Best Practices Guide for Developing LIBS Calibrations

Overview: This document outlines how to generate good calibration curves using Profile Builder and
SciAps LIBZ analyzers. Most of the information has been developed from alloy applications but should be
applicable to other sample types.

1) Choose a group of samples and create a calibration model

a. Group by matrix type/major element(s)
2) Decide which elements to set curves for
3) Determine lines to use in the intensity ratio denominator
a. Pick element(s) that are constant across all samples
b. Lines should be strong and prominent in all samples
4) Determine lines to use in the intensity ratio numerator
a. In same spectrometer as, and near, denominator lines
b. Strong, but not saturating across calibration range
c. Little overlap or interference from other elements
d. Choose lines that produce good calibration curves
i. Linear, low RMSE, small/no offset preferred
e. Use multiple lines if signal is weak or curve improves
5) Determine regions and preprocessing
a. Region should fit whole peak, plus ~0.05-0.1nm wiggle room
b. 0th derivative/SGOLAY Baseline/≥7pts smoothing
c. 1st derivative/SGOLAY abs value/≥5pts smoothing
d. More smoothing at higher WLs
e. Consider changing region when using preprocessing
f. Try multiple preprocessing parameters, pick set that gives best curve
6) Determine fit parameters
a. Linear – narrow calibration range, poor precision, too few data points
b. Quadratic – wide calibration range, many data points, non-constant denominator
c. Force Zero – too few blanks in calibration, too much scatter on blanks
d. Allow offset – Many blanks included clustered at some nonzero value, background peak,
1st derivative used
7) Remove any outliers from the calibration curve
a. Causes
i. Bad tests
ii. Sample contamination/bad assay
iii. Overlapping elements
iv. Matrix variation or effects
v. Analyzer drift
b. Solutions
i. Reshoot
ii. Remove from curve or calibration model
Step 1) Choose a group of samples and create a calibration model

A calibration model, e.g. Al alloy

Is a grouping of samples of similar matrix type. A matrix type is characterized by similar bulk element
composition (i.e. similar content of major elements), similar density, similar target analytes, and similar
preparation (e.g. solid, pressed pellet, etc.). Curves within a matrix type generally use the same
denominator in the IR and have good correlation (low RMSE) on each element’s calibration curve.

Profile Builder allows for two variations on the calibration model:

1) Sub-base: For most purposes, a sub-base is its own independent calibration model. Sub-bases
are grouped together during base ID in Alloy Mode, and are sorted during fingerprinting within
each base. After sub-base is determined, it acts just like an independent calibration model.
2) Sub-type: A sub-type is a way to divide a calibration model (or base) so that some elements put
all samples within the group on the same curve, and others use different curves for the different
sub-types. For example, within StainlessStd, samples are broken into 300, 400, and Duplex sub-
types. Most elements do not use sub-types, meaning that all StainlessStd samples are plotted on
a single curve for each element. Cr uses sub-types, so if a sample matches to a sub-type (e.g.
300) during fingerprinting, Cr will be calculated using the Cr sub-type curve, not the main
StainlessStd curve. The format for a sub-type model name is StainlessStd-300, where
StainlessStd is the base/sub-base, and 300 is the sub-type.

Step 2) Decide which elements to set curves for

Choose elements using the element selector in Profile Builder

Your highest concentration sample should be well above the LOD for that element in that sample type. If
the LOD for your sample type hasn’t been determined yet, it may be necessary to attempt to create a
calibration curve to validate the application.
Step 3) Determine lines to use in the intensity ratio denominator

In the intensity ratio (IR), the denominator serves as a normalization factor for the numerator (analyte)
signal. The normalization corrects for variations in signal caused by variations in mass ablated, laser
focus, laser pulse energy, sample density, and other plasma effects.

For many sample types, a single matrix element can be used in the denominator. For alloys, this is the
alloy base material – e.g. Al for aluminum alloys, Fe for stainless steels, etc. For geochemical samples, it
is the element or elements that comprise the bulk of the sample. The idea is to pick one or more
elements that have high and constant concentration for all samples. There may only be one line in the
denominator, e.g. Al alloy

Or there may be multiple lines, e.g. Ni alloy

Where Ni may not be constant, but the sum of Ni + Cr + Fe is constant.

Common elements used as denominators include:

Sample Type Denominator Element(s)
Aluminum Alloy Al
Copper Alloy/Brass Cu
Low Alloy Steel Fe
Stainless Steel Fe
Ni
Ni + Cr + Fe
Ni + Cr + Fe + Mo
Nickel Alloy Ni + Cr + Co
Titanium Alloy Ti
Si
Soil Si + O
Si + Al + Ca + Fe
Si + O
Igneous Rock Si + Fe + Ca + Fe + O
Fe
Iron Ore Fe + O
Choosing denominator elements can be especially difficult for geochemical samples, since often no one
element dominates the matrix. In some cases, different element curves may need different
denominators, especially if the concentration range is different. For example, a trace element analyte
may only need one element in the denominator, while a matrix element >> 10% may need multiple
elements in the denominator.

After choosing which elements should be in the denominator, it is necessary to choose which lines to
use. It is best to choose lines that are closest to the analyte lines in the numerator, which will be chosen
in step 4, so step 3 and 4 may be a somewhat iterative process. The criteria are, in order of importance:

1) Line should be strong enough that repeatability is good

2) Line should be within the same spectrometer as the numerator line(s)
3) Line should be as close as possible to the numerator line(s)

For example, don’t choose a denominator line 10nm from the numerator line if it is very weak; choose a
stronger line further away. If the numerator line is near a spectrometer boundary, it is better to choose
a denominator line within the same spectrometer than one that is fewer nm away but in a different
spectrometer.

Some matrix elements have many lines to choose from. Fe in a steel has dozens of lines across all
spectrometers, while elements like Si only have a few (Si251 and 288).

It is common to use the same denominator for all curves within a base, though this is not universal. For
example, in a binary Ni/Mo alloy, where Ni ranges 60-80%, most elements may use Ni only as the
denominator, while Ni and Mo may use Ni+Mo. It is also possible that preprocessing and regions may be
different for the same line used in the numerator and denominator (see section 5).

Advanced: It is also possible to normalize to spectral features other than matrix element emission lines.
These can include:
1) Normalization to the integrated intensity of the whole spectrum, or the spectrometer
2) Normalization to some atmospheric line, e.g. Ar (with purge) or N (in air, a N feature at
500.51nm can be useful)
3) No normalization (normalization to 1.0). This is highly sensitive to intensity drift and sample
preparation and presentation.

Below is a table of common denominator lines to use on the Z-500.

A (180-255nm) B (255-315nm) C (315-420nm) D (420-900nm)

Al 236.71, 237.31 308.22, 309.27 394.4 466.31*
Si 251.61 288.16 - -
Ca - 317.93* 396.85 422.67*
Ti 226.12 308.8 336.12, 368.52 498.17
Fe 238.2 259.94 373.49 438.35, 534.1
Ni 221.65 300.25* 341.48 508.05
Cu 224.43 296.12* 327.4 510.55
* Line is weak and may not be visible in all spectra
And on the Z-300
K (185-365nm) L (365-620nm) P (620-950nm)
236.71, 237.31, None
Al 308.22, 309.27 394.4 Determined Yet
Si 251.61, 288.16
Ca 317.93*
Ti 226.12, 308.8
Fe 238.2 259.94
Ni 221.65 300.25*
Cu 224.43 296.12*

Step 4) Determine lines to use in the intensity ratio numerator

The analyte element, which is the element of interest that will be plotted on the calibration curve, will
have one or more lines in the numerator. In general, if a line is strong, only one is necessary. If the
emission lines available are weak, or the concentration is low, more than one may be needed. Criteria to
consider for choosing analyte lines include:

- Strong emission lines with high intensity

- Emission lines that do not saturate at high concentration (On LIBZ analyzers, signal < 25k
intensity)
- Low overlap of neighboring peaks
- Good internal standard (denominator) line choices nearby
- If more than one line is used, keep within one spectrometer if possible

Potential lines to use can be found in the “Common LIBS Wavelengths” table, as well as the standard
Alloy and GeoChem calibrations. In Profile Builder, a set of SciAps validated LIBZ lines as well as NIST
lines are available to choose from

Note that the NIST database of lines is large and may include lines not visible in all spectra. SciAps LIBZ
lines have been validated for use, but there may be usable lines not on the LIBZ list.

If choosing between multiple numerator lines, it may be useful to create calibration plots of each
individually to determine which one gives the best correlation. If multiple lines reduce RMSE significantly
and produce a better curve, then it is better to use multiple lines than a single line. When considering 3
or more lines in the numerator, Profile Builder allows users to turn off ( → Disable) each region to
see the impact of removing one line on the overall calibration curve.
Step 5) Determine regions and preprocessing

The region for each line should be wide enough to fit the highest concentration sample (i.e. widest
peak). If possible, it is best to draw the region wide enough to accommodate up to 0.05 – 0.1nm of
wavelength drift, which is the maximum typically seen between WL calibrations.

A common region width for peaks < 400nm is 0.4nm, so 0.2nm surrounding the peak maximum. Below is
an example of a typical region selection for Ti336 in Al alloys

At high WLs, particularly >> 400nm, regions may need to be much wider, up to 1nm or more, as seen
below for Zn481 in Al alloy

Regions should not normally be narrower than 0.25nm, but in some cases where there is a peak overlap
or unusual background features, narrow regions are unavoidable. See, for example, Re227 in Ni alloy,
which is very weak and on top of a background of Ni peaks:
Wide regions allow for some WL shifting, but may inadvertently include neighboring peaks or
background features. Narrow regions avoid neighboring features and can give optimal background
subtraction when using 0th derivative, but are very sensitive to small changes in WL.

Since Profile Builder recalculates the calibration curve in real time, users can shift the region boundaries
left and right to see the impact of small WL shifts on the calibration. An ideal region will allow shifting of
±0.05nm with only small change to the calibration curve and RMSE.

Profile Builder allows users to click and drag the region boundary (green, above) or use the up and down
arrows on the region to adjust the boundary. For simplicity, standard SciAps region boundaries are
rounded to the nearest 0.05nm. There is not much benefit in setting regions with any greater precision,
since peak resolution, detector pixel spacing, and small WL shifts between each WL calibration make
finer divisions less meaningful.

Next, choose preprocessing for the region. Profile Builder allows users to apply Savitzky-Golay
smoothing and derivative, as well as several algorithms for integration. The control panel for each region
looks like this:

The algorithm options are:

Peak Area Minus Baseline – the simplest background subtraction. The software draws a linear baseline
between the two region boundaries and subtracts off the background based on this approximation. The
peak above that baseline is then integrated. See example Ti336:
This option should rarely be used. Because the baseline subtraction is applied to each shot (e.g. each
laser pulse, or a small number of averaged laser pulses) the StDev calculated from using this method can
be high.

SGOLAY minus baseline – Applies Savitzky-Golay smoothing and then performs baseline subtraction as
above. Normally used with 0th order derivative.

SGOLAY with Absolute Values – Applies Savitzky-Golay smoothing and then integrates region with no
baseline subtraction. Normally used with 1st order derivative or higher.

Local Spectrometer Normalization (LSN) – Integrates the signal within a spectrometer, with the given
WL boundaries. This would only be used in the denominator.

Value 1.0 – Uses a constant value of 1.0. Use this in the denominator to create an unnormalized
calibration curve.

For most lines, one of the two SGOLAY algorithms will be used, depending on whether 0th or 1st
derivative is applied. Choosing LSN or Value 1.0 for the denominator lines is an advanced option (see
section 3). In rare cases, SGOLAY with Absolute Values may be used with 0th derivative, and vice versa,
but this is not recommended.

For each region, choose whether 0th order (no) derivative or 1st order derivative is more appropriate. 0th
order derivative is more straightforward to use, and can be useful when concentrations (or signal) are
low or when different samples within the calibration have varying background shapes. 1st derivative can
be helpful when a background is flat or constant and noise is low. 1st Derivative is less sensitive to WL
shifting, since no background subtraction is needed. See the example of Ti336 below. If the peak were to
shift 0.1nm to the red (higher WL), the background subtraction would be skewed:

Where the 1st derivative integrates the absolute value, a similar shift results in less loss of signal.

In practice, users may wish to try both 0th and 1st derivatives on each region to see which gives a better
curve. If both curves are comparable, 1st derivative is usually preferred, but this is a judgement call left
to the user.

If applying Savitzky-Golay smoothing, the user will have to determine the number of smoothing points,
or WinSize. Larger WinSize means more smoothing. Note that the algorithm only accepts odd values for
WinSize. When using 0th derivative, a minimum of 7 or 9 smoothing points should be used, even if this
slightly degrades the curve. Greater number of smoothing points may improve repeatability further,
though eventually the signal itself will be smoothed out. Even a small amount of smoothing will reduce
the StDev significantly, thus improving repeatability. For 1st derivative, 5 smoothing points is often used.
More smoothing can improve the curve, and may improve repeatability.

For narrow peaks, mostly <400nm, less smoothing is generally applied – usually 9-13 points for 0th
derivative, and 5-9 points for 1st derivative. For peaks >>400nm, more smoothing can be applied, with
values 25-41 not uncommon. This is because the data points on the spectrum are at fixed intervals of
0.033nm, even though peaks >>400nm are wide and resolution is worse than <400nm.

Smoothing can reduce noise, but too much smoothing will reduce the signal as well. Below is the Ti336
peak with 0th derivative, 9 points of smoothing:

And with 15 points applied:

Note that as more smoothing is applied, the region boundaries may need to be widened. Also, as more
smoothing is applied, neighboring peaks may blend into the region of interest. See an example with this
Zr339 peak with 7 and 13 points of smoothing applied. At 7 points (left), the two peaks are well defined,
but at 13 points (right), the neighboring Ti peak at 339.45 spills into the Zr339 region:
When using the 1st derivative, it is also possible to apply too much smoothing such that the peak signal is
reduced or neighboring peaks are blended into the region. However, this can be more difficult to
visualize. Using 1st derivative with minimal smoothing can be useful in separating one peak from an
adjacent peak, since it is possible to only integrate part of the peak. In practice, though, this can be
difficult, and often 0th derivative works better for separating adjacent peaks.

When setting the region for 1st derivative preprocessing, ensure that it is wide enough to include the
whole peak. Recall that regions using 1st derivative are less sensitive to WL shifting, so region boundaries
need not account for this. It is important not to make the region too wide, since any signal within the
region is integrated.

It is possible to integrate only half of the derivative peak, but this is more advanced and is not commonly
done. It is also possible to use 2nd and higher order derivatives in Profile Builder, but this offers little
benefit and often degrades the signal too much to be useful.

Preprocessing and regions may be different depending on whether the line is used in the numerator or
denominator. For example, a Ni peak like Ni341 may be used in the numerator and denominator in the
Ni base, but a user may find that one set of preprocessing parameters works better in the numerator,
and another in the denominator.

Savitzky-Golay smoothing allows users to change the polynomial function used for fitting, which is given
by the Degree option. This is normally kept at 2, since changing it has little effect on the calibration.
SciAps does not recommend trying to optimize the degree used.

Step 6) Determine fit parameters

Once the intensity ratio is determined, the calibration is fit with a polynomial. The free parameters are
Degree, which determines a linear (degree = 1) or quadratic (degree = 2) fit, and Force Zero, which
determines whether the y-intercept (curve offset) will be zero or whether it can be calculated using the
data points.

LIBS calibration curves should in general be linear, so this is preferred. Linear curves are especially the
best choice if:

- The concentration range is small

- Precision of measurement (or RMSE) is poor
- Extrapolation beyond the calibration range is needed
- The high end of the calibration range only has 1 or 2 data points

Quadratic fits may be used for some calibration plots. These are most useful when:

- The calibration range is wide, spanning more than an order of magnitude of concentration
- There are many data points across the calibration range
- The denominator is not necessarily constant, e.g. for a binary alloy

Quadratic fits can model real phenomena like self-absorption and variable denominator. But they can
also model statistical variation or other deviation from linearity that may not apply to all samples, which
is to say, they model trends in the data that aren’t real. Profile Builder allows calibration curves of
degree >2, but these are not recommended.

RMSE will always be lower when using a quadratic fit. Consider the below Ti in Al base plots

While the quadratic fit (right) offers a better fit, the concentration range is low and the fit’s curvature is
heavily weighted by the two data points at higher concentration. Linear would be preferred in this case.

For many curves, there is only one high concentration sample. These are effectively “one-point
calibrations” and should always be linear, e.g. Zn in Al base
To define a quadratic curve, a minimum of 3 data points are needed (2 if force Zero). In practice,
normally at least 4-5 data points should be used, spread at regular intervals over the calibration curve to
make a good quadratic fit. Consider Si in Cu base

There are too few data points at high concentration to say whether the curvature is real, so linear fit is
preferred.

The improvement to RMSE must also be significant to choose a quadratic fit over linear. Consider Mn in
Cu base:
The concentration range is wide enough to consider quadratic fit. However, the change in RMSE is not
significant, and the curvature is strongly influenced by the single point of CDA 958 which lies just off of a
linear curve. Linear fit would be preferred in this case, though there is little risk in using a quadratic.
However, if the calibration range is wide enough, and the number of data points large enough, quadratic
fit may be preferred, as in Zn in Cu base

The change in RMSE is small, but the curve is so well-characterized that quadratic would be preferred. In
the case of Zn in Cu, it is also known that the denominator of the IR is not constant (Zn468/Cu510 in this
case) since Cu content changes significantly as Zn content increases.

Quadratic fits should normally be concave up, especially if the curvature is due to self-absorption.
Concave up indicates that the IR numerator is increasing more slowly at high concentration and the
denominator is relatively constant. Concave down fits are allowable in cases where the denominator is
not constant, or contains lines from the numerator element, e.g. Ni in Ni base

A quadratic fit that is only slightly concave down is usually not a problem, but in this case the user
should consider whether a linear fit is more appropriate.

In some cases, a quadratic fit is clearly better, even over narrow calibration ranges or with few data
points, as in Ni in Al base
Where only two data points anchor this fit. In this case, which fit to use is a judgement call by the user
(SciAps uses quadratic for this fit).

Quadratic fits should normally never have a local minimum or maximum within the calibration range, as
in this Cu in Ti base example:

This case is a one-point calibration, so linear fit is more appropriate. If a quadratic fit is more clearly
needed and the fit gives a local minimum, try adjusting the regions, preprocessing, or denominator
choices to improve the curve.

Force Zero should be used when there are few blanks in the calibration (few samples with zero or near-
LOD content) or when the impact on RMSE is very small. Offsets are frequently due to inadequate
background subtraction, but could be due to other effects. Allowing an offset (Force Zero off) is more
often used with 1st derivative regions in the numerator, since the SGOLAY with absolute value algorithm
guarantees that the IR will never be zero.

When few blanks are present, force zero may be necessary to preserve a more realistic fit, especially
when using quadratic fits. Take, for example, Cr in Stainless
Both are quadratic fits, but because the concentration range begins at 16%, the curvature is dominated
by scatter in the 16-19% range rather than real trends across the whole calibration range. Note that
allowing an offset (left) puts a local minimum within the calibration range, which is not desirable. This
plot is best fit with quadratic/force zero (right), or linear/offset (not shown).

Force zero is especially useful when there are very few samples in the calibration. Consider Si in CuNi

The calibration range here is very low, and only 2 data points with Si assays are included. The lower
concentration sample may not represent a real offset, but may lie off the force zero linear fit due to
statistical variation or some other matrix effect. There are too few data points to know for sure, and the
deviation from the curve is not large. So Force Zero is appropriate here.

Force zero should not be used when the IR for blanks is strongly negative, which can happen when an
emission peak is weak and the region boundary may catch a neighboring peak, as in V326 in Fe
The calibration curve at low concentration has negative IRs, but correlation across the whole calibration
range is otherwise good

Having a blank IR << 0 is not ideal, but is allowable if the calibration curve is otherwise good. Note,
strongly negative IR is also common for Si, among other elements.

In some cases where the calibration range is all at the % level (no blanks), e.g. Cr in StainlessStd, the best
choices for calibration may be between linear/allow offset and quadratic/force zero. Which to use is a
judgement call left to the user

To summarize good fit practices:

 - Linear is preferred over quadratic
- Force Zero is preferred over allow offset
- Quadratic fits should not show excessive curvature and should not have a local
minimum/maximum within the calibration range
- Quadratic fits should normally be concave up, rarely concave down
- RMSE should be minimized as long as the curve parameters can be justified

Region selection and curve fitting is often an iterative process. First, determine a set of acceptable
regions, examine the fit, and then adjust the regions and preprocessing to improve the fit by lowering
RMSE and sticking to the best practices outlined above.

Step 7) Remove any outliers from the calibration curve

Outliers can happen for all sorts of reasons:

1) Bad test
a. Dirty surface
b. Sample not flush with window
c. Ran out of Argon
2) Peak from some other element within the region
3) Sample contamination
4) Overlapping peak on the region boundary – especially when using 0th derivative
5) Statistical variation
6) Matrix differences
a. Denominator not applicable
7) Matrix effects due to secondary absorption
a. Presence of one element is absorbing or quenching signal
8) Bad assay
9) Analyzer drift

1) Bad Test – check spectrum to see if overall signal is much lower than other samples. If it is,
consider reshooting. Check that the sample is clean, and grind, Dremel, or swab with alcohol if
not. Check analyzer Argon level – Profile Builder should warn when Ar is low, but check in the
alloy app. Ensure that the sample is held flush with the window.
a. Repeatedly low signal could indicate a matrix difference, where the sample does not
ablate in the same way as other samples in the grouping, or a hardware problem on the
analyzer.
2) Peak from another element – most regions commonly used in LIBS have few cases of overlap
from other elements, which is why those regions have become popular. Overlaps can happen
for uncommon elements, analytes with few strong peaks (thus fewer opportunities to switch to
an overlap-free line), and weak matrix lines (weak lines from the bulk element – Fe and Ti are
known to have many weak lines that appear at high concentrations).
Profile Builder does not presently have the ability to correct for other elements’ contributions to
a given region, but that functionality will be added.
3) Sample contamination – related to (2), can be due to contamination on the surface or
unassayed content in the sample. Laser cleaning helps with surface contamination, but cleaning
methods recommended in (1) can be helpful.
4) Overlapping peak on region boundary – Adjust the region or preprocessing to avoid the
overlap, especially if the overlapping peak is only in some samples, not all. Consider alternate
peaks.
5) Statistical variation – examine individual spectra or data points to see if there is significant test-
to-test variation that may have caused the average to be an outlier. This is especially common at
low concentrations. Consider reshooting the sample.
6) Matrix differences – the most common case of this is when the denominator is not consistent
with other samples in the grouping. An example may be a Fe ore sample with 50% Fe content
when other samples in the grouping are closer to 35% Fe. Also common is a case where the
denominator is a sum of several matrix elements, e.g. Ni+Fe+Cr, but one of those is not present
in the sample. Consider removing this sample from the calibration, or changing the denominator
to better include this sample on the curves.
7) Other matrix effects – high content of some other element can have absorption or plasma
quenching effects that suppress signal of other elements. A common example is Cu in Al base,
where high Zn content is known to cause deviation on the Cu curve

E993 and 7050 both have 5% Zn content and consistently lie off of the curve. They are excluded
for now, though Profile Builder will soon have the ability to enter interelement correction
factors.
Other matrix effects and matrix differences account for most cases of outliers when sample or
instrument effects are ruled out and no other explanation is apparent.
8) Bad assay – if all other sample effects are considered, bad assay information is a possible cause.
This more commonly happens with customer samples that may have only been tested once,
may be inhomogeneous, or has some error when written down. If a peak is present when the
sample should be a blank, or the data point is an outlier on an otherwise good curve when
denominator signal is consistent, then bad assay may be to blame. Bad assays on commercially
available CRMs are uncommon but not unheard of.
9) Analyzer drift – If some time has elapsed between calibration shots, e.g. the analyzer was
calibrated at the factory and samples are added at some later date, instrument drift can cause
outliers on the calibration curve. This can also happen if there has been some environmental or
temperature change or the analyzer has been shipped. Presently, Profile Builder has no way to
account for drift within a calibration model, only correct for it after the calibration has been
done. If drift is suspected and the calibration must be modified, a recalibration of that model
may be necessary.
What to do about outliers

1) Reshoot sample
2) Exclude point from that curve
3) Remove sample from the model (remove from all curves)