plotting PAP phi-rho-z

[Brian Joy]

Hi Sander,

I'm glad I could help.  No need to apologize!  I'm not that important.  I just got my hackles up a little because I'm extremely picky about units.  I carry them through all calculations and use them to check my work.

Brian

[sem-geologist]

Personally, I've basically given up on the full PAP in favor of XPP. It is simpler and generally gives better answers.

While XPP is by default on DTSA-II, I saw that NeXL has also X-PHI. I am quite satisfied with X-PHI on Cameca Peaksight (default matrix correction there). What I get impression of X-PHI is that it works pretty reasonable with under voltage. Nicholas, could you give some opinion on XPP vs X-PHI?

[Sander]

I'm glad I could help.  No need to apologize!  I'm not that important.  I just got my hackles up a little because I'm extremely picky about units.  I carry them through all calculations and use them to check my work.

I wish "certain other people" would be as picky in their publications... :-)

[Nicholas Ritchie]

Sander, what do you mean "as a function of WD?"

Brian & Sander:  With respect to the choice of XPP vs PAP vs XPHI, I personally almost always use XPP.  There may be algorithms that in certain circumstances that might work better.  However, it bothers me when people "shop for matrix correction algorithms" which work for a particular problem.  If you have a principled reason a priori to favor one algorithm over another, then great use that algorithm.  Selecting an algorithm because it gives the answers you expect seems suspect - not good scientific method.

That isn't to say that you can't do studies and based on those studies select one algorithm for a specific type of problem.  However, since my microanalysis is in multiple domains (geological, material science, forensics, nuclear etc.) and since I don't have time to study each algorithm with each class of problem in each domain, I've settled on XPP as a "good enough" algorithm for all domains.   (One of my hopes for the k-ratio database that Aurelien Moy is championing is that it would allow us to really evaluate the algorithms against a broad swath of data types.)  It has served me well but there are other viable alternatives including PAP, XPHI, CITZAF etc.  If I want to improve my accuracy, I endevour to find a better matched standard rather than abandon XPP.

[John Donovan]

If you have a principled reason a priori to favor one algorithm over another, then great use that algorithm.  Selecting an algorithm because it gives the answers you expect seems suspect - not good scientific method.

I completely agree with Nicholas on these points.   

It's also important to keep in mind that most of these matrix correction methods were tuned to one particular data set of another. PAP and XPP were tuned to the Pouchou k-ratio dataset, Bastin's matrix method was tuned to his k-ratio dataset, while John Armstrong tuned his Phi-Rho-z to the Shaw dataset. Others were tuned to others.

And that is why his CITZAF (and now CalcZAF/Probe for EPMA) offer all these matrix correction methods for comparison.  Paul Carpenter spent many years of work plotting these datsets against the 10 matrix methods in CalcZAF and also the 6 different MAC tables!  It's an exercise worth pursuing:

https://probesoftware.com/smf/index.php?topic=924.0

As an example (using just one MAC table) here is a GaSb synthetic (which should be 50:50 atomic concentrations) analyzed using GaAs and Sb metal as primary standards for all 10 matrix corrections in Probe for EPMA:

Summary of All Calculated (averaged) Matrix Corrections:
Un    2  GaSb as unk
LINEMU   Henke (LBL, 1985) < 10KeV / CITZMU > 10KeV

Elemental Weight Percents:
ELEM:       Sb      Ga      In      As   TOTAL
     1  62.757  38.023   -.014    .043 100.809   Armstrong/Love Scott (default)
     2  63.973  32.855   -.014    .038  96.852   Conventional Philibert/Duncumb-Reed
     3  63.144  33.355   -.014    .040  96.525   Heinrich/Duncumb-Reed
     4  62.761  35.664   -.014    .058  98.468   Love-Scott I
     5  62.904  36.426   -.014    .069  99.384   Love-Scott II
     6  64.539  37.036   -.014    .035 101.596   Packwood Phi(pz) (EPQ-91)
     7  63.281  37.885   -.014    .055 101.207   Bastin (original) Phi(pz)
     8  63.826  36.720   -.014    .063 100.594   Bastin PROZA Phi(pz) (EPQ-91)
     9  63.718  36.467   -.014    .039 100.211   Pouchou and Pichoir-Full (PAP)
    10  63.725  36.151   -.014    .043  99.905   Pouchou and Pichoir-Simplified (XPP)

AVER:   63.463  36.058   -.014    .048  99.555
SDEV:     .589   1.720    .000    .012   1.757
SERR:     .186    .544    .000    .004

MIN:    62.757  32.855   -.014    .035  96.525
MAX:    64.539  38.023   -.014    .069 101.596

Atomic Percents:
ELEM:       Sb      Ga      In      As   TOTAL
     1  48.570  51.387   -.012    .054 100.000   Armstrong/Love Scott (default)
     2  52.699  47.262   -.012    .051 100.000   Conventional Philibert/Duncumb-Reed
     3  51.996  47.963   -.012    .054 100.000   Heinrich/Duncumb-Reed
     4  50.161  49.776   -.012    .075 100.000   Love-Scott I
     5  49.683  50.241   -.012    .088 100.000   Love-Scott II
     6  49.931  50.037   -.012    .044 100.000   Packwood Phi(pz) (EPQ-91)
     7  48.861  51.082   -.012    .069 100.000   Bastin (original) Phi(pz)
     8  49.850  50.082   -.012    .080 100.000   Bastin PROZA Phi(pz) (EPQ-91)
     9  49.995  49.966   -.012    .050 100.000   Pouchou and Pichoir-Full (PAP)
    10  50.213  49.743   -.012    .055 100.000   Pouchou and Pichoir-Simplified (XPP)

AVER:   50.196  49.754   -.012    .062 100.000
SDEV:    1.266   1.261    .000    .015    .000
SERR:     .400    .399    .000    .005

MIN:    48.570  47.262   -.012    .044 100.000
MAX:    52.699  51.387   -.012    .088 100.000


It's interesting to see which methods do a better job...

The point being (as Nicholas mentions further on) that depending on the material, different matrix methods may or may not work very well. That's because none of them are really "universal" methods, as they've all been tuned to one dataset over another.  If one is analyzed a stoichiometric standard material, one can see that, but if the sample is an unknown, one must be careful not to "pick and choose" what one is hoping for.

The Si-Ir alloy is a classic example of this accuracy issue:

https://probesoftware.com/smf/index.php?topic=158.0

My personal hope is that a more physics based fundamental parameters methods will finally get us a universal matrix correction method.

[Probeman]

If anyone wants to play with the binary k-ratio datasets from Heinrich and Pouchou they are installed in the CALCZAFDATData folder in the UserData folder. Look for these files:

NISTBIN.DAT
NISTBIN2.DAT
NISTBIN3.DAT

Pouchou.dat
Pouchou2.dat

These are easily re-processed using the menus in the CalcZAF app using the Analytical | Calculate Binary Intensities... menus.  Some discussion on this is found here:

https://probesoftware.com/smf/index.php?topic=1256.0

Attached below are some other binary k-ratio data files from Bastin using the same CalcZAF binary k-ratio ASCII format.

[Brian Joy]

Brian & Sander:  With respect to the choice of XPP vs PAP vs XPHI, I personally almost always use XPP.  There may be algorithms that in certain circumstances that might work better.  However, it bothers me when people "shop for matrix correction algorithms" which work for a particular problem.  If you have a principled reason a priori to favor one algorithm over another, then great use that algorithm.  Selecting an algorithm because it gives the answers you expect seems suspect - not good scientific method.

Hi Nicholas,

I definitely don’t “shop” for a matrix correction method and then choose the one that gives the results I like.  I analyze a huge range of materials (mostly natural), and, with reasonable choices of standards, I find that PAP produces acceptable results for most oxides (including silicates), sulfides, sulfosalts, arsenides, selenides, tellurides, etc.  I like P&P’s internal consistency, their detailed treatment/improvement of the atomic number correction, and also their accurate phi(rho*z) curves, which should produce more accurate results for cases in which f(chi) is relatively small.  For the sake of consistency, I use PAP (with MAC30) exclusively.  Other models that appear acceptable to me are XPP, X-Phi, and PROZA96, but not Armstrong.  I was simply wondering why you seem to express a preference for XPP over PAP.  (After all, you did write that you had “given up” on PAP and that XPP “generally gives better answers.”)

It is also one of my “dreams” to collect new k-ratio data and expand that database greatly…  if I ever get the time.

Brian

[Sander]

Sander, what do you mean "as a function of WD?"

I have access to a system with a motorized Z sample holder and collected spectra of some pure elements at varying working distance.  This obviously changes the absorption path length towards the detector a little bit, and the ratio between the K and L lines is different in each spectrum.  My XPP implementation didn't account for the whole effect I was seeing, which is what set me off on this whole investigation.  I don't want to say too much yet because I can't even get my units right, so it's probably something stupid somewhere.

[Sander]

I definitely don’t “shop” for a matrix correction method and then choose the one that gives the results I like.

I think EPMA (especially EDS) is slowly becoming a "mainstream technology".  Especially on the current batch of more affordable systems, end-users expect a "point and shoot" functionality.  We should realize that the people on this forum are not your average analysts, and neither are the people using DTSA or CalcZAF.  For us, having a huge panel of buttons and algorithms to choose from is different: We'll select various algorithms and try to understand the results.  I can tell you from experience that not everybody works like this.  I have been in many customer escalations where someone would say "I am not getting the right results" and the "solution" was to select different ZAF models, different peak fitting strategies, different background fits, until the results were in line with their expectations.  Including people hitting crunchy particles and whatnot.  Most people on this forum would pull their hair out and tell the customer "No!! This is not how any of this works!!" but you can just hear them think "OK, nerd!" and continue with their day.

[Probeman]

Sander,
You have a point.

Reminds me of the old joke: "Half of all Americans are below average intelligence".  Yeah, yeah, it should be "median" but "average" sounds better!   

But it's worse than that because as you just alluded to: half of all scientists are below average in scientific ability.  And that might be an underestimation...

It's one reason NIST and others have been pushing for using actual standards in SEM-EDS, with a corresponding honest to goodness analytical total. Yes, it won't help in the case of "crunchy particles", then again, it certainly won't hurt!

And then there's Dale Newbury's studies about mis-identification of emission lines when using the auto-ID from some (all?) vendors...  so before we get to quantification, we're already not even wrong!

[Nicholas Ritchie]

Sander,
    You have to be really careful when interpreting data when you change the working distance.   You are not simply changing the take off angle.  I'm going to assume EDS and an ultra-thin window.  Maybe this is relevant, maybe it isn't.  Ultra thin windows are a polymer layer on a venician blind-like support.  The support grid is actually very thick.   100's of microns of Si usually and are very sensitive to orientation.  As long as you are on axis, the X-rays see the full 80% or so open area and 20% of the grid.   As soon as you start to go off the optimal axis, the incident X-rays start to strike the sides of the support.  The open area drops and a larger and larger fraction of the X-rays strike the support.   
   Of course, how this will effect your working distance data depends on the orientation of the window support grid (vertical vs horizontal vs .)
   I'm less clear about the construction of silicon nitride windows and if there is a similar issue with them.  Good old Be windows wouldn't have this problem.

Nicholas

[Sander]

Hi Nicholas,

That is a very interesting remark, which definitely goes on the list of "possible causes of what I'm seeing".  I am indeed using a silicon nitride window, but as far as I know it's mounted on a "honeycomb" support.  This support structure is in the order of 15 microns thick, according to the manufacturer.  Also, the distance of the detector to the point of beam impact is in the order of 4cm.  I take into account the fact that X-rays arriving from different working distances don't reach the detector window perpendicularly, and the absorption path through the window goes with cos(real_TOA - optimum_TOA), which in my case was less than 2% difference - but I didn't take the support structure into account.

"Good old" Be windows probably would eat up most of that FeLa anyway :-)

Thanks a lot for your thoughts.  I'll do some goniometry homework.

[Sander]

So I did the homework, and threw a detector cap under a microscope.  It indeed has a honeycomb structure and each "cell" is about 200 microns side-to-side, and the silicon support structures are about 20 microns wide.  If they are indeed only 15 microns thick as per the manufacturer's specification, then the effect of non-perpendicular incidence should not as big as you scared me for.

Some back-of-the-envelope calculations:  At the shortest WD, I have a TOA of 24 degrees, at the longest, 36 degrees.  The optimum (design) TOA is 28.5 in my system.  At the shortest WD, I would get a factor of 0.076 extra occlusion, at the longest, 0.137.  Calculating the average extra Si absorption would be hard because these numbers are "pure shadow" numbers and the X-rays would not pass to the full thickness of the Si.  Also, I'm seeing much more FeLa absorption at the shorter working distance as opposed to the longer one.

The extra "virtual thickness factor" of the window itself under those angles would be 1.003 and 1.009 respectively, so that can safely be ignored.