Measuring O (or C or N)

[Nicholas Ritchie]

I'm curious what are people's expectations when measuring O (not O-by-stoichiometry) using WDS.

What sort of accuracy / precision do you expect for a careful measurement with simple standards / or with matrix-matched standards?

[Probeman]

I'm curious what are people's expectations when measuring O (not O-by-stoichiometry) using WDS.

What sort of accuracy / precision do you expect for a careful measurement with simple standards / or with matrix-matched standards?

This might be a good topic for a poll...

That said, if people utilize all the tools available (in Probe for EPMA) for empirical mass absorption coefficients (MACs), Area Peak Factors (APFs), non-linear backgrounds and spectral interference corrections, and possibly time dependent intensity (TDI) corrections, the accuracy achievable with WDS for light elements is similar to other elements (~2% relative).

But this does take more effort than just your typical WDS analysis.  I discuss in detail what is required for accurate WDS light element analyses here in a YouTube video from a webinar last month:

https://www.youtube.com/watch?v=wrxPaZdK-Rg&ab_channel=ProbeSoftwareInc

It's a bit long but lots of details and examples. And I will also be giving a (much shorter) talk on light element WDS analysis at M&M next month.

[JonF]

That said, if people utilize all the tools available (in Probe for EPMA) for empirical mass absorption coefficients (MACs), Area Peak Factors (APFs), non-linear backgrounds and spectral interference corrections, and possibly time dependent intensity (TDI) corrections, the accuracy achievable with WDS for light elements is similar to other elements (~2% relative).

Looking back at some O measurements I made recently, ~2% relative error is pretty much what I got on ~1wt% O totals using PfE and the array of options listed above.

Also worth noting (as a general O measurement comment) that the attenuation coefficient of O by C is pretty big, so you need to make sure the standards and samples have comparable (or at the very least, known) carbon coat thicknesses.

 

[Nicholas Ritchie]

Looking back at some O measurements I made recently, ~2% relative error is pretty much what I got on ~1wt% O totals using PfE and the array of options listed above.

Jon, Do you really mean "95% of measurements in the range 0.98% to 1.02% for a nominal mass fraction of 1.00%?"

[Probeman]

Here's an example I did from earlier this year in getting ready for my light element talk this summer at M&M.

Here are some simple oxides (averages from two sets of measurements) in my standard mount using MgO (912) as a primary standard for O ka plotting measured concentrations (y axis) versus "published" oxygen concentrations (x axis) in these oxides (assuming single crystal stoichiometry).  First without using empirical MACs and APFs and plotted as a 1:1 plot in concentration units:



Note that the 1:1 line intersects MgO which is the primary standard. And again but this time plotting the concentration differences to the 1:1 line normalized to the highest concentration standard (SiO2):



Note that TiO2 (922) is the largest outlier as the absorption of oxygen Ka by Ti is very large and nominal table values are not accurate. And plotting once again, but this time using empirical determined MACs and APFs from Bastin and Pouchou:



The empirical MACs helped a lot particularly for TiO2. The only big outlier is ZnO (930) and maybe it's not exactly stoichiometric? The others are all within +/- 2% relative or so.  The oxide standard numbers are:

860 Set   1 Al2O3 + 0.40 Cr single crystal (#260)
870  Hematite Sennin (#370)               
895  Magnetite (std #395)                 
912  MgO (elemental) (#12)                 
914  SiO2 (elemental) (#14)               
922  TiO2 (elemental) (#22)               
930  ZnO (elemental) (#30)                 

[JonF]

Looking back at some O measurements I made recently, ~2% relative error is pretty much what I got on ~1wt% O totals using PfE and the array of options listed above.

Jon, Do you really mean "95% of measurements in the range 0.98% to 1.02% for a nominal mass fraction of 1.00%?"

Not quite, I was referring to the 1 sigma analytical error in relative percent, per analytical point i.e. the output PfE gives you in the output excel spreadsheet when you select the "Analytical Errors in Relative Percents" radio button in the User Specified Custom Output page in Analyze!.

[Probeman]

Yes, we need to distinguish between accuracy and precision in responding to Nicholas' question...

My post above refers to the question of accuracy for light element (oxygen) analyses.  Jon's post refers to the analytical sensitivity question, which the calculation of in Probe for EPMA is based on the Scott-Love expression:

https://probesoftware.com/smf/index.php?topic=1307.msg12149#msg12149

Because LDE diffractors yield high intensities, though with lower P/B ratios, it makes sense that light element precisions would be somewhat similar to other element sensitivities. 

The point I was making about light element accuracy is that unless one is utilizing the empirical MAC/APF/TDI and interference corrections in Probe for EPMA, one will not attain similar levels of accuracy with light elements as with other (non-light) elements.

Of course one can get bad results with other (non-light) elements as well, depending on the situation, e.g., curved backgrounds, spectral interferences, volatile (TDI) issues, etc...

Basically I agree with Grok:

https://probesoftware.com/smf/index.php?topic=233.msg12654#msg12654

 

[Nicholas Ritchie]

I think you are right.  We do seem to mean something different by "relative error"

I'd differentiate between:
    Precision - If a measurement is replicated N times what is the dispersion of the results.   sqrt( (1/(N-1)) sum( (x_i - mean(x_i))^2))  (Sample standard deviation)
    Accuracy - Relative Deviation from Expected Value (X)       RDEV = (mean(x_i) - X)/X

Maybe I spend too much time measuring materials that I know the true composition, but it is the RDEV that I'm interested in. 

What is the expected RDEV for O (or C or N)?

[John Donovan]

I think you are right.  We do seem to mean something different by "relative error"

I'd differentiate between:
    Precision - If a measurement is replicated N times what is the dispersion of the results.   sqrt( (1/(N-1)) sum( (x_i - mean(x_i))^2))  (Sample standard deviation)
    Accuracy - Relative Deviation from Expected Value (X)       RDEV = (mean(x_i) - X)/X

Maybe I spend too much time measuring materials that I know the true composition, but it is the RDEV that I'm interested in. 

What is the expected RDEV for O (or C or N)?

My post:

https://probesoftware.com/smf/index.php?topic=1643.msg12689#msg12689

is close to what you want.  You'll just need to subtract and divide the axis values and assume the x-axis values are the "expected" values.

[Nicholas Ritchie]

Thanks.  This is very interesting data.