nonsensical negative values / restrictions
Dear all, I have the following problem: In some fits using Artemis (Demeter 0.9.25) I will get negative values for parameters where it makes no sense at all. For example, when trying to fit two first shell paths, Pt-Pt and Pt-O, for a Pt nanoparticle, sigma^2 for the Pt-O path becomes negative. Obviously, sigma^2 has to be a positive number. My questions: 1. Why does the fit algorithm do this? 2. What can I do to avoid this? I know that there are most likely significant problems with my fit if this behaviour occurs. As far as I know it's possible to impose restrictions in Artemis and if anyone could point me towards a tutorial explaining how to implement restrictions that might be a first step in the right direction. Currently, the chapter "Constraints and restraints" in the Artemis documentation is "Coming soon......". With kind regards, Felix
Felix, When you fit your EXAFS data using Artemis + Ifeffit or Larch, you are using a Levenberg-Marquardt non-llinear, steepest descent algorithm. The fit does kn ow anything /at all/ about what you are trying to do. It doesn't know what a sensible value for sigma^2 or any other parameter is. All it knows is that you have data and a model that you are throwing at the data. Armed with data and model, the algorithm optimizes the parameters that you have flagged as the variables. It changes those values until it makes the sum of the paths look as much like the data as possible. It's goal is to make the red line look like the blue line. That's it. That's all it does. That's how sigma^2 can come out as a negative number. A negative sigma^2 or any other nonsensical parameter doesn't represent a failure of Artemis. It is information that is useful to you, the person doing the fit. A nonsense parameter is trying to tell you about a way that the model you have proposed does not represent the data that you have measured. And it is a hint about what changes you need to explore in your model to more properly represent the data. An example: suppose you have an unknown oxide phase and you have reason to believe it is some sort of highly disordered phase. Just to pull an example out of the air, you might think that your sample is filled with FeO(OH) when in fact it is filled with FeO. FeO(OH) is pretty disordered, whereas FeO is rocksalt. If you run Feff on Feo(OH), you will get a spread of paths at several distances for the first shell. When you make the fit with the FeO(OH) model, the structural contribution to the disorder is MUCH higher than in the actual sample. The fit is likely to use a negative sigma^2 to counteract that effect. That's a simple and contrived "explaining example", but it makes the point (I hope). Artemis didn't fail. The fit is trying to tell you that you assumptions were not realized in the data. In short, I agree that your results make no sense, but I suggest that the problem is that you are making a poor assumption in your fitting model. It also might be the case that there is a problem with the measurement of your data that introduces a systematic error that the fit compensates for by applying a weird parameter value. Or maybe you have a sulfide instead of the oxide you were expecting. Or maybe you ran Feff for the Pt K edge rather than the Pt L3 edge. (Oops! I actually do that with embarrassing frequency....) Or something else I haven't thought of.... HTH, B On 06/16/2017 02:09 AM, Felix E. Feiten wrote:
Dear all,
I have the following problem:
In some fits using Artemis (Demeter 0.9.25) I will get negative values for parameters where it makes no sense at all. For example, when trying to fit two first shell paths, Pt-Pt and Pt-O, for a Pt nanoparticle, sigma^2 for the Pt-O path becomes negative.
Obviously, sigma^2 has to be a positive number.
My questions:
1. Why does the fit algorithm do this?
2. What can I do to avoid this?
I know that there are most likely significant problems with my fit if this behaviour occurs. As far as I know it's possible to impose restrictions in Artemis and if anyone could point me towards a tutorial explaining how to implement restrictions that might be a first step in the right direction. Currently, the chapter "Constraints and restraints" in the Artemis documentation is "Coming soon......".
With kind regards,
Felix
_______________________________________________ Ifeffit mailing list Ifeffit@millenia.cars.aps.anl.gov http://millenia.cars.aps.anl.gov/mailman/listinfo/ifeffit Unsubscribe: http://millenia.cars.aps.anl.gov/mailman/options/ifeffit
-- Bruce Ravel ------------------------------------ bravel@bnl.gov National Institute of Standards and Technology Synchrotron Science Group at NSLS-II Building 743, Room 114 Upton NY, 11973 Homepage: http://bruceravel.github.io/home/ Software: https://github.com/bruceravel Demeter: http://bruceravel.github.io/demeter/
Dear Bruce, thank you very much for your quick and thorough answer. I used to do I/V-LEED to do structural characterization of surfaces and there it was pretty common to get stuck in local minima. So my fear was that this negative R-factor corresponds to a local minimum and that there is actually a better minimum with reasonable values. I have noticed that there is some dependence of the fitting results on the start values used. Also, in one of the very useful videos of you giving talks about EXAFS you mentioned that the Levenberg-Marquardt algorithm can end up in a local minimum. However, I'm starting to get the feeling that this is much less of a problem in EXAFS analysis than in my previous field. I guess one of the reasons might be that a fit in IFEFFIT/Artemis takes seconds whereas a single datapoint (i.e. a fully dynamical scattering calculation for one trial structure) in I/V-LEED would take ~30 mins. After having done some systematic variation of parameters with a little script I wrote (more on that in the respective thread) I'm starting to realize that the hyperspace in EXAFS fitting seems rather shallow (at least for my samples). I don't think that the reason for the physically unreasonable results is a systematic error, as I have a number of different samples measured in one beamtime and some of them are fine while others have varying issues (mostly with too high/low sigma^2 and it's correlation with N). With kind regards Felix On 17/06/2017 00:10, Bruce Ravel wrote:
Felix,
When you fit your EXAFS data using Artemis + Ifeffit or Larch, you are using a Levenberg-Marquardt non-llinear, steepest descent algorithm. The fit does kn ow anything /at all/ about what you are trying to do. It doesn't know what a sensible value for sigma^2 or any other parameter is. All it knows is that you have data and a model that you are throwing at the data.
Armed with data and model, the algorithm optimizes the parameters that you have flagged as the variables. It changes those values until it makes the sum of the paths look as much like the data as possible. It's goal is to make the red line look like the blue line. That's it. That's all it does. That's how sigma^2 can come out as a negative number.
A negative sigma^2 or any other nonsensical parameter doesn't represent a failure of Artemis. It is information that is useful to you, the person doing the fit.
A nonsense parameter is trying to tell you about a way that the model you have proposed does not represent the data that you have measured. And it is a hint about what changes you need to explore in your model to more properly represent the data.
An example: suppose you have an unknown oxide phase and you have reason to believe it is some sort of highly disordered phase. Just to pull an example out of the air, you might think that your sample is filled with FeO(OH) when in fact it is filled with FeO. FeO(OH) is pretty disordered, whereas FeO is rocksalt. If you run Feff on Feo(OH), you will get a spread of paths at several distances for the first shell. When you make the fit with the FeO(OH) model, the structural contribution to the disorder is MUCH higher than in the actual sample. The fit is likely to use a negative sigma^2 to counteract that effect.
That's a simple and contrived "explaining example", but it makes the point (I hope). Artemis didn't fail. The fit is trying to tell you that you assumptions were not realized in the data.
In short, I agree that your results make no sense, but I suggest that the problem is that you are making a poor assumption in your fitting model.
It also might be the case that there is a problem with the measurement of your data that introduces a systematic error that the fit compensates for by applying a weird parameter value.
Or maybe you have a sulfide instead of the oxide you were expecting.
Or maybe you ran Feff for the Pt K edge rather than the Pt L3 edge. (Oops! I actually do that with embarrassing frequency....)
Or something else I haven't thought of....
HTH, B
On 06/16/2017 02:09 AM, Felix E. Feiten wrote:
Dear all,
I have the following problem:
In some fits using Artemis (Demeter 0.9.25) I will get negative values for parameters where it makes no sense at all. For example, when trying to fit two first shell paths, Pt-Pt and Pt-O, for a Pt nanoparticle, sigma^2 for the Pt-O path becomes negative.
Obviously, sigma^2 has to be a positive number.
My questions:
1. Why does the fit algorithm do this?
2. What can I do to avoid this?
I know that there are most likely significant problems with my fit if this behaviour occurs. As far as I know it's possible to impose restrictions in Artemis and if anyone could point me towards a tutorial explaining how to implement restrictions that might be a first step in the right direction. Currently, the chapter "Constraints and restraints" in the Artemis documentation is "Coming soon......".
With kind regards,
Felix
_______________________________________________ Ifeffit mailing list Ifeffit@millenia.cars.aps.anl.gov http://millenia.cars.aps.anl.gov/mailman/listinfo/ifeffit Unsubscribe: http://millenia.cars.aps.anl.gov/mailman/options/ifeffit
participants (2)
-
Bruce Ravel -
Felix E. Feiten