GAK

November 20, 2006

XRF Measurement Confidence Issues in Testing Alloys

Or Any Product to Engineering Specifications
When testing any physical or chemical parameter of any substance, material, alloy or other,
there are inherent measurement errors, regardless of the measurement technique used. These
errors are the result of many key factors:
Type of Measurement Instrument Used and its Measurement Stability over Time
Measurement Sensitivity of Instrument Used (smallest measurement unit possible)
Measurement Conditions (temperature, instrument calibration, operator experience)
Accuracy of Calibration Standards Used (including measurement errors)
Counting Statistics (the longer the measurement time, the better the precision)
Fundamentally, testing or measuring is a statistical process, and having a basic understanding
of statistics is important when making measurements and determining their reliability. For
example, when measuring an alloy for chemical composition, say 316L SS, if you measure it
under the same conditions with an x-ray spectrometer for a few or one hundred or more times,
you will get different results for every measurement. If you were to plot the results for any
element measured, taking Cr for example, on a graph, you would see a distribution of values
known as a standard bell or Gaussian curve. The Apex of this curve is the average or Mean
value of all the measurements made. The further you move to the left or right of the Mean
value, the frequency of measurements declines, with most of the values clustering around the
Mean. Statistically, 68%, 95%, and 99.7% of all measurements made will lie correspondingly
within 1, 2, or 3 standard deviations (Sigma) from the Mean. See the illustrative Graph and
table on the next page.
Say the measured Niton value for Cr is 14.5% +/- 0.4 (which is the Niton 2 Sigma measurement error). This means that 95% of all measurement values would lie within 14.1% and
14.9%. If you wished to expand the Cr measurement confidence interval to 3 Sigma (multiply
the 2 Sigma +/-0.4 error by 1.5), so 99.7% of all measurement values would lie between 13.9%
and 15.1%. Hence, you could be highly confident that the true value for Cr was between
13.9% and 15.1%. If you wanted to reduce the XRF measurement error by a factor of 2x, you
would need to increase the measurement time by a factor of 4x. For example, by counting for
60 seconds instead of 15 seconds, the 2 Sigma error would drop to +/-0.2.
Say the Engineering Specification for 316L SS for Cr is 16% min to 18% max, and we have
measured a Mean value of 16.5% with a 2 Sigma error of +/-0.4. Does this mean the analyzed
sample is in or out of Spec? Given the instrument measurement value of 16.5% with a 2 Sigma
error of +/- 0.4, you could be 95% confident that the Cr value lay between 16.1% and 16.9%,
which is within the 16% to 18% values specified, and the sample would be in Spec. Also, any
alloy engineering specification should have a stated range of acceptable values, for alloy
composition varies from heat to heat and by manufacturer, and every standard used for
reference purposes will have +/- measurement errors as well, whether shown or not.
The critical issue when testing alloys by XRF is to be sure the alloys of interest can be safely
discerned from one another. For example, from common SS alloys being tested, the typical
problem areas are between 304 and 304L and 316 and 316L. The reason these alloys cannot

GAK

November 20, 2006

be separated from one another by XRF is due to the very small difference in Carbon valves in
304L and 316L and in 304 and 316, and Carbon cannot be measured by the Thermo/Niton XRF
system. The other alloys of interest vary in composition enough to avoid problems of
misidentification. If there is a measurement confidence issue, the instrument tested chemistry
values should be directly compared with a known standard sample measured under the same
conditions. Also, by increasing the measurement time by a factor of 4x, the measurement
precision can be increased (error reduced) by a factor of 2x, if necessary.

Standard Gaussian Curve

2
X

Typical Measurements with Thermo/Niton XLt for 316L

(15 second Measurement Time)

Run #
1
2
3
4
5
6

Cr
17.1%
*16.2%
16.8%
16.4%
17.4%
*17.7%

2 Sigma
+/-0.4%

3 Sigma
+/-0.6%

Standard
16% to 18%

In or Out of Spec
In
Out?
In
In?
In
Out?

GAK

November 20, 2006

Typical Niton Analyzer Screen Results for 316 SS Test

# 11 All Alloy
Time 1st Run 21.3 sec
316SS

Ele
Cr
Mo
Ni
Fe
Mn

0.3

>

2nd Run 64.7 sec

3rd Run 88.2 sec

Chi Square Figure

Note: If Chi Square Figure is 2.0 or less, the identified alloy fit
is considered Good in relation to the stored 316 Standard in the
Niton Memory; from 2.0+ to 3.0 the fit is Probable; from 3.0+
to 4.0 the fit is Possible; and if 4.0+ there is No Fit or Match.

+/-

*18.44
17.66
17.83
2.68
2.72
2.68
12.86
12.93
12.46
63.41
63.90
63.80
1.95
1.71
2.08

0.34
0.18
0.16
0.14
0.08
0.07
0.63
0.34
0.30
0.79
0.43
0.19
0.39
0.21
0.19

316 Standard
Ref. Values
16 18%
2 3%
10 14%
Bal
2.0%

Si

Not Measured

1.0%

Not Measured

0.03%

Not Measured

0.045%

While the Chi Square Fit above (0.3) for this sample is Good for 316SS, this sample
tests Inconclusive in the 1st Run because the Cr measurement value *18.44 +/-0.34
exceeds the 18% max value for the reference standard, factoring in the 2 Sigma
counting error. Longer analysis times will reduce the measurement error and improve
the testers confidence in the measurement, as shown above in Runs 2 and 3.