0, 1}, {{initRH, 0}, 0,
> 1}, {{ScaleFactor, 10000}, 0, 1000000}, {{kadD, 0.1}, 0,
> 1}, {{kdeD, 0.1}, 0, 1}, {{kox, 0.1}, 0, 1}, {{kadR, 0.1}, 0,
> 1}, {{kdeR, 0.1}, 0, 1}]
This works also:
Manipulate[solution = model[a];
Show[ListPlot[data], Plot[f[t] /. solution, {t, 0, 1}]],
{{a, 1}, 0, 3}]
>
> What I cannot seem to do is fit this model to experimental data using
> FindFit, or by generating a least-squares error function and using
> Minimize, NMinimize, or FindMinimum.
>
> For example:
>
> FindFit[XPFluor, (Model2 =
> model[kadD, kdeD, kox, kadR, kdeR, initDHR, initRH, initC60,
> ScaleFactor];
> Fluorescence[t] /. Model2), {{initC60, 0.018}, {initDHR,
> 1}, {initRH, 0}, {ScaleFactor, 134000}, {kadD, 0.012}, {kdeD,
> 0.118}, {kox, 1}, {kadR, 0.012}, {kdeR, 0.134}}, t]
>
> generates ReplaceAll::reps errors and InterpolatingFunction::dmval
> errors.
The simplified version of this command is
FindFit[data, mdl = model[a]; f[t] /. mdl, {{a, 1.}}, t]
But there are several problems with this. First, not that FindFit does
not have any Hold* attributes, so 'mdl = model[a]; f[t] /. mdl' is
evaluated before passed to FindFit[]. The result of the evaluation is
f[t] /. model[a], and a ReplaceAll::reps error because model[a] is not a
list of replace rules.
You can see this if you evaluate 'mdl = model[a]; f[t] /. mdl separately.
The solution is to define
fitfun[a_?NumericQ, t_] := f[t] /. model[a]
and use
FindFit[data, fitfun[a, t], {{a, 1.}}, t]
>
> LSE[kadD_?NumberQ, kdeD_?NumberQ, kox_?NumberQ, kadR_?NumberQ,
> kdeR_?NumberQ, initDHR_?NumberQ, initRH_?NumberQ, initC60_?NumberQ,
> ScaleFactor_?NumberQ, XPT_?VectorQ,
> XPF_?VectorQ] := (LSE[kadD, kdeD, kox, kadR, kdeR, initDHR, initRH,
> initC60, ScaleFactor, XPT,
> XPF] = (solution =
> model[kadD, kdeD, kox, kadR, kdeR, initDHR, initRH, initC60,
> ScaleFactor]; M = Fluorescence[XPT] /. solution; Norm[XPF - M]))
>
> FindMinimum[
> LSE[kadD, kdeD, kox, kadR, kdeR, initDHR, initRH, initC60,
> ScaleFactor, XPTime,
> XPFluor], {{initC60, 0.018}, {initDHR, 1}, {initRH,
> 0}, {ScaleFactor, 134000}, {kadD, 0.012}, {kdeD, 0.118}, {kox,
> 1}, {kadR, 0.012}, {kdeR, 0.134}}]
>
> generates NDSolve::icfail, First::first, ReplaceAll::reps, and
> FindMinimum::nrnum errors.
I didn't take the time to figure out what went wrong here but
lse[a_?NumericQ, data_] :=
Norm[fitfun[a, data[[All, 1]]] - data[[All, 2]]]
FindMinimum[lse[a, data], {a, 1.}]
works correctly.
Just one more comment: try not to use global variables inside functions.
If you don't want to write rep /. fun[x] because it is too long, then
use With[{ff = fun[x]}, rep /. ff], but *not* ff = fun[x]; rep /. ff.
And don't forget to Clear[] on functions before redefining them if you
use memoization.
>
> I have tried dozens of other formulations, and can evaluate fits
> one-at-a-time or by using Manipulate, and then evaluate LSE (or other
> formulations) one-at-a-time or by using Manipulate, but I cannot seem to
> get Mathematica to perform an optimization.
>
> Does anyone have an idea of how to help here? Thanks!
--
Szabolcs