Polydispersity parameters: Users don't know what's being reported.

[toqduj]

I've noticed that 44% of the SasView submissions to the Data Analysis Round Robin say that SasView reports volume-weighted size distribution parameters (means and widths in particular). 51% note that these are number-weighted, and some hedge their bets and don't say anything at all.

This means it's unclear as to what the values mean that come out of SasView. That may be a problem with the user, but it could be resolved by being more explicit in the user interface. @timsnow noticed in the documentation that it should be number-weighted but might be volume-weighted for Schultz distributions?

Anyway, PEBCAK, but given that even the more "advanced users" that, I suppose, contributed to the Round Robin can't figure this out, perhaps it's worth improving in the UI. These things shouldn't be hidden in the docs.

If I understand correctly, @pkienzle uses volume-weighting internally for the integrations (as that makes more sense in terms of signal weighting anyway). Perhaps having a slider to switch between "number-weighting" and "volume-/mass-weighting" in the Polydispersity panel would solve both the awareness and the flexibility issues?

[smk78]

Interesting!

We currently have this on the FAQ (because we've been asked before!!!):
https://www.sasview.org/faq/#is-the-schulz-zimm-polydispersity-function-used-in-sasview-a-number-average-or-a-weight-average-distribution
https://www.sasview.org/faq/#what-about-the-other-polydispersity-functions-used-in-sasview

And this in the docs (USER DOCUMENTATION > Fitting & Other Analyses > Model Fitting > Polydispersity Distributions):
https://www.sasview.org/docs/user/qtgui/Perspectives/Fitting/pd/polydispersity.html

Both explicitly say our distributions are implemented as number distributions.

Is there something else you think we could do to further clarify our implementation?

Would you be willing to drop a line to a random selection of those who reported them as volume weighted and ask why they thought that?

[RichardHeenan]

But the "scale" parameters in sasview are a "volume fraction" if the sld's are correct, as we normalise the number distribution polydispersity to unit volume. We do this as the volume fraction is what an experimenter is most likely to know, as well as being helpful when thinking about S(Q). Ought to modify the polydispersity.html page linked above.

[pkienzle]

Distributions use number density even though integrations use volume weighting. Documentation here. Extended discussion with derivations here.

It would be quite a bit of work to use volume fraction distributions. You would need to specify it to the UI and send it through the fitting perspective into sasmodels. Within sasmodels you would need to send this flag to the kernel calculators so they remove the volume weighting for the individual integration points.

[butlerpd]

I suspect that most of those reporting as a volume weighted (or z average) distribution are in fact the more "advanced users"? I've fount that most material scientists do not know about anything but number averages and therefore intrinsically assume number average without realizing it. This is exacerbated by the rise of microscopy techniques which intrinsically report number averages and lead to very confused users when the number is different from the volume weighted average obtained by scattering.

Over the years I've also noticed that the concept of higher moments (weight average, z average etc) tends to be confusing to many who should be at least aware of them (polymer scientist?). Note that if polydispersity is not provided, or form_volume and Iq are not properly supplied, the result will be a "z" or volume weighted average, something that comes up in writing custom models.

[toqduj]

@butlerpd I should maybe clarify that, while the reports said "volume-weighted", the actual values reported still screamed "number-weighted" to me. So it's not that people did a conversion, they just don't know what the numbers mean, advanced or not.

[butlerpd]

Thanks @toqduj. I would indeed be shocked if anybody tried to convert the fitting results they receive back into volume (or some other) averages 😄.

It could be just semantics here, but in my experience it is not that people don't know what the numbers mean so much as they don't understand the concept of anything beyond a number average - very often, even if they've heard the terms weight average, z average and/or volume average.

[smk78]

Like I said, contact some of those that reported volume-weighted and ask them why they did so! You might actually be doing them a service... 😄

[lucas-wilkins]

This might be as simple as putting the right labels on the plots. See #2472

[toqduj]

That would be a good start, yes. Maybe also indicate the volume weighting by adding a subscript _n or _v to the distribution parameters in the fitting panel.