Peletier Article

focal point
M. J. PELLETIER
JET PROPULSION LABORATORY
4800 OAK GROVE DRIVE
PASADENA , CALIFORNIA 91109-8099
Quantitative Analysis
Using Raman
Spectrometry
INT RODUCTIO N
aman spectroscopy has become m ore popular in recent

years because improvements
in instrumentation have allowed users to focus on their applications
rather than on the operation and limitations of the instrument. A single
Raman spectrum can provide a large
amount of information about a sample. Its m any well-resolved spectral
features often provide good speci city for qualitative analysis and good
analyte selectivity for quantitative
analysis. Comm on m atrix or container materials, such as water, pharm aceutical excipients, glass, and cataly st sup ports, hav e w eak R am an
spectra, and therefore generate relatively little interference during analyte determination.
For m any practical applications
the greatest strength of Raman spectroscopy is exible sampling. Fiberoptic probes allow Raman measurements to be made hundreds of meters from the Raman analyzer, an impo rtant b en e t for h ars h process
chemistry applications. M ost forms
of sample, such as powders, slurries,
pellets, bers, emulsions, or lms,
can be analyzed as is without sample
20A
Volume 57, Number 1, 2003
preparation. Raman samples can be

transparent, scattering, or opaque.
They can be a liquid, a solid, or a
gas. A Raman probe only needs to
illuminate the sample and collect
scattered light from the sample. No
contact with the sample is required.
The sample can even be inside a
closed container as long as light can
reach the sample. Confocal optics
can strongly attenuate the measured
Raman signal from material in front
of or behind the intended sample,
giving Raman analysis strong spatial
discrimination. Raman can also be a
m icrosco pic tech niq ue h avin g an
analysis volume of only a few cubic
micrometers.
Raman spectroscopy is more often
associated with the determination of
molecular structure and with qualitative analysis than it is with quantitative analysis. Applications of Raman spectroscopy for the quantitative analysis of sample composition
have, however, been described in the
technical literature for more than 50
years. A naly te concentra tion s are
typically in the range of 0.1 to 100%,
but in favorable cases, concentrations below 100 ppm can be accurately measured. A list of more than
200 published quantitative Raman
applications is given in the Appendix, organized by application and analytical method. This list is by no
means exhaustive of the published
quantitative Raman applications.
The result of a quantitative analysis is a value specifying the amount
of something in a sample, an estimate of the uncertainty in that value,
and if possible, independent information that tests the validity of those
results. The trend toward automating
analytical techniques in order to address very large data sets and to reduce the impact of operator bias on
the analytical results places greater
importance on robustness and validity testing. Raman experts are not always available to interpret spectra
and make sure the analytical results
are reasonable. Even when experts
are available, how do they know
their expectations for what is reasonable are correct? W hy do the analysis if the operator has greater con dence in his expectation than in the
analytical results? The real value of
the analytical method depends on its
credibility when providing surprising
results. Raman is well suited for
quantitative analysis because a single
R am an sp ectru m o ften contains
enough information to address all
three of these analytical needs for

several components of the sample simultaneously.
This Focal Point article reviews
the techniques that are used to extract quantitative information from
Raman spectra. Only spontaneous
Raman scattering will be considered.
Other forms of Raman scattering,
su ch as su rface-enh an ced R am an
spectroscopy (SERS), coherent antiStokes Raman spectroscopy (CARS),
reson ance R am an scattering , an d
stimulated Raman scattering have all
been used for quantitative analysis,
but will not be reviewed here. The
rest of this paper is organized as follows: The second section explains
the fundamental basis for quantitative analysis using simple Raman
spectra. The third section describes
comm on types of noise in Raman
spectra. The value and utilization of
noise is emphasized. The fourth section describes preprocessing operations that are used to enhance extraction of analytically useful information from Raman spectra. The
fth and sixth sections describe the
quantitative determination of Raman
band descriptors such as band area
and band position that can be related
to analyte concentration. The seventh and eighth sections describe
univariate calibration and multivariate calibration, respectively. The Appendix lists over 200 published examples of quantitative analysis using
Raman spectroscopy organized by
analyte identity and data processing
method.
2. BASIS FO R QUANTITATIVE
RAMAN SPECTROSCOPY
The intensity of Raman scattered
light from a sample, I (n) R , is given 1,2
by Eq. 1.
I (n) R 5
2 4p 3
hI N (n 0 2 n) 4
L
2
4
45 3 c mn(1 2 e 2 h n /k T )
2 1 7(g9) 2 ]
3 [45(a9)
a
a
(1)
where
c
h
IL
N
5
5
5
5
Speed of light
Plancks constant
Excitation intensity
Number of scattering m olecules
n 5 M olecular vibrational frequency in Hertz

n0 5 Laser excitation frequency
in Hertz
m 5 Reduced mass of the vibrating atoms
k 5 Boltzmann constant
T 5 Absolute temperature
a9a 5 Mean value invariant of the
polarizability tensor
g9a 5 Anisotropy invariant of the
polarizability tensor
Equation 1 shows that the Raman
scattering intensity is proportional to
the number of m olecules, N, in the
sample volume being probed by the
Raman instrument. I (n) R is actually
th e inten sity in tegrated ov er the
width of the Raman band. The proportional relationship between Raman scattering intensity and analyte
concentration is the basis for m ost of
the quantitative analysis done using
Raman spectroscopy.
Quantitative analysis can also be
done using changes in Raman band
position or band shape. A Raman
band is often composed of two or
more unresolved overlapping bands
from spectroscopically non-equivalent species in the sample. These
bands have maxima at slightly different wavenumber positions. The
observed Raman band is broadened
by the superposition of the bands
from all the overlapping non-equivalent sp ecies. T his spectrosco pic
ban d-broaden ing m ech an ism is
called inhomogeneous broadening.
A change in the relative concentrations of the non-equivalent species
will cause the observed Raman band
po sition and /or shap e to chang e.
These changes in observed Raman
band position, width, or shape can be
empirically related to concentration
changes in the sample, but the fundamental m echanism for this type of
quantitative analysis is still the proportional relationship between Raman scattering intensity and species
concentration.
Equation 1 provides additional insight into the use of Raman spectroscopy for quantitative analysis.
Raman intensity is proportional to
the intensity of the excitation source,
I L , so analyzer sensitivity can be in-
creased by using a stronger laser.

Most important, however, is the fact
that the product of the constants in
Eq. 1 is a ver y sm all number. Production of a single Raman photon
may require 10 8 excitation photons.
This small cross-section for spontaneous Raman scattering limits its
sensitivity and leaves the Raman signal vulnerable to being obscured by
more ef cient optical processes such
as uorescence.
Quantitative Raman analysis relies
on the principle of linear superposition. The Raman spectrum of a mixture is equal to the weighted sum of
the Raman spectra of the components that make up the mixture. The
weights are determined by the Raman cross-sections and quantities of
the components in the sample. Notice that a component making up the
mixture may not have the same Raman spectrum as that same component in its pure state. Chemical interactions between sample components can signi cantly m odify their
Raman spectra. In this case, the pure
com p onen t effectiv ely becom es a
different material with a different
Raman spectrum when it is incorporated into the mixture. The principle of linear superposition applies
to the different material actually
in the mixture, not necessarily to the
pure com ponents.
The principle of linear superposition works because Raman scattering
cross-sections are ver y small. The
probability of a Raman scattered
photon being lost because of another
Raman scattering interaction is essentially zero. The reduction of the
excitation light intensity by Raman
scattering is also essentially zero.
There is no optical coherence between components in the sample,
leaving no mechanism for Raman
scattering by one component in the
sample to in uence the Raman scattering of another component. The
principle of linear superposition does
break down, however, when the absorption spectrum of one or m ore
components signi cantly affects the
transm ission of excitation or Raman
scattered light to or from the analyte.
This breakdown is the Raman analog
APPLIED SPECTROSCOPY
21A
focal point
to the inner- lter effect in uorescence spectroscopy.
3. NOISE IN RAM AN
SPECTRA
A Raman spectrum can be divided
into two parts: the signal and the
noise. The signal is the part of the
Raman spectrum that contains the
desired inform ation. The noise is the
part of the Raman spectrum that contains unwanted information. Since
the subject of this article is the determination of analyte concentration
from a Raman spectrum, a Raman
band that can be used for this purpose is an example of a signal. Random intensity uctuations could not
be used to determine analyte concentration and would, therefore, be an
example of noise. A value for analyte concentration has very limited
utility, however, without estimates of
the uncertainty and robustness of
that value. Often, random intensity
uctu atio ns rep re sen t the m ajo r
source of uncertainty in the analysis
of the Raman signal. Then, for the
calculation of concentration uncertainty, the random intensity uctuations become the signal and the analyte Raman band becomes the noise!
In keeping with the subject of this
Focal Point, intensity in a Raman
spectrum that is not used to determine analyte concentration is considered noise. That noise, however, will usually provide valuable
inform ation about the uncertainty
and robustness of the analysis. The
lack of various types of noise can
also be used to improve con dence
in analytical results.
The analysis of the Raman signal
requires its isolation from the noise.
This process is never complete, so it
is generally desirable to maximize
th e R am an sig nal and m in im ize
noise in the Raman spectrum. An
important metric describing the quality of a Raman spectrum is the signal-to-noise ratio, or S/N ratio. The
uncertainty in a reported S/N ratio is
usually dominated by the uncertainty
in the noise measurem ent because
noise is generally m ore dif cult to
quantify than the signal. An understanding of the types of noise typi22A
cally present in Raman spectra can

be useful for isolating the Raman
signal from the noise and maximizing the S/N ratio. Noise in a Raman
spectrum , or collection of Raman
spectra, can be classi ed into ve
different groups: photon shot noise,
sample-generated noise, instrum entgen erated n oise, com pu tatio nally
generated noise, and externally generated noise.
Photon Shot Noise. Photon shot
noise is a result of the statistical nature of light. Its intensity is equal to
the square root of the number of detected photons. For example, if a
constant-intensity light source delivers 1 000 000 photons to a noise-free
detector in one second, m any replicate one-second measurements will
have an average value of 1 000 000
photons with a standard deviation of
1000 photons. Photon shot noise is
an inescapable noise source in Raman spectra. W hen the S/N ratio of
a Raman spectrum is at the photon
shot-noise limit, nothing can be done
to improve the S/N ratio without collecting more photons. Photon shotnoise calculations can therefore provide valuable insight during the development of an analytical method.
Photon shot noise can also be used
to determine the gain of a chargecoupled device (CCD) detector.3
Sample-G enerated Noise. Sample-generated noise includes Raman
intensity changes caused by sample
changes not related to analyte concentration. F or ex am ple, R am an
band shape, position, and integrated
intensity can all change with sample
temperature. These changes tend to
be small, but they can become important when doing spectral subtraction or multivariate m odeling. Sample heterogeneity can also create
sample-generated noise because the
sample composition at the point of
analysis does not necessarily represent the entire sample correctly. This
occurs in coarsely m ixed static samples or dynamically in m oving system s. Co m m on exam ples in clude
owing liquids containing bubbles
or particulate matter and owing
powders having variable solids-to-air
ratios.
Sample-generated noise also inclu des un w anted op tical intensity

generated by the sample. The m ost
important example of this is uorescence. F luo rescen ce in a R am an
spectrum usually looks like a slowly
chang ing cur ved b aselin e. B lack body radiation from hot samples also
looks like a slowly changing curved
baseline. Both uorescence intensity
and b lackbo dy radiation intensity
can range from zero to orders of
magnitude greater than the analyte
Raman signal. Transition metal complexes can produce spectrally sharp
luminescence, especially in the near
infrared, that can be mistaken for
Raman bands. Unlike Raman bands,
the emission wavelengths of these
bands do not change with changing
excitation wavelength.
Instrument-Generated Noise. Instrument-generated noise depends on
the speci cs of the instrum ent design. This type of noise includes various detector noises such as thermal
noise, xed pattern noise, wavelength dependence of quantum ef ciency, and read n oise. D rift in
wavenumber or intensity axis calibration adds noise to Raman spectra
that can be important for repetitive
measurem ents and for the long-term
stability of analytical models. Variability in the presentation of the sample to the instrument can strongly affect the measured Raman intensity
and can even affect different parts of
the Raman spectrum differently due
to chromatic aberration in the collection optics. In most of these examples, the instrum ent-generated noise
can be detrimental to the determination of analyte concentration from
the Raman spectrum, but it can also
provide valuable information about
the performance of the instrument.
F or ex am ple, ex cessiv e rand om
noise in the baseline may indicate
that a detector has overheated. Good
analytical m ethods use this type of
inform ation to improve the robustness of an analysis against degraded
instrument performance.
C o m p u t a t i o n a l ly G e n e r a t e d
N o ise. Com putationally generated
noise refers to noise that is added to
the Raman spectrum after the output
of the detector is digitized. A lot of

signal processing m ay be involved in
converting the digitized detector data
into the raw spectrum the operator
sees. Operations as innocent as dark
subtraction, apodization, or intensity
calibration can add intensity to a Raman spectral region in a way that is
detrimental to quantitative analysis.
For example, random noise in the
spectrum used for dark subtraction
or intensity calibration becomes a
non-random noise when applied to
sets of spectra. Raman band broadening due to apodization can shift intensity from a non-analyte band into
an analyte band. Data processing designed to remove noise from a Raman spectrum can also add new
noise. For example, baseline correction using polynomials or fractals
tends to produce functions that oscillate around the true value of the
background they are trying to remove. When these functions are subtracted from the Raman spectrum,
the residual intensity variations appear in the baseline-corrected Raman
spectrum. Again, these noise sources
provide information that can be used
to improve the robustness of an analysis. For example, strong bands of
negative intensity in a Raman spectrum resulting from dark subtraction
can compromise an analysis. They
usually indicate a faulty dark spectrum or a failed spectrograph shutter.
An analytical method that tests for
and nds negative Raman bands can
report that the analysis results are
questionable and propose possible
rem edies.
Externally Generated Noise. Externally generated noise is an unwanted signal in the Raman spectrum that does not originate from the
Raman instrument or the sample. It
is usually caused by a source of external light that scatters into the optical path of the Raman instrument.
P roperly design ed R am an in stru ments and sample compartments are
generally imm une to external light
sources. Good sample compartments
are not always available for in situ
analysis, however. W hen doing reaction m onitoring in glass laboratory
reactors, uorescent room lighting is
a comm on source of externally generated noise. In the near-infrared

spectral region, the signal from uorescent room lighting is m ostly due
to atomic lines from argon and mercury. In the visible spectral region,
stro ng ato m ic m ercu r y lines at
404.66, 435.83, 546.07, 576.96, and
579.07 nm are superimposed on the
broad background from the phosphors used in the uorescent lighting
tube. An analytical method that tests
for these atomic lines in each Raman
spectrum is less likely to report
faulty analytical results w itho ut
w arn ing w h en u orescent ro om
lighting is a problem. Sunlight is another potential source of externally
generated noise. The spectrum of
sunlight is a broad continuum containing m any spectrally sharp features due to m olecular absorption in
the earths atmosphere and atomic
absorption in the atmosphere of the
sun (Fraunhoffer lines).
A non-optical source of externally
generated noise is intensity spikes
created when high-energy particles
such as cosm ic rays pass through the
detector of the Raman instrument.
These cosmic ray events release a
large number of electrons that are
electrically indistinguishable from
photoelectrons. These electrons are
usually con ned to one or occasionally a few detector elements in an
array detector. The result is a huge
intensity spike in the dispersive Raman spectrum that is generally one
detector element wide. These spikes
occur infrequently, at random times,
and in random locations of the Raman spectrum. Analytical methods
th at ign ore cosm ic ray intensity
spikes will sooner or later produce
erroneous results due to a cosm ic ray
event. In FT-Raman instrum ents, the
cosmic ray events occur in the interferogram of FT-Raman instruments
rather than in the Raman spectrum.
The effect on the FT-R aman spectrum is, therefore, not localized to a
small spectral region.
4. PREPROCESSING OF
RAMAN SPECTRA
Any set of m athematical operations applied to the data prior to
analysis is called preprocessing. Preprocessing improves the accuracy

and robustness of quantitative analysis and m akes analytical data easier
to understand and process. Preprocessing can separate the signal from
some of the noise by taking advantage of the differences in their properties. Properties commonly used for
this purpose include (1) information
content of spectral segments, (2) the
frequency content of the Fourier
transform ed Raman spectrum, (3)
the intensity variance in replicate
Raman spectra, (4) intensity variance
in a set of Raman spectra having different analyte concentrations, and (5)
known or predictable intensity patterns in the Raman spectra. Preprocessing can also compress a set of
spectra by removing non-informative
data. M ultiplicative methods to correct Raman spectra for changing laser power or sample opacity, such as
normalization or multiplicative scatter correction, are considered in the
fth section as part of the analytical
method rather than as part of the preprocessing.
Information Content of Spectral
Segments. The signal in a Raman
spectrum often has signi cant intensity in a relatively small portion of
the entire spectrum. The rest of the
spectrum is noise. A large fraction of
the noise is often concentrated in
sm all p ortio ns of th e spectru m .
Eliminating parts of the spectrum
that contain little signal or excessive
noise can enhance calculations for
determ ining analyte concentration.
This conceptually simple form of
preprocessing is especially effective
for Raman spectra because Raman
bands from various molecular vibrations tend to be well resolved from
one another. As a result, signal intensities representing a vanishingly
small fraction of the total Raman
spectrum intensity can often be used
for accurate quantitative analysis.
A detailed knowledge of the analyte Raman spectrum and the noise
sources in the Raman spectra of the
samples can be used to select limited
spectral regions for quantitative analysis. Spectra of sample components
can often be determined experimenAPPLIED SPECTROSCOPY
23A
focal point
tally or found in published collections of Raman spectra. 46 W hen a
spectrum cannot be found, spectra of
similar m aterials can guide the selection of an analytically useful spectral
region. Spectral information can also
be estimated from the functional
groups they contain and published
spectral locations for various functional g ro up s. 6 S pectral regio ns
known to contain excessive noise,
perhaps due to other sample components, can be eliminated from analyte concentration calculations.
Another way to identify spectral
regions that are useful for quantitative analysis is to analyze differences
between the Raman spectra of samples having different analyte concentrations. Spectral regions that are
predictive of analyte concentration
must change with analyte concentration. The variance of Raman intensity from a set of samples containing
differing amounts of analyte can
therefore be used to target spectral
regions that are likely to be predictive of analyte concentration. A plot
of this variance as a function of Raman shift is called a variance spectrum. Spectral regions in the variance spectrum having large values
are m ore likely to be predictive of
analyte concentration than those regions having small values.
A ny R am an inten sity ch an ges
among the set of samples, whether
related to analyte concentration or
not, will increase the corresponding
values in the variance spectrum. A
more informative spectrum is a plot
of how m uch of the variance is consistent with the changes in analyte
con cen tration . This spectrum is
called a correlation spectrum or correlogram.79 At each wavenumber
point in the correlation spectrum, the
Raman intensities from the set of
samples are regressed against the actual analyte concentrations. The correlation coef cient, R, of the regression is then plotted as a function of
Raman shift. Regions of the correlation spectrum having values approaching 1 or 21 are predictive of
analyte concentration, while those
approaching 0 are not. Other regression statistics can be used to m ake
24A
similar plots. Simply selecting wavenumber regions that correlate with

analyte concentration may not give
optimal results, however, because
wavenumber regions that correct for
the in uence of interfering signals
may be omitted.
A third approach for identifying
the optimal spectral segments for
quantitative analysis is to create and
test predictive models using subsets
of the entire Raman spectral region.1013 An exhaustive search of all
combinations of Raman spectral elements is not practical, however. For
example, comm on CCD-based Raman spectrometers produce Raman
spectra having 1024 Raman spectral
elem ents. A n exhau stiv e searc h
would require creating and testing
2 1024 2 1 predictive m odels. Search
strategies to avoid this combinatorial
explosion have been developed using methods such as data compression, simulated annealing, or genetic
algorithms. A relatively new search
strategy, interval partial least-squares
(iPLS), 13 uses a graphical interface to
involve the analyst in the search process. Analytical models are built using partial least squares (PLS, see
eigh th section) o n eq ually sized
spectral segments. Analytically useful segments are optimized for size
and location.
Fourier Frequency Content. The
Fourier transform of a Raman spectrum is a linear combination of sine
waves with varying phase relationships that exactly reproduces the Raman spectrum . Figure 1 illustrates
some waveform s that are important
in Raman spectra and their frequency content. Figure 1a shows simulated Raman bands of equal height
and width having a Gaussian and a
Lorentzian pro le. The derivative of
the Gaussian multiplied by 2 is also
shown. Figure 1b shows that the
Lorentzian band has a greater portion of its frequency content at lower
frequencies than the Gaussian band.
Taking the derivative of the Gaussian band greatly reduces its low-frequency content. Figure 1c shows a
typical Raman baseline caused by
uorescence. Figure 1d shows that
nearly all of the intensity in Fig. 1c
is derived from the rst few Fourier

frequencies. The rst few Fourier
frequencies of the Lorentzian and
Gaussian bands are also the strongest
in their Fourier spectrum, but they
make a m uch sm aller contribution to
the total intensity. These lowest Fourier frequencies have negligible intensity in the d erivative o f the
Gaussian band.
The signal typically has different
relative amounts of the various sine
w av e frequen cies th an the no ise
does. Attenuating those frequencies
that make little contribution to the
signal, but signi cant contribution to
the noise, separates some of the
noise from the signal. A mathematical operation that attenuates some
frequencies more than others, and/or
shifts their relative phases, is called
a lter. Several types of lters are
useful for separating the Raman signal from different types of noise.
These lters can be implemented by
carr ying out operations on the Fourier transformed Raman spectrum
followed by an inverse Fourier transform, but they are more comm only
im p lem ented by the equivalent
method of convoluting the Raman
spectrum with a ltering function.
A low-pass lter, also called a
smoothing lter or apodizing lter,
attenuates low frequencies less than
high frequencies. Low-pass lters
are useful for separating random
noise from the Raman signal. The
points in a random noise waveform
are uncorrelated, so the intensity of
the waveform is fairly evenly distributed across the range of Fourier frequencies. The highest frequencies in
the Fourier transform of the Raman
signal contribute less to the Raman
signal than m iddle-range frequencies, so a low-pass lter attenuates
random noise m ore than it attenuates
the Raman signal. Notice, however,
that the rem aining random noise now
more closely resembles the Raman
signal. Attenuation of high frequencies in the Fourier transform of the
Raman signal broadens its Raman
bands and reduces the spectral resolution.
Spectroscopists often prefer a Raman spectrum having a wavenumber
FIG. 1. Synthetic Raman signals and baseline, and their Fourier transforms. Each waveform consists of 2048 points (pixels) covering a spectral range of 4330 cm 2 1 . Only the signicant Fourier frequencies of the total 1024 are shown. (a) Gaussian band with
FWHM of 25 cm 21 is red, Lorentzian band with FWHM of 25 cm 2 1 is blue, and derivative of red Gaussian band times 2 is green.
(b) Intensity of Fourier frequencies of waveforms shown in (a). (c) Shape of typical Raman baseline. (d) Intensity of Fourier frequencies making up the baseline shown in (c).
separation between adjacent points

that is much smaller than required to
fully utilize the spectral resolution of
the Raman instrument. The highest
frequencies in the Fourier transform
of the Raman spectrum then unambiguously represent noise rather than
Raman signal. An estimate of the
noise intensity in the Raman spectrum is useful for estimating uncertainty and for distinguishing between
real Raman bands and noise in qualitative analysis. It is therefore useful
to examine the part of the Raman
spectrum removed by a low-pass lter as well as the part rem aining after
the ltering operation. Unfortunately, many commercial Raman instruments apply apodizing or smoothing
operations before making the Raman
sp ectra availab le to th e operator.

Useful information in the noise is
then irretrievably lost.
A high-pass lter attenuates high
frequencies less than low frequencies. High-pass lters are useful for
separating the Raman signal from
curved baselines. A curved baseline
is an intensity that changes slowly
with Raman shift. It is often caused
by the uorescence spectrum of impurities in the sample and may be
many times stronger than the Raman
signal. S im p le h igh -p ass ltering
that affects only a few lowest frequencies in the Fourier transform of
th e R am an sp ectru m usually removes enough of the baseline from
a Raman spectrum to facilitate analysis without seriously distorting the
Raman bands. Higher frequencies

present in the curved baseline, while
much weaker than the lowest frequencies, are less attenuated by the
high-pass lter and can distort bands
rem aining in the Raman spectrum.
Attenuation of low frequencies in the
F ou rier transform of the R am an
spectrum narrows Raman bands and
can therefore improve spectral resolution, but also causes Raman bands
to exhibit negative overshoot and
ringing.
First and second derivatives of
Raman spectra are often used for
quantitative analysis. Since the number of molecules, N, probed by the
Raman analyzer in Eq. 1 is not a
function of the scattered light frequency, the derivative of Raman
25A
focal point
scattered intensity with respect to
Raman shift rem ains proportional to
the analyte concentration. Derivatives are actually a special kind of
high-pass lter. Taking the rst derivative of a Raman spectrum multiplies the amplitude of each frequency component in the Fourier
transform of the Raman spectrum by
a value proportional to its frequency.
The rst-derivative operation also
shifts the phase of each frequency
component by 908. Derivatives provide a fairly robust and automatic
baseline correction and spectral resolution enhancement to Raman spectra. They also strongly enhance highfrequency noise and can make analytical models more sensitive to temperature variation.14
A bandpass lter attenuates both
the high and low frequencies relative
to the middle-range frequencies. Applying a bandpass lter is equivalent
to sequentially applying a low-pass
lter and a high-pass lter. Since the
Fourier transform of a typical Raman
signal is composed mainly of middle-range frequencies, bandpass lters are usually m ore effective at
separating the Raman signal from
the noise than high-pass or low-pass
lters alone. A derivative is usually
used in conjunction with a low-pass
lter in order to reduce the enhancement of high-frequency noise. This
combination is actually a special
kind of bandpass lter.
Representing a Raman spectrum
as a linear co m binatio n of sine
waves provides powerful bene ts,
but sine waves are not always the
best basis set. Each sine wave is a
single frequency that extends over
the entire Raman spectrum, so a ltering operation used to separate signal and noise at one location in the
Raman spectrum affects all the rest
of the spectrum as well. Raman
bands are localized in sm all regions
of the Raman spectrum, so representing the Raman spectrum as a linear combination of more localized
functions m ay be bene cial. This is
the idea behind the windowed Fourier transform or short-time Fourier
transform .15 Fourier transform s and
ltering operations are carried out
26A
independently on segments of the

Raman spectrum to reduce longrange effects. S im ilar reason ing
leads to the use of wavelets1517 to
represent a Raman spectrum. Wavelets have variable extent across the
Raman spectrum and a variable frequency content. Wavelet ltering operations are similar to Fourier transform ltering operations, but m ay
sometimes be better suited to the analytical needs of Raman spectra. 18,19
Intensity Variance of Replicate
Spectra. The reproducibility of Raman spectra from a single sample
provides information that can be
used to separate the Raman signal
from various sources of noise. Replicate m easurements of the same
sample provide the same Raman signal, bu t d iffering ra n dom no ise.
Summ ing n replicate spectra causes
the signal intensity to increase linearly with n. The random noise and
the resulting signal-to-noise ratio increase linearly with the square root
of n. In addition to summing the n
replicate spectra, the standard deviation of the n values for each spectral element can be calculated, giving
an estimate of the random noise in
the spectral element intensity. The
standard deviation can also be used
to detect outliers in the set of n intensities at any spectral element. The
spectral element sum mation, with
app ro priate scaling , can ex clud e
these outliers and thereby be more
rob ust against rand om in tensity
spikes. 2024 It is even possible to t
the n values for each spectral element to a line or curve in order to
detect and qu an tify non -ra n dom
trends in the replicate m easurements,
such as photobleaching of uorescent impurities. If such trends are detected, a weighted summation with
outlier rejection may provide better
intensity values than the simple summation. Clearly, there is much to be
gained by paying attention to the
noise in replicate Raman spectra.
It is also possible to detect and
eliminate cosmic ray intensity spikes
without using variance in replicate
measurem ents. 2528 Cosm ic ray intensity spikes in Raman spectra are generally one to two detector elements
wide. If all Raman bands are known

to be wider, this property can be used
to distinguish between a cosm ic ray
intensity spike and Raman bands.
This condition is often true because
the spectral resolution of many Raman instruments corresponds to several detector elements on the optical
dispersion axis. W hen a cosmic ray
intensity spike is detected, the entire
spectrum can be discarded, or, m ore
often, the effected points in the Raman spectrum can be replaced by estimates of what the intensity should
have been.
Intensity Variance in a Set of
Raman Spectra Having Different
Analyte Concentrations. The intensity variance in a set of spectra can
also be used to eliminate redundant
inform ation from the Raman spectra.
This elimination can m ake the data
easier to interpret and can discriminate between the Raman signal and
the noise. Raman spectra typically
contain many more points than are
needed to represent the analytical
signal. Adjacent points are not independent of each other for m any
reasons including spectral interpolation or sm oothing, the natural linewidth of Raman bands, and the spectral resolution of the Raman instrument. Even points that are well separated from each other m ay not be
independent because most materials
have m ultiple Raman bands.
One way to reduce the number of
points in a Raman spectrum is to express each spectrum in the set of
spectra as a linear com bination of a
small number of orthogonal functions. The scaling constants for the
orthogonal functions are a set of
numbers that can replace each Raman spectrum without loss of analytical information. A Fourier transform can do this, but m any sine
waves are required to adequately
represent the Raman spectrum. Principal components analysis (PCA) is
a mathematical tool that can determine the best orthogonal functions
(called fa cto rs , prin cipal co m ponents, or PCs) to represent a set of
Raman spectra. Conceptually, PCA
rst determines the function (called
PC1) that simultaneously represents
sets that are becoming comm on in

applied Raman analysis. Some automated baseline-correction methods
attempt to identify points in the
spectrum belonging to Raman bands,
and then t the remaining points to
an appropriate baseline-estim ation
function. 3234 Such methods are usually iterative. Other properties of the
baseline that can sometimes be used
to separate it from the rest of the Ram an spectrum inclu de excitation
w av elength indep en dence 3 5 , 3 6 and
baseline intensity ch anging w ith
time.37
5. RAM AN BAND AREA O F
TH E ANALYTE
FIG. 2. Integrated signal intensity (green), noise intensity (blue), and S/N ratio (green)
for a synthetic Lorentzian band (FWHM
50 cm 2 1 ) plus random baseline noise as a
function of the distance from peak maximum that is included in the integration. The
peak of the Lorentzian band was 1 unit high and the noise had a standard deviation
of 0.2 units. The S/N ratio is optimized for integration including all points in the isolated Lorentzian greater than about 1 3 of the peak height when random baseline noise
is the dominant noise type.
as m uch variance in the set of spectra as possible. PC1 is then removed

from all the spectra and the process
is repeated to create PC2, which represents less variance in the set of
spectra than PC1. Successive PCs
are calculated until all the variance
in the set of spectra has been represented. Usually the rst several PCs
represent nearly the entire analytical
signal, and the rest of the PCs represent mostly noise.
One of the dif culties with PCA
is determining how many PCs to use,
or stated another way, determining
how m uch of the variance in the data
set is useful. A low-variance signal
(weak Raman band) in the presence
of much higher variance signals can
vanish into the noise of higher PCs.
Appropriate preprocessing prior to
PCA, or limiting PCA to appropriate
segments of the Raman spectra, can
minimize these problems. Analysis
of the excluded higher PCs (assumed
noise) can som etimes detect analyte
inform ation and other trends in the
data set.
K nown or Predictable Intensity
Patterns. Spectral stripping 29 can be

a powerful preprocessing m ethod. If
the spectrum of a component in the
sample is known, it can be subtracted from the Raman spectrum of the
sample. The remaining Raman spectrum of the sample contains less intensity that might interfere with the
analysis of other signals. Spectral
stripping can also be used for quantitative analysis. It is described in
greater detail in the eighth section.
Curve tting methods can be used
to estimate the shape of the baseline.
The baseline estimate can then be
subtracted from the Raman spectrum. M anual baseline estimation using curve tting requires an operator
to specify points in the Raman spectrum that should be zero if no baseline perturbation were present. A
po lyn om ial fun ctio n, 30 m ultiple
polynom ial functions,31 or a spline
function may be t to those points to
estimate the baseline. Operator intervention is undesirable, however, because it is likely to introduce operator bias into the data and because it
is not practical with the large data
Determination of Band Area.

The analyte concentration in a sample is proportional to the integrated
intensity of an analyte Raman band.
Raman bands are often well resolved
from one another, but usually sit on
top of a non-zero baseline. The integrated intensity of the Raman band
is, therefore, a simple summation of
th e m easu re d intensity ov er the
wavenumber range of the Raman
band after the baseline intensity is
subtracted. The integrated intensity
of a Raman band is proportional to
its height as long as its band shape
does not change. Quantitative Raman methods are frequently based
on either band area or band height.
Noise in the Raman spectrum determines whether band area or band
height can be m easured m ore precisely. If shot noise from the Raman
band is the dominant source of noise,
band area can be m easured more precisely than b and h eigh t b ecause
more photons are included in the
m easurem ent. S om etim es, though,
band height provides greater accuracy.38 40 Band-height measurements
often have an advantage over bandarea m easurements when there is
partial overlap between the analytical band and another band in the Raman spectrum. Including more photons by integrating the whole Raman
band can actually reduce the S/N ratio when the main source of noise is
random noise in the baseline. This is
demonstrated numerically in Fig. 2,
where the extra signal measured in
27A
focal point
the wings of the Raman band is less
than the extra noise included into the
measurem ent.
Smoothing prior to band-height
measurem ent com bines the properties of a simple band-area measurement and a simple band-height measurement. A band-height measurement after sm oothing is equivalent to
a band are a m easurem ent w here
points are weighted by their proximity to the peak intensity. This approach uses m ore of the detected
photons than a simple band-height
measurem ent, but also discriminates
against both low signal-to-noise-ratio points at the wings of the Raman
band and intensity from other partially overlapping Raman bands. The
smoothing function can be adjusted
to optimally compromise the bene ts
and weaknesses of band-area and
band-height measurem ent.
Another way to improve upon
simple band-area or -height measurement is to utilize additional information about the shape of the Raman
band by curve tting. Raman band
shapes are often Lorentzian because
vibrational coherence decays exponentially with time and the Fourier
transform of an exponential decay is
Lorentzian. The shape of spectrally
narrow er R am an ban ds is m o re
strongly affected than broader bands
by the response function of the Raman instrument, which often looks
like a Gaussian, but may have asymmetry due to optical aberration. The
shape of simple, moderate-spectralresolution gas-phase Raman bands
may be largely determined by unresolved vibrational-rotational bands. 41
Curve tting provides an estimate of
the Raman band area, height, band
width, and baseline that is consistent
with the user-supplied band shape. It
uses all the detected photons in the
Raman band, discriminates against
other partially overlapping Raman
bands, and can improve baseline estimation. Curve tting must be used
cautiously, though. It is an iterative
process that can provide different results for the same data when the
starting co nditio ns are ch an ged.
Also, it often uses several operatoradjustable param eters that can be
28A
used by an expert operator to unintentionally re inforc e his incorrect

prejudices.
Uncertainty and Error in Band
Area. Once the area of the analyte
Raman band is determined it is necessary to estimate its uncertainty.
The best way to do this is to determine the standard deviation of bandarea m easurements from replicate
spectra. A good replicate measurement repeats all the operations that
could contribute to a change in the
Raman band area, possibly including
sam p le preparation an d sam ple
placement in the Raman instrument.
A different approach for estimating
band area uncertainty is to identify
the major sources of noise, quantify
them, and calculate their effect on
band area. Ideally, both approaches
for estimating band-area uncertainty
should be used. Agreement between
these approaches indicates that the
sources of measurement uncertainty
are understood and may provide insight into how the measurem ent precision might be improved. For example, if the standard deviation of
the band area is equal to that predicted due to photon shot noise, the
only ways to improve the measurement precision are to increase the
number of detected photons from the
analyte or to decrease the number of
detected ph oto ns fro m the back ground.
Several factors other than analyte
concentration can contribute to the
analyte Raman band area. Their in uence on the accuracy of quantitative analysis needs to be kept small
by good experiment design and appropriate data analysis. The Raman
intensity for an analyte band changes
with the refractive index and the
temperature of the sample. Refractive index in uences the electric eld strength from the exciting light
that the analyte molecule experiences. In the absence of large dipole
moments, the m ultiplicative correction, L, for integrated band area due
to the refractive index of the sample
is given by:
L 5 (n s /n 0)(n s2 1 2) 2(n 02 1 2) 2 /81 (2)
where n s is the refractive index of the
sample for the Raman scattered light,

and n 0 is the refractive index of the
sample for the exciting light. Temperature also affe cts the R am an
cross-section by changing the fraction of analyte molecules in excited
vibratio nal states. T his effect is
quanti ed by the Boltzmann factor
in Eq. 1 (1 2 e 2 h n/ kT ). Lower frequency vibrations are m ost strongly
affected. Temperature can also alter
Raman band areas by changing the
sample density and therefore the
number of analyte molecules in the
sampled volume.
Interactions between the analyte
and its solvent or m atrix typically affect analyte bands that are involved
with the interaction. Bands whose
areas are modi ed usually experience shifts in the band positions and
so m etim es chan ges in the ban d
widths as well. A good example of
interactions between the analyte and
the solvent causing changes in analyte band intensity 42 is shown in Fig.
3. The intensities of two acetone
bands are plotted as a function of acetone concentration in chloroform .
One band shows a linear relationship
with acetone concentration, and the
other does not.
Optical absorption by the sample
can affect Raman bands by attenuating the laser intensity reaching the
sample or by attenuating Raman
photons returning from the sample.
This effect is called self-absorption.
Changes in the concentration of the
absorbing sample component, or in
the depth of the Raman analyzer focus in the sample, can change the
Raman band area of the analyte.
S elf-ab so rp tio n o ften alters som e
parts of the Raman spectrum m ore
than other parts, distorting the relative band areas. Near-infrared absorption bands of com mon solvents
can signi cantly reduce the intensity
of analyte Raman bands,43 45 especially at longer wavelengths. Compilations of near-infrared absorption 46,47
spectra are useful for anticipating, detecting, or correcting near-infrared
self-absorption. Self-absorption is a
more common problem when ultraviolet excitation is used or when resonance Raman measurements are
FIG. 3. Relative band areas of the C O (green) and CH (red) stretching vibrations of
acetone diluted with chloroform. The intensity of the C O vibration is affected by its
micro-environment. Adapted from Ref. 42.
made. Raman spectra can sometimes

be corrected for self-absorption.48,49
Band Area Normalization. Probably the most important factors in uencing analyte Raman band area other than analyte concentration are
those that change the overall ef ciency of getting photons to and
from the sample. These are multiplicative factors due to changes in laser
power, optical alignment of the Raman analyzer, optical coupling of the
sample to the Raman analyzer, sample opacity, material buildup on the
window between the sample and the
Raman analyzer optics, and m any
other experimental variables. Analyte band area determinations that
rely on these experimental variables
rem aining constant are said to be
based on absolute Raman intensity. Quantitative analysis using absolute Raman intensity is often successful for single data sets collected
in the laboratory. Several published
examples are referenced in the Appendix.
Several things can be done to stabilize the experimental variables for
quantitative analysis using absolute

Raman intensity. Drift in Raman analyzer perform ance can be minimized by doing the entire calibration
and experiment in a short time, operating the instrument in a constant
temperature environment, and avoiding any instrum ent adjustments during the calibration and experiment.
Lasers inside Raman analyzers often
use optical feedback to reduce laserpower drift and to keep noise below
the 0.5% level. Some instrument param eters or sample properties m ay
be dif cult to control, however, so it
is desirable to minimize the sensitivity of the m easu rem en t to their
changes. For example, the effect of
variation in sample placement relative to the Raman analyzer on measured Raman intensity can be reduced by using collection optics having a larger depth of eld or by using
special apparatus to m ore precisely
position the sample. Flow cells provide extremely precise sample placement. The effect of variation in sample transmission can be reduced by
minimizing the distance that light
travels through the sample. This can

be achieved by focusing the Raman
analyzer near the sample-cell window or reducing the depth of eld of
collection optics.
If experimental variables that affect Raman band area do change,
quantitative analysis is still possible
by measuring the change and applying a correction to the measured
area. An external standard is often
used for this purpose. An external
standard is an unchanging sample
that is periodically m easured in order
to quantify changes in the Raman
analyzer that affect quantitative analysis. The ratio of analyte band intensity to external standard band intensity is proportional to analyte concentration and is ind epend en t of
many experimental variables, such
as laser power. The external standard
is not necessarily similar to the materials being analyzed, but similarity
can provide added robustness. Ideally, the entire optical path from the
laser to the detector for the external
standard should be identical to that
of the samples being analyzed.
External standards can be used
with several different interfaces between the sample and the Raman analyzer. A ow cell can provide nearly identical optical paths for presenting liquid external standards and
samples to the Raman analyzer. Different cuvettes holding the external
standard and the sample provide
nearly identical optical paths, but
correction errors can occur if the Raman analyzer focus is not at the
same location in each cuvette. One
channel of a multi-channel Raman
analyzer can m easure the external
standard while other channels simultaneously m easure the samples. The
lack of a time delay between measurement of the external standard
and the samples eliminates error due
to short-term variations in instrument parameters such as laser power.
Th e m ulti-chann el ap proach also
bene ts from having an external
standard m easurement for each sample measurem ent. No extrapolation
of instrument performance between
external standard measurem ents is
needed. External standards work
29A
focal point
well and are frequently used for
qu antitativ e R am an analy sis. D e
Paepe et al.38 report coef cients of
variance below 1% using an external
standard. Many other applications
using an external standard are described in the Appendix. The primar y weakness of the multi-channel
external standard approach is potential non-equivalence of the optical
paths of the external standard and of
the sample. For example, a focusing
change at the sample or fouling of a
collection optic would change the
analyte band area without changing
the band area of the external standard.
The use of an external standard
eliminates the in uence of many experimental variables that have to be
carefully held constant when doing
quantitative analysis using absolute
Raman intensity. External standards
eliminate the effect of long-term
drift, allowing a quantitative analysis
method to work reliably over a long
period of time. The external standard
can also provide veri cation of proper instrum ent perform ance. Unfortunately, the external standard does not
correct for changes in sample properties such as opacity, refractive index, or moving inhomogeneity. External standards m ay not be practical
nor fully corrective for single-channel kinetics or for on-line measurements.
The limitations of external standards can be addressed by mixing
the standard into the sample. Sample
properties that affect analyte band
area equally affect the band area of
the standard. Ideally, no Raman intensity from th e stand ard w o uld
overlap the analytical band of the
sample. This type of standard is
called an internal standard. The concept of an internal standard was developed m ore than 70 years ago 50,51
to address similar concerns for quantitative analysis using atomic emission. Internal standards provide correction for changes in the sample
properties as well as for changes in
the Raman analyzer. The measurements of the standard and the analyte
can be simultaneous even for singlechannel analyzers and are made un30A
der identical m easurem ent con ditions. Unfortunately, the addition of

a foreign material such as the internal standard to the sample can cause
changes in the analyte, resulting in
measurem ent error. Equilibria m ay
be shifted or chemical reactions m ay
occur. Addition of a standard m ay
not be practical for continuous online analysis and can be time consuming for laboratory analysis. Adding a foreign material to the sample
also sacri ces two of the m ost important bene ts for many applications of Raman spectroscopy: noninvasive analysis and no need for
sample preparation.
M o st R am an spectra contain
bands that do not change with the
concentration of the analyte. By definition, these bands are a type of
noise. However, all form s of noise
contain potentially useful information and these bands are not an exception. They can be used as internal
standards. The bene ts of using an
internal standard can then be obtained without adding a foreign material to the sample.
Solvent bands often make good
internal standards. The ratio of the
analyte band area to the solvent band
area is proportional to the ratio of the
analyte m olar concentration to the
so lven t m o lar co ncentration. T he
solvent concentration is essentially
constant, as long as it is much larger
than the changes in analyte concentration. To the extent that the solvent
concentration does change with analyte concentration, the use of a solvent band as an internal standard
will cause the analyte band area to
have a nonlinear relationship with
analyte concentration. Fortunately,
the m agnitude and functional form
of this nonlinearity are predictable
and can be corrected when predictive
analytical m odels are created.
Water is a particularly good internal standard for aqueous solutions
because of its high m olar concentration and its minimal interference
with most of the Raman spectral region. Water has two Raman bands
that are commonly used as internal
standards: a broad, weak band centered near 1640 cm 2 1 and a broad,
stro ng ban d cen tere d near 3 450

cm 2 1 . The strong band is very sensitive to ionic strength, temperature,
and anything else that changes hydrogen bonding. The weaker band is
much less sensitive to these factors,
so it is the preferred internal standard
band when it is not too weak for accurate analysis.
The Ram an band of a functional
group that is comm on to all the
changing m aterials in a sam ple can
be used as an internal standard.
Even though the m olecules containing that functional group are changing in concentration, the m olar concentration of that functional group
is not. It can therefore be used as an
internal standard. A good application for this type of internal standard is the quantitative analysis of
m onom er concentration during a vin y l p o lym eriza tio n . T h e R am an
band from the C5C stretching vibration can be used as the analyte
band and a Raman band from a
m onom er side-group can be used as
the internal standard. The internal
standard Ram an band m ust be chosen carefully, how ever, because the
Raman cross-section of som e m onom er side-groups changes after polym erization. For example, the ring
breathing m ode of the styrene phenol group is m uch stronger than the
corresp onding vibration in polystyrene.5254
T h e inter na l stan da rd R am an
band does not have to be a single
Raman band. It can be the sum of
several band areas, the integrated
intensity of a whole spectral region,
or even the integrated intensity of
the entire Ram an spectrum . 14,5557
Since different m olecular vibrations
can have greatly different Raman
cross-sections, it is not obvious that
the sum of several integrated band
areas would function ver y well as
an internal standard. The success of
this approach is probably due to a
combination of factors including an
averaging of cross-sections, a m inimal change in the intensities of
m ost of the spectrum area w ith
changing analyte concentration, and
m anagem ent of the correction error
by the calibration step that follow s.
The use of several bands sum med

together as the internal standard is
m o re f un d am en ta lly sou n d w he n
each band is weighted by its relative
cross-section.58 These relative crosssections are usually not known initially, but can be determ ined from
the Raman spectra of m ixtures of
standards.
Som etimes the desired quantitative information about a sample is
the concentration ratio of two analytes. The ratio of their Raman band
areas provides the sam e type of
m ultiplicative correction as ratioing
a single analyte to an internal standard. If band-area ratios and relative
Raman cross-sections are known for
each com ponent of the sample, mass
balance can be used to calculate the
actual concentrations of each com ponent. An interesting variation on
quantitative analysis using analyte
band ratios occurs when the two
analyte bands are not spectrally resolved, but have different band-center positions. The band center of the
unresolved bands is then approxim ately proportional to the fraction
of one analyte in the m ixture of the
two. This approach has been used to
determine the conversion of one
crystalline polym orph into another
in a slurry 59 and the concentration
of aqueous phosphoric acid. 60
6. QUANTITATIVE ANALYSIS
USING RAM AN BAND
POSITION
Quantitative Raman analysis of
sample composition and other properties d eterm ined b y the sam ple
composition is almost always based
on or derived from band area. Raman band position (or width) is m ost
often used for the quantitative determination of sample properties not
determined by composition, such as
crystallinity, stress, or temperature.
Raman band position depends on
atomic m asses and the strength of
the chemical bond that joins them.
The atomic m asses for a given composition do not change, so Raman
band position is a measure of the
strength of the chemical bond for a
given material. Local molecular order, external force, and temperature
can all change the chemical bond

strength, however. To the extent that
they do, they can be quantitatively
measured by the resulting Raman
band position. For example, stress in
silicon lms is proportional to the
freq uency shift o f the p ho non
band 61,62 near 521 cm 2 1 . Nylon temperature can be determined from the
band position of the NH 2 stretching
vibration.63 Polymer crystallinity is
often determined from the shape of
a Raman band.56,64,65 Less crystalline
polymers have a broader range of
micro-environments, causing a broader range of band positions for those
vibrations affected by micro-environment. Band position data are usually correlated with sample property
empirically, so precision and consistency throughout the data set are
usually more important than absolute
accuracy in the determination of
band position.
Raman band position may be de ned as the location of the maximum
intensity or the location that separates half of the band intensity from
the other half. Both are equivalent
for symmetric bands. The location of
maximum intensity is more commonly used. The Raman band position can be determined by simply locating the maximum intensity point.
More precise results are obtained by
tting a quadratic function to points
near the peak of the band and determining the m aximum of the function. The zero-crossing point of a derivative can also be used to determine band position, but factors such
as band asymm etry and smoothing
associated with the derivative calculation can cause the calculated
band position to differ from the actual point of maximum intensity.
Curve tting can also be used to
measure band position very precisely. Curve tting utilizes all the points
in the Raman band and knowledge
about the band shape to enhance the
precision of Raman band-position
determination.
Fourier interpolation can be used
prior to determining band position,
especially if few er than 10 points
de ne the band. Fourier interpolation adds no inform ation to the Ra-
m an spectrum , but allows the band

m aximum to be accurately located
within a fraction of the interval separating adjacent points in the Raman
spectrum . The uncertainty in band
position can be as much as 50 times
sm aller than the spectral resolution
of the Ram an analyzer. Fourier interpolation does not provide m uch
bene t for cur ve tting, however.
7. UNIVARIATE
CALIBRATIO N
The goal of calibration is to nd
a mathematical relationship between
the metric extracted from a Raman
spectrum , such as band area, and the
desired property, such as analyte
concentration. This mathematical relationship is called an analytical
model. W hile it is theoretically possible to calculate an analytical model
from rst principles, in practice several necessary constants such as absolute Raman cross-section and optical collection ef ciency are rarely
known. As a result, analytical m odels are almost always created by
measuring Raman spectra of known
samples (the training set) and empirically relating the Raman m etric to
the known property. The accuracy of
the resulting analytical model is tested by a second set of known samples
(the validation set) and by a statistical analysis of how well the analytical m odel ts the training-set data.
The analytical model is also examined for how robust it is likely to be
against expected variation in future
samples.
Since Raman band area is proportional to analyte concentration, a plot
of analyte concentration vs. analyte
Raman band area produces a straight
line. The analytical m odel is then y
5 a 1 bx where x is the analyte concentration and y is the area of the
analyte band in its Raman spectrum.
Assuming all the error is norm ally
distributed random noise in the y
variable, the standard deviation, s,
for the slope, b, and intercept, a, is
given by: 66
31A
focal point
FIG. 4. Plot of normalized Raman intensity from the analyte as a function of analyte
concentration. The green curve results from normalization using an internal standard
whose concentration does not change with changing analyte concentration. The red
curve results from normalization using an internal standard whose concentration decreases with increasing analyte concentration. In the case shown, the mole fraction of
the analyte plus that of the internal standard equals one and both materials have the
same Raman cross-section.
1
s a2 5 1
n
s b2 5
s e2
(x i 2 x ) 2
2
(x i 2 x )
where
1
n 22
s e2
i5 1
s e2 5
n
i5 1
the root m ean square error of prediction, RMSEP, given by:
x 2
(3)
i5 1
y i2 2 a
i5 1
2b
i5 1
xi yi
yi
The assumption of negligible error in

the analyte concentrations of the
samples used to m ake the training set
is often a good one since these concentrations are often prepared or determined under ideal laboratory conditions. W hen the error in analyte
concentration is not negligible, linear
least-squares tting and uncertainty
estimation become more complex.6770
The uncertainty in predicted concentration can also be estimated by
32A
RM SEP 5
i5 1
(c i 2 c i ) 2
n
(4)
where n is the number of samples in

the validation set, c is the true analyte concentration, and c is the calculated concentration of the analyte.
This statistic gives an error estimate
that is independent of analyte concentration. W hen the error is known
to be dependent on analyte concentration RMSEP can still be used, but
it should be calculated over a limited
concentration range if enough data
points are available. The RMSEP
statistic can also be used to identify
potential outliers in the calibration
data set. Individual points whose error of prediction, (c i 2 c i), is m uch
larger than the RM SEP are potential
outliers that could introduce unnecessary error into the analytical m odel.
Another valuable way to test the
accuracy and robustness of an ana-
lytical m odel is to insert noise into a

sample spectrum and determine the
resulting error in the concentration
calculated by the analytical model.
The noise is chosen to represent variations that m ight be present in measurements of future samples. For example, if a Raman analyzer has a
calibration shift speci cation of 0.1
cm 2 1 per 8C and a temperature range
of 610 8C is expected during future
m easurem ents, sh iftin g a sam ple
spectrum by 61 cm 2 1 and applying
the analytical model will provide an
estimate of how this variation will
affect the accuracy of the analysis.
Another way to improve the robustness of an analysis is to use a
second analytical m odel on the same
data set. The second analytical model may be based on a different Raman band or may use a different calculation on the same analytical band
used in the rst analytical model.
Lack of agreem ent between the two
analytical models provides a warning that the results of the analysis are
suspect. A common cause for disagreem ent between two models is a
sample that is not well represented
by the training set used to create the
models. This often happens when
process conditions are changed, new
feed streams are used, or problems
develop in the Raman analyzer. The
simultaneous use of two or more analytical m odels also provides some
robustness against software bugs.
Calibration Curve Nonlinearity.
A plot of analyte concentration vs.
analyte Raman band area should be
linear, but sign i cant d eviations
from linearity can occur. Nonlinearity can result from normalizing analyte band areas with an internal standard band whose area changes with
analyte concentration. Figure 4 demonstrates this behavior for an analyte
dissolved in a solvent. The green calibration curve results from using an
internal standard whose concentration does not change with analyte
concentration. The red calibration
curve results from using the solvent
as the internal standard. At low analyte concentration the solvent concentration is approximately constant
and the calibration curve is approx-
imately linear. Higher analyte concentration displaces the solvent from

the sample volume causing nonlinearity in the calibration curve. One
way to avoid the nonlinear behavior
is to eliminate the normalization and
build an analytical m odel based on
absolute Raman intensity. A second
approach is to incorporate the concentration dependence of the internal
standard on the analyte concentration into the analytical model.71 A
third approach is to use the linear
model of analyte concentration vs.
normalized analyte band area and
then correct the error due to internal
standard dilution using the known
relationship between analyte concentration and solvent concentration. A
fourth approach is to use a piecew ise analy tical m o del or locally
w eighted re g re ssio n analy sis, 72 ,7 3
where a subset of the standards having nearly the same analyte concentration as the unknown are used to
predict the analyte concentration in
the unknown. A classi cation step is
needed prior to prediction in order to
decide which standards in the training set to use. The piece-wise analytical m odel essentially breaks a
nonlinear calibration curve into a series of straight-line segments.
Nonlinearity in the plot of analyte
con cen tration vs. analy te R am an
band area also occurs when changing
analyte concentration alters the scattering cross-section of the analyte or
internal standard band. This can occur when the atoms involved in the
vibration interact with their microenvironment and the m icro-environment changes with analyte concentration. Binary m ixtures of acetone
and chloroform demonstrate this effect, as shown earlier in Fig. 3. M ore
subtle effects due to interactions between the analyte and the rest of the
sample can be identi ed by correlating the analyte prediction error with
the concentration of other components in the sample.
Another source of nonlinearity is
the existence of the analyte as m ore
than one species in equilibrium. If
the analyte Raman band comes from
only one of the equilibrium species,
Raman band area will not usually be
proportional to total analyte concentration. For example, a weak organic

acid in water is mostly dissociated at
low concentration, but is increasingly associated with increasing concentration. The Raman band area of the
CO 22 vibration therefore has a nonlinear relationship with total acid
concentration, while the band area of
a different vibration of the acid molecule m ay not.
Sample changes not included in
the training set can cause an analytical method to fail. This problem is
especially comm on in process analysis. The optimal samples for the
training set are often not available
w itho ut causing a plan t u pset.
Changes in feedstock sources m ay
m od ify feedstock com p ositio n in
ways that are detrimental to an established analytical m ethod. Even
the temperature at the m easurement
point m ay vary with the season of
th e y ear. T hese un co ntro lled or
weakly controlled variables m ake error detection and robustness critical
for analytical methods used for process control.
8. M ULTIVARIATE ANALYSIS
OF RAM AN SPECTRA
Som etim es quantitative analysis
requires the use of m ore than just
one Raman band. Consider the prediction of octane num ber in gasoline
from its Raman spectrum. Gasoline
is a complex m ixture of hydrocarbons. Its octane number is determined by characteristics such as arom aticity an d b ran chin g th at are
shared by many gasoline components. Raman spectra are sensitive to
these characteristics, but no single
Raman band is suf cient to determine octane number. Octane number
can be determined, however, by the
simultaneous analysis of the entire
Raman spectrum or of spectral segm ents con tain ing several R am an
bands. Single analytical m ethods that
utilize m ultiple quantities or signals
are called multivariate m ethods. The
mathematics of multivariate methods
differ from those of the univariate
methods described in the previous
sections, but the underlying principles for relating Raman spectra to
quantitative information about the

sample remain the same. Raman
band areas are proportional to analyte concentration. Raman band position, width, and shape sometimes
contain useful information as well.
Even when univariate analytical
models can be used, multivariate analytical m odels have several advantages over them. Including m ore
than one analytical Raman band in
the analytical model provides a signal averaging bene t. M ore analyte
photons in the analytical model improves the shot-noise-limited signalto -n oise ratio. Ev en non -ra n dom
noise, such as detector xed-pattern
noise, can be reduced by including
more of the Raman spectrum in the
analy sis. M ultiv ariate analytical
methods use m ore information than
is necessary to determine a sample
property. The excess information enables diagnostics, outlier detection,
and error estimation that are not possible with univariate methods when
a single piece of information is used
to predict a single sample property.
Similar diagnostics and error estimation could be obtained by doing
two or more independent univariate
analyses and then examining regions
of the spectrum that do not contain
analyte Raman bands. Multivariate
software packages do this autom atically.
M ultivariate analysis can aid in
the identi cation of the spectral regions that are useful for quantitative
analysis, as shown in the fourth section. Variance or correlation spectra
depend strongly on how the Raman
spectra were normalized, however.
Consider what would happen if an
analyte Raman band were used to
normalize a data set. Analyte band
areas would not correlate with analyte concentration. Raman bands that
normally were unchanging with analyte concentration, such as solvent
bands, would correlate with and be
predictive of analyte concentration.
Such a model would become unstable near the detection limit, though.
Normalizing a data set with a signal
that changed, but in an uncorrelated
way with analyte concentration,
33A
focal point
could hide analytically useful information.
The strength of multivariate analy tical m od els o ver u niv ariate
ones their use of more of the information present in the data can
also be a weakness. Noise in the
training set that happens to correlate
with analyte concentration can be incorporated into the analytical model.
The internal consistency of such an
analytical m odel may be very good,
but it m ay fail miserably at predicting analyte concentration in new
samples when the chance correlation
breaks down. Noise can also be the
dominant source of intensity variation between spectra in the training
set, obscuring a weak, localized signal due to the analyte. Even when
such noise is not strong, it can degrade the perform ance of an analytical model. Hirschfeld 74 reports an
interesting example of this in qualitative analysis. The quality of library
searching was observed to improve
when randomly selected portions of
the spectrum were excluded from the
search. The improved searching apparently resulted from minimizing
the impact of impurities in the library spectra! M ultivariate analytical
models can also be very sensitive to
x-axis error. Small shifts in wavenumber calibration create derivativelike intensity variation between spectra that some multivariate methods
will try to correlate with analyte concentration. Powerful integrated multivariate analysis software packages
can provide plausible, but wrong, results. It is important to carefully test
multivariate m odels and to understand the basis for their predictions.
S everal differe n t m ultivariate
methods that have been used to analyze Raman data are brie y described and contrasted below. A detailed description of these methods is
beyond the scope of this article, but
can be found in chemometrics textbooks7577 and in the speci c references noted.
Spectral Stripping. The spectral
stripping operation consists of multiply ing the pu re analyte R am an
spectrum by a constant and subtracting the resulting product from the
34A
Raman spectrum of the sample. 29

The analyte concentration is proportional to the value of the constant
that eliminates the analyte Raman intensity from the Raman spectrum of
the sample. A calibration curve and
some form of spectral normalization
are usually required to convert this
constant into an absolute analyte
concentration. Spectral stripping has
been described 78 as a mathematical
titration. It is often used m ultiple
times on the same sample spectrum
in conjunction with library searching
to determine the identities of multiple components in a sample.
There are several different ways to
determine the value of the spectralstripping constant. In the simplest
case, an isolated analyte band exists
in the Raman spectrum of the sample. The correct value of the spectral-stripping constant rem oves this
band, and produces a synthetic Raman spectrum of the sample at zero
analyte concentration. If no isolated
analyte band can be found, the value
of the spectral-stripping constant that
minimizes the complexity of the
sample spectrum can be used. Spectral complexity, the indicator for the
mathematical titration, is measured
by the rank of the training set spectral matrix. Ideally, the rank of this
matrix equals the number of components in the sample. The term
component may include not only
the different chemical species in the
sample, but also variables such as
temperature that are not constant in
the data set. Complete subtraction of
the spectrum of one component reduces the rank of the matrix by one.
This m ethod of determining spectral
complexity is called rank annihilation.7882 Spectral stripping can also
be done using an iterative leastsquares method,82 by minimizing the
absolute area of a derivative, 83 or by
a ratio method.84 An important assumption in the spectral-stripping
operation is that the analyte spectrum is not changed by the rest of
the sample. This assumption is often
violated when the analyte interacts
with other components in the sample
through hydrogen bonding or dipoledipole interactions.
Cross-Correlation. Raman band

area can also b e d eterm ined by
cross-correlating 85 the Raman spectrum of the sample with the Raman
spectrum of the analyte.86 Equation 5
below describes this operation.
`
C (x) 5
S (t) A ( x 1 t) dt
(5)
2`
C(x) is the correlation function, S (t)

is the Raman spectrum of the sample, and A(x 1 t) is the Raman spectrum of the analyte. The value of the
correlation function at x 5 0 is proportional to the analyte band area.
The main bene t of Raman bandarea determination using cross-correlation is that it does not require
separation of the analyte band from
the Raman spectrum of the matrix.
Another bene t of cross-correlation
is the improved signal-to-noise ratio
that can result from using the entire
Raman spectrum rather than a small
spectral segment. The value of the
correlation function at x 5 0 for zero
analyte concentration depends on the
spectrum of the analyte matrix and
is generally not equal to zero. Analyte band area determination using
cross-correlation should, therefore,
probably be limited to cases where
the analyte m atrix spectrum does not
change very m uch.
Direct Classical Least Squares
(D C L S ). D irect classical least
squares nds the linear combination
of spectra from the pure components
making up the sample that m ost
closely matches the Raman spectrum
of the sample. The m ultiplicative
constants used on the pure spectra
are proportional to the concentrations of their respective components.
The difference between the sample
spectrum and the linear combination
of pure component spectra is called
the residual spectrum. The residual
spectrum is the portion of the sample
spectrum that is not described by the
pure component spectra and is useful
for diagnostic purposes. DCLS requires Raman spectra of all the pure
materials that m ake up the sample.
Like spectral stripping, DCLS assumes that the pure component spectra are not changed when the pure
components are mixed together.
Indirect Classical Least Squares

(IC L S). Ind irect classical least
squares rst estimates the Raman
spectra of the pure components that
make up the sample from a series of
mixture spectra and then applies
DCLS. ICLS is useful when pure
component spectra are not available
or when interactions between the
sample components change the pure
com p onen t spectra. Lik e D C LS ,
ICLS needs to account for all the
variation in the sample spectrum using only the component spectra that
make up the sample. Intensity in the
sample spectrum from anything other than the estimated pure component spectra creates error in the analysis or shows up in the residuals
spectrum.
Inverse Least Squares (ILS). Inverse least-squares methods express
analyte concentration as a function
of spectrum variables. Each intensity
value in a digitized spectrum is a
spectrum variable. Notice that unlike
classical least-squares m ethods, ILS
does not require inform ation about
the other components that make up
the sample. This is a big advantage
since such information is often not
available. An important requirement
for all ILS methods is that the number of samples in the training set
must be equal to or greater than the
number of spectrum variables. Raman spectra often contain thousands
of spectrum variables, making the
number of samples required for the
training set unreasonably large. Several ILS methods are differentiated
by the way they reduce the size of
the training set to a m anageable
number of samples. ILS methods
used for Raman analysis include
multiple linear regression, principal
com p onen ts analy sis, and partial
least squares.
A subtle, but very important point
about ILS m ethods is that they correlate analyte concentration with the
sample spectrum. This is different
from a direct cause-and-effect relationship. Random noise may happen
to correlate with sample concentration in a given training set. Poor experiment design m ay cause spectral
features that are not from the analyte
to correlate with analyte concentration. In both cases, internal diagnostics may indicate good predictive
performance, but the resulting model
may have poor robustness or m ay
even fail completely when applied to
new samples.
Can random noise really look like
the analytical signal? Consider the
digits of p. They can be used as a
random number generator since they
have no particular pattern . . . or do
they? If p is calculated in base 26,
each digit can be assigned to a letter
of the alphabet. Starting at the 8th
digit after the decimal point we can
nd the phrase R am an IsB est.
Starting at digit 54 is the phrase
To IR isH u m an and at dig it
136 290 is the rest of the phrase
ToRamanDivine. Although many
may consider these phrases to be an
analytical signal worthy of attention,
they are really just examples of
chance correlation.87
M u ltip le L in ear R eg ressio n
(M LR). Multiple linear regression is
one example of an ILS model. MRL
reduces the number of samples required in the training set by using
only a limited number of the spectral
variables in the Raman spectra. Selection of the spectral variables may
be based on spectral knowledge. It is
also possible to optimize MLR models by iteratively selecting variables
and testing the predictive performance of the resulting model. A bene t of MLR models is their simplicity. The elimination of most of the
spectral variables, however, reduces
m ultiv ariate sig nal av eraging an d
greatly limits m ultivariate diagnostic
capabilities.
Principal C om p onent R egression (PCR). Principal com ponent regression is another type of ILS model. PCR reduces the number of samples required in the training set by
using principal components analysis
(PC A) to compress the spectra without signi cant loss of information.
The thousands of variables in each
original spectrum are replaced by a
linear combination of several new
variables (factors or loadings). PCR
can be thought of as PCA followed
by M LR, but the actual computation
does these operations simultaneously. Unlike MLR, PCR retains the

multivariate signal averaging bene t
and the diagnostic capabilities typical of an over-determined system.
Pa rtial L ea st S q u ares (P L S).
Partial least squares is probably the
most popular ILS m odel for quantitative Raman analysis. Like PCR,
PLS uses factor analysis to compress
the size of the spectra and to remove
redundant information. The difference between PCR and PLS is in the
detail of how the compression is
done. PLS uses the analyte concentration inform ation along with sample variance in the compression process to create factors that are correlated w ith analyte con cen tratio n.
PCR just uses sample variance. If the
variance in the training set due to analyte concentration is sm all compared to other sources of variance,
PLS is expected to have an advantage over PCR. In most other cases
it is not clear that PLS offers a signi cant bene t over PCR.75,88 Like
PCR, PLS retains the multivariate
signal averaging bene t and the diagnostic capabilities typical of an
over-determined system.
9. CONCLUSIO N
Raman spectroscopy is a powerful
and well-established tool for quantitative analysis. The high information
content and good spectral resolution
typical of Raman spectra often allow
simple analytical models to be accurate and robust. Spectral preprocessing, noise analysis, and multivariate methods extend the quantitative capabilities to more complex
samples. The recent availability of
rugged, turnkey analyzers together
with exible, non-contact sampling
bene ts m ake the future for quantitative R am an an alysis lo ok ver y
promising indeed.
1. G. Placzek, Rayleigh-Streuung und Ram an-Effekt in Handbuch der Radiologie, in Acadeische-Verlag, E. M arx, Ed.
(Leipzig, 1934), vol VI.2, pp. 205. Translation: The Rayleigh and Raman Scattering (University of California Radiation
Laboratory (UCRL), trans 526(L), 1962).
2. D. A. Long, Raman Spectroscopy (McGraw-Hill, New York, 1977).
35A
focal point
3. D. H. Gudehus and D. J. Hegyi, Astronom. J. 90, 130 (1985).
4. B. Schrader, Raman/Infrared Atlas of Organic Compounds (John Wiley and Sons,
New York, 1989), 2nd ed.
5. Sadtler Research Laboratories Ram an
Spectra. A collection of 4400 Raman
spectra, Philadelphia, PA (19731976).
6. D. Lin-Vien, N. B. Colthup, W. G. Fateley, and J. G. Grasselli, The Handbook of
Infrared and Raman Characteristic Frequencies of Organic Molecules (Academ ic Press, Boston, 1991).
7. D. Jouan-Rimbaud, B. Walczak, D. L.
M assart, I. R. Last, and K. A. Prebble,
Anal. Chim. Acta 304, 285 (1995).
8. J. M . Brenchley, U. Horchner, and J. H.
Kalivas, Appl. Spectrosc. 51, 689 (1997).
9. Galactic Industries Corporation, PLSplus/
IQ Users Guide (Galactic, Salem, NH,
1997), p. 85.
10. C. B. Lucasius and G. Kateman, Trends
Anal. Chem. 10, 254 (1991).
11. A. S. Bangalore, R. E. Shaffer, G. W.
Small, and M. A. Arnold, Anal. Chem.
68, 4200 (1996).
12. M . J. McShane, B. D. Cameron, G. L.
Cote, M . M otamedi, and C. H. Spiegelm an, Anal. Chim. Acta 388, 251 (1999).
13. L. Norgaard, A. Saudland, J. Wagner, J.
P. Nielsen, L. Munck, and S. B. Engelsen,
Appl. Spectrosc. 54, 413 (2000).
14. O. Svensson, M. Josefson, and F. W.
Langkilde, Eur. J. Pharm. Sci. 11, 141
(2000).
15. F. T. S. Yu and G. Lu, Appl. Opt. 33, 5262
(1994).
16. B. K. Alsberg, A. M. Woodward, and D.
B. Kell, Chemom. Intell. Lab. Syst. 37,
215 (1997).
17. B. Walczak, W avelets in Chemistry (Elsevier, Amsterdam, 2000).
18. D. Jouan-Rimbaud, B. Walczak, R. J.
Poppi, O. E. de Noord, and D. L. Massart,
Anal. Chem. 69, 4317 (1997).
19. W. Cai, L. Wang, Z. Pan, J. Zuo, C. Xu,
and X. Shao, J. Raman Spectrosc. 32, 207
(2001).
20. H. Takeuchi, S. Hashim oto, and I. Harada, Appl. Spectrosc. 47, 129 (1993).
21. D. Zhang, K. N. Jallad, and D. BenAmotz, Appl. Spectrosc. 55, 1523 (2001).
22. R. Bhargava and I. W. Levin, Anal.
Chem. 74, 1429 (2002).
23. L. Kay and D. A. Sadler, M eas. Sci. Technol. 2, 532 (1991).
24. L. Kay, M eas. Sci. Technol. 3, 400
(1992).
25. G. R. Phillips and J. M. Harris, Anal.
Chem. 62, 2351 (1990).
26. W. Hill and D. Rogalla, Anal. Chem. 64,
2575 (1992).
27. A. W. Moore, Jr., and J. W. Jorgenson,
Anal. Chem. 65, 188 (1993).
28. W. Hill, Appl. Spectrosc. 47, 2171 (1993).
29. J. L. Koenig, Appl. Spectrosc. 29, 293
(1975).
30. M . G. Shim and B. C. Wilson, J. Raman
Spectrosc. 28, 131 (1997).
31. T. J. Vickers, R. E. Wambles, Jr., and C.
36A
32.
33.
34.
35.
36.
37.
38.
39.
40.
41.
42.
43.
44.
45.
46.
47.
48.
49.
50.
51.
52.
53.
54.
55.
56.
57.
58.
59.
K. M ann, Appl. Spectrosc. 55, 389

(2001).
S. M. Haight and D. T. Schwartz, Appl.
Spectrosc. 51, 930 (1997).
D. L. Wilson, K. S. Kump, S. J. Eppell,
and R. E. Marchant, Langmuir 11, 265
(1995).
G. A. Pearson, J. Mag. Res. 27, 265
(1977).
A. P. Shreve, N. J. Cherepy, and R. A.
M athies, Appl. Spectrosc. 46, 707 (1992).
J. J. Baraga, M. S. Feld, and R. P. Rava,
T. Hasegawa, J. Nishijo, and J. Umemu ra,
Chem. Phys. Lett. 317, 642 (2000).
A. T. G. De Paepe, J. M. Dyke, P. J. Hendra, and F. W. Langkilde, Spectrochim.
Acta, Part A 53, 2267 (1997).
A. Brookes, J. M. Dyke, P. J. Hendra, and
A. Strawn, Spectrochim. Acta, Part A 53,
2303 (1997).
K. M. Cunningham, M. C. Goldberg, and
E. R. Weiner, Anal. Chem. 49, 70 (1977).
W. H. Weber, M. Zanini-Fisher, and M. J.
Pelletier, Appl. Spectrosc. 51, 123 (1997).
P. Hendra, C. Jones, and G. Warnes, Fourier Transform Raman Spectroscopy: Instrumentation and Chemical Applications
(Prentice Hall, London, 1991), p. 150.
C. J. Petty, Vib. Spectrosc. 2, 263 (1991).
D. D. Archibald and P. Yager, Appl. Spectrosc. 46, 1613 (1992).
N. Everall, J. Raman Spectrosc. 25, 813
(1994).
M . Buback and H. P. Vogele, FT-NIR Atlas (VCH, Weinheim, 1993).
J. Workman, Jr., Handbook of Organic
Compounds (Academic Press, New York,
2000).
J. D. Womack, T. J. Vickers, and C. K.
M ann, Appl. Spectrosc. 41, 117 (1987).
J. D. Womack, C. K. Mann, and T. J.
Vickers, Appl. Spectrosc. 43, 527 (1989).
W. Gerlach, Z. Anorg. Allg. Chem. 142,
383 (1925).
B. F. Scribner and M. Margoshes, Emission Spectroscopy, in Treatise on Analytical Chemistry, I. M. Kolthoff and P. J.
Elving, Eds. (John Wiley and Sons, New
York, 1965), part I, vol. 6, Chap. 64.
B. Chu, G. Fytas, and G. Zalczer, Macromolecules 14, 395 (1981).
W. M. Sears, J. L. Hunt, and J. R. Stevens, J. Chem. Phys. 75, 1599 (1981).
E. Gulari, K. McKeigue, and K. Y. S. Ng,
M acromolecules 17, 1822 (1984).
O. Svensson, M. Josefson, and F. W.
Langkilde, Chemom. Intell. Lab. Syst. 49,
49 (1999).
N. Everall, K. Davis, H. Owen, M. J. Pelletier, and J. Slater, Appl. Spectrosc. 50,
388 (1996).
I. S. Helland, T. Naes, and T. Isaksson,
Chem om . Intell. Lab. Syst. 29, 233
(1995).
P. Marteau, N. Zanier, A. Aou , G. Hotier,
and F. Cansell, Vib. Spectrosc. 9, 101
(1995).
F. Wang, J. A. Wachter, F. J. Antosz, and
K. A. Berglund, Org. Process Res. Dev.
4, 391 (2000).
60. H. Abderrazak, M . Dachraoui, M. J.

Ayora Canada, and B. Lendl, Appl. Spectrosc. 54, 1610 (2000).
61. F. H. Pollak, Characterization of Sem iconductors by Raman Spectroscopy, in
Analytical Raman Spectroscopy, J. G.
Grasselli and B. J. Bulkin, Eds. (John Wiley and Sons, New York, 1991), pp. 172
174.
62. I. DeWolf, Semiconductors, in Analytical Applications of Raman Spectroscopy,
M . J. Pelletier, Ed. (Blackwell Science,
Oxford, 1999), pp. 463 469.
63. J. D. Dupee, On-line Crystallinity and
Temperature M easureme nts of Nylon-66
Using a Remote Ram an Probe, Dissertation, University of London, London,
UK (1998).
64. K. P. J. Williams and N. J. Everall, J. Ram an Spectrosc. 26, 427 (1995).
65. F. Adar and H. Noether, Polymer 26, 1935
(1985).
66. R. Khazanie, Elementary Statistics in a
W orld of Applications (Addison-Wesley,
New York, 1990), 3rd ed., pp. 376377.
67. J. M . Listy, A. Cholvadova, and J. Kutej,
Comput. Chem. 14, 189 (1990).
68. J. Riu and F. X. Rius, J. Chemom. 9, 343
(1995).
69. H. Bruzzone and C. M oreno, Meas. Sci.
Technol. 9, 2007 (1998).
70. A. Martinez, J. Riu, and F. X. Rius, Chem om. Intell. Lab. Syst. 54, 61 (2000).
71. Y. Melendez, K. F. Schrum, and D. BenAm otz, Appl. Spectrosc. 51, 1176 (1997).
72. W. S. Cleveland and S. J. Devlin, J. Am.
Stat. Assoc. 83, 596 (1988).
73. S. Chang, E. H. Baughman, and B. C.
M cIntosh, Appl. Spectrosc. 55, 1199
(2001).
74. T. Hirschfeld, Computerized Infrared:
The Need for Caution, in Computerized
Quantitative Infrared Analysis, G. L. McClure, Ed. (ASTM , 1984), pp. 169179.
75. K. R. Beebe, R. J. Pell, and M. B. Seasholtz, Chemometrics: A Practical Guide
(John Wiley and Sons, New York, 1998).
76. R. Kramer, Chemometric Techniques for
Quantitative Analysis (M arcel Dekker,
New York, 1998).
77. H. Martens and T. Naes, Multivariate Calibration (John Wiley and Sons, New
York, 1989).
78. J. A. Kleimeyer, P. E. Rose, and J. M.
Harris, Appl. Spectrosc. 55, 690 (2001).
79. C.-N. Ho, G. D. Christian, and E. R. Davidson, Anal. Chem. 50, 1108 (1978).
80. C.-N. Ho, G. D. Christian, and E. R. Davidson, Anal. Chem. 53, 92 (1981).
81. E. Sanchez and B. R. Kowalski, Anal.
Chem. 58, 496 (1986).
82. P. C. Gillette and J. L. Koenig, Appl.
Spectrosc. 38, 334 (1984).
83. S. Banerjee and D. Li, Appl. Spectrosc.
45, 1047 (1991).
84. T. Hirschfeld, Anal. Chem . 48, 721
(1976).
85. G. Horlick, Anal. Chem. 45, 319 (1973).
86. C. K. M ann, J. R. Goleniewski, and C. A.
Sism anidis, Appl. Spectrosc. 36, 223
(1982).
87. B. J. Pelletier, p calculation program

(California Institute of Technology).
88. E. V. Thomas and D. M. Haaland, Anal.
Chem. 62, 1091 (1990).
MSC
SNV
FRACT
APPENDIX
Examples of Quantitative Analysis Using Raman Spectroscopy.
Raman spectroscopy has been used
for quantitative analysis for several
decades. The table in this appendix
and its references summarize over
20 0 exam ples from th e peer-review ed technical literature. N um bered references in the table refer to
the reference list contained in this
appendix, not the reference list for
the body of the Focal Point article.
Descriptive information for the table
colum n headings is given below.
ID (classi cation used to group similar applications)
11 Polymer, polybutadiene
12 Polymer, polystyrene
13 Polymer, epoxy
14 Polymer, poly(methylmethacr ylate)
15 Polymer, density and crystallinity
16 Polymer, miscellaneous
17 Polymer, co-polymers
21 Hydrocarbons, C 8 separation
22 Hydrocarbons, xylenes
23 Hydrocarbons, fuel composition and
classi cation
24 Hydrocarbons, fuel properties
25 Hydrocarbons, miscellaneous
30 Glucose, biological applications
40 Fats and oils
50 Pharmaceutical applications
60 Gas-phase applications
80 Inorganic applications
Metric (spectral property related to analyte
concentration)
H
Band height
A
Band area
PCA
Principal components analysis
PLS
Partial least squares
SHIFT Change in band position
PCR
Principal components regression
CORR
Cross-correlation
Norm (method used to normalize intensity
of Raman spectrum)
NONE
Absolute intensity, no normalization
IN STD
Internal standard already part
of sample
1 IN STD
Internal standard added to
sample
S SPEC
Area of whole Raman spectrum set equal to 1
S SEG
Area of spectral segm ent set
equal to 1
w S BAND Weighted sum of Raman
bands
EX STD
RATIO
OTHER
Multiplicative scatter correction

Standard normal variate
transformation
Fraction of analyte band remaining
External standard
Simple ratio of concentrations or mole ratio
Other methods of normalization
Ref (Literature referen ces for applications)

1. K. P. J. Williams and S. M. Mason,
Spectrochim. Acta, Part A 46, 187
(1990).
2. K. D. O. Jackson, M . J. R. Loadman, C.
H. Jones, and G. Ellis, Spectrochim.
Acta, Part A 46, 217 (1990).
3. D. D. Tunnicliff and A. C. Jones, Spectrochim. Acta 18, 579 (1962).
4. J. B. Cooper, K. L. Wise, J. Groves, and
W. T. Welch, Anal. Chem. 67, 4096
(1995).
5. K. P. J. Williams, R. E. Aries, D. J. Cutler, and D. P. Lidiard, Anal. Chem. 62,
2553 (1990).
6. J. J. Heigl, J. F. Black, and B. F. Dudenbostel, Jr., Anal. Chem. 21, 554 (1994).
7. G. R. Strobl and W. Hagedorn, J. Polym.
Sci. 16, 1181 (1978).
8. K. W. J. Williams and N. J. Everall, J.
Raman Spectrosc. 26, 427 (1995).
9. W. A. England, S. N. Jenny, and D. A.
Greenhalgh, J. Raman Spectrosc. 15,
156 (1984).
10. S. W. Cornell and J. L. Koenig, Macrom olecules 2, 540 (1969).
11. E. Gulari, K. McKeigue, and K. Y. S.
Ng, Macromolecules 17, 1822 (1984).
12. H. J. Sloane and R. Bram mston-Cook,
13. A. Bader and J. R. Wunsch, Macrom olecules 28, 3794 (1955).
14. C. E. Miller, D. D. Archibald, M. L.
M yrick, and S. M. Angel, Appl. Spectrosc. 44, 1297 (1990).
15. A. C. R. Pipino, Appl. Spectrosc. 44,
1163 (1990).
16. E. N. Lewis, V. F. Kalasinsky, and I. W.
Levin, Anal. Chem. 60, 2309 (1988).
17. H. G. M. Edwards, A. F. Johnson, I. R.
Lewis, and S. J. Wheelwright, Spectrochim. Acta, Part A 49, 457 (1993).
18. M. L. Scheepers, J. M. Gelan, R. A. Carleer, P. J. Adriaensens, D. J. Vanderzande, B. J. Kip, and P. M. Brandts, Vib.
Spectrosc. 6, 55 (1993).
19. O. Florisson and P. H. Burrie, J. Phys.
E: Sci. Instrum. 22, 123 (1989).
20. M. B. Seasholtz, D. D. Archibald, A.
Lorber, and B. R. Kowalski, Appl. Spectrosc. 43, 1067 (1989).
21. X. Dou, Y. Yamaguchi, H. Yamamoto,
H. Uenoyama, and Y. Ozaki, Appl.
Spectrosc. 50, 1301 (1996).
22. X. Dou and Y. Ozaki, Appl. Spectrosc.
52, 815 (1998).
23. R. C. Harney, US Patent 4,068,953
(1978).
24. R. H. Clarke, US Patent 5,139,334
(1992).
25. D. A. Leonard, US Patent 3,723,007

(1973).
26. G. R. Abel and C. E. Gillespie, US Patent 3,625,613 (1971).
27. T. L. Williams, R. B. Martin, and T. W.
Collette, Appl. Spectrosc. 55, 967
(2001).
28. C. J. de Bakker and P. M. Fredericks,
Appl. Spectrosc. 49, 1766 (1995).
29. C. G. Kontoyannis, N. Ch. Bouropoulos
and P. G. Koutsoukos, Appl. Spectrosc.
51, 64 (1997).
30. J. F. Aust, K. S. Booksh, and M. L. Myrick, Appl. Spectrosc. 50, 382 (1996).
31. J. B. Cooper, P. E. Flecher, T. M. Vess,
and W. T. Welch, Appl. Spectrosc. 49,
586 (1995).
32. C. M. Stellman, J. F. Aust, and M . L.
M yrick, Appl. Spectrosc. 49, 392
(1995).
33. M. Wunderling, B. Fischer, and S. Kastle, European Patent Application EP 0
947 825 A1 (1999).
34. D. Bruck and U. Wolf, European Patent
Application EP 0 897 933 (1999).
35. J. B. Cooper, R. R. Bledsoe, Jr., K. L.
Wise, M . B. Sumner, W. T. Welch, and
B. K. Wilt, US Patent 5,892,228 (1999).
36. K. Sano, M . Shimoyama, M. Ohgane, H.
Higashiyam a, M . Watari, M . Tomo, T.
Ninomiya, and Y. Ozak i, Appl. Spectrosc. 53, 551 (1999).
37. R. H. Clarke, S. Londhe, W. R. Prem asiri, and M. E. Womble, J. Raman Spectrosc. 30, 827 (199).
38. T. M. Niemczyk, M . M . Delgado-Lopex,
and F. S. Allen, Anal. Chem. 70, 2762
(1998).
39. T. H. King, C. K. M ann, and T. J. Vickers, J. Pharm. Sci. 74, 443 (1985).
40. T. M. Niemczyk, M. Delgado-Lopex, F.
S. Allen, J. T. Clay, and D. L. Arneberg,
41. C. M. Deeley, R. A. Spragg, and T. L.
Threlfall, Spectrochim. Acta, Part A 47,
1217 (1991).
42. L. P. Powell and A. Campion, Anal.
Chem. 58, 2350 (1986).
43. J. D. Houston, S. Sizgoric, A. Ulitsky,
and J. Banic, Appl. Opt. 25, 2115
(1986).
44. R. P. Schnell, R. W. Sampson, R. F. Pacanowski, and D. J. Bruggema, US Patent 4,620,284 (1986).
45. L. M. Smith, G. Osaki, and T. Peters,
Adv. Instrum. 45, 711 (1990).
46. S. G. Skoulika, C. A. Georgiou, and M.
G. Polissiou, Appl. Spectrosc. 53, 1470
(1999).
47. N. A. Arley, C. K. Mann, and T. J. Vickers, Appl. Spectrosc. 38, 540 (1984).
48. M. Shimoyama, H. Maeda, H. Sato, T.
Ninomyia, and Y. Ozak i, Appl. Spectrosc. 51, 1154 (1997).
49. P. J. Watts, A. Tudor, S. J. Church, P. J.
Hendra, P. Turner, C. D. Melia, and M.
C. Davies, Pharm. Res. 8, 1323 (1991).
50. J. R. Walton and K. P. J. Williams, Vib.
Spectrosc. 1, 339 (1991).
51. G. Ellis, A. Sanchez, P. J. Hendra, H. A.
Willis, J. M. Chalmers, J. G. Eaves, W.
37A
focal point
52.
53.
54.
55.
56.
57.
58.
59.
60.
61.
62.
63.
64.
65.
66.
67.
68.
69.
70.
71.
72.
73.
38A
F. Gaskin, and K.-N. Kruger, J. Mol.

Struct. 247, 385 (1991).
A. S. Gilbert, K. W. Hobbs, A. H.
Reeves, and P. P. Jobson, Proc. SPIE-Int.
Soc. Opt. Eng. 2248, 391 (1994).
G. Ellis, M. Claybourn, and S. E. Richards, Spectrochim. Acta, Part A 46, 227
(1990).
J. K. Agbenyega, G. Ellis, P. J. Hendra,
W. F. Maddams, C. Passingham, H. A.
Willis, and J. Chalmers, Spectrochim.
Acta, Part A 46, 197 (1990).
R. E. Benner, J. D. Andrade, R. A. Van
Wagenen, D. R. Westenskow, US Patent
4,648,714 (1987).
R. A. Van Wagenen, J. D. Geisler, D. E.
Gregonis, D. L. Coleman, US Patent
4,784,486 (1988).
D. R. Westenskow, K. W. Smith, D. L.
Coleman, D. E. Gregonis, and R. A. Van
Wagenen, Anesthesiology 70, 350
(1989).
C. J. de Bakker, G. A. George, N. A. St
John, and P. M. Fredericks, Spectrochim.
Acta, Part A 49, 739 (1993).
H. G. M. Edwards, A. F. Johnson, I. R.
Lewis, and M. A. Macleo d, Polymer 34,
3184 (1993).
H. Sadeghi-Jorabchi, R. H. Wilson, P. S.
Belton, J. D. Edwards-Webb, and D. T.
Coxon, Spectrochim. Acta, Part A 47,
1449 (1991).
J. Clarkson, S. M . Mason, and K. P. J.
Williams, Spectrochim. Acta, Part A 47,
1345 (1991).
J. K. Agbenyega, M. Claybourn, and G.
Ellis, Spectrochim. Acta, Part A 47,
1375 (1991).
J. Y. Qu, B. C. Wilson, and D. Suria,
Appl. Opt. 38, 5491 (1999).
J. F. van Staden, M . A. Makhafola, and
D. de Waal, Appl. Spectrosc. 50, 991
(1996).
T. M. Niemczyk, M . M . Delgado-Lopex,
and F. S. Allen, Anal. Chem. 70, 2762
(1998).
Y. Ozaki, R. Cho, K. Ikegaya, S. Muraishi, and K. Kawau chi, Appl. Spectrosc.
46, 1530 (1992).
S. Damo un, R. Papin, G. Ripault, M.
Rousseau, J. C. Rabadeux, and D. Durand, J. Ram an Spectrosc. 23, 385
(1992).
M. Kam ath, W. H. Kim, L. Li, J. Kumar,
S. Tripathy, K. N. Babu, and S. S. Talward, Macromolecules 26, 5954 (1993).
T. L. Rapp, W. K. Kowalchyk, K. L. Davis, E. A. Todd, K. Liu, and M. D. Morris, Anal. Chem. 64, 2434 (1992).
S. M. Adler-Golden, N. Goldstein, F.
Bien, M. W. Matthew, M. E. Gersh, W.
K. Cheng, and F. W. Adams, Appl. Opt.
31, 831 (1992).
K. E. Chike, M. L. Myrick, R. E. Lyon,
and S. M. Angel, Appl. Spectrosc. 47,
1631 (1993).
H. G. M. Edwards, A. F. Johnson, I. R.
Lewis, N. J. Ward, and S. Wheelwright,
J. Raman Spectrosc. 24, 495 (1993).
J. J. Freeman, D. O. Fisher, and G. J.
74.
75.
76.
77.
78.
79.
80.
81.
82.
83.
84.
85.
86.
87.
88.
89.
90.
91.
92.
93.
94.
95.
96.
97.
98.
99.
Gervasio, Appl. Spectrosc. 47, 1115

(1993).
H. Wag, C. K. M ann, and T. J. Vickers,
E. J. Hutchinson, D. Shu, F. LaPlant, and
D. Ben-Am otz, Appl. Spectrosc. 49,
1275 (1995).
H. G. M. Edwards, J. M. C. Turner, and
V. Fawcett, J. Chem. Soc., Faraday
Trans. 91, 1439 (1995).
H. G. M . Edwards and V. Fawcett, J.
M ol. Struct. 326, 131 (1994).
N. Wen and M. H. Brooker, Can. J.
Chem. 72, 1099 (1994).
J. F. Aust, M. K. Higgins, P. Groner, S.
L. Morgan, and M. L. Myrick, Anal.
Chim. Acta 293, 119 (1994).
C. Tseng, C. K. M ann, and T. J. Vickers,
S. Y. Wang, C. E. Hasty, P. A. Watson,
J. P. Wicksted, R. D. Stith, and W. F.
M arch, Appl. Opt. 32, 925 (1993).
J. F. Ford, C. K. Mann, and T. J. Vickers,
N. J. Everall, J. M. Chalmers, R. Ferwerda, J. H. van der Maas, and P. J. Hendra, J. Raman Spectrosc. 25, 43 (1994).
E. Gantner, M . Beck, H. G. Muller, D.
Steinert, and H. J. Ache, J. Raman Spectrosc. 25, 31 (1994).
S. Poshyachinda and V. Kanitthanon,
(1994).
P. E. Flecher, W. T. Welch, S. Albin, and
J. B. Cooper, Spectrochim. Acta, Part A
53, 199 (1997).
X. Dou, Y. Yamaguchi, H. Yamamoto,
S. Doi, and Y. Ozak i, Vib. Spectrosc. 14,
199 (1997).
C. G. Kontoyannis, N. C. Bouropoulos,
and P. G. Koutsoukos, Vib. Spectrosc.
15, 53 (1997).
A. Brookes, J. M. Dyke, P. J. Hendra,
and A. Strawn, Spectrochim. Acta, Part
A 53, 2303 (1997).
T. Ozpozan, B. Schrader, and S. Keller,
Spectrochim. Acta, Part A 53, 1 (1997).
B. F. Dudenbostel, Jr., and N. J. Linden,
US Patent 2,527,121 (1950).
H. S. Carm an, Jr., D. C. Alsmey er, C. H.
Juarex-Garcia, A. W. Garrett, B. E. Wilson, and V. A. Nicely, US Patent
5,850,623 (1998).
D. C. Alsm eyer, M . J. Pearce, and V. A.
Nicely, World Patent W O 98/08066
(1998).
Y. Yam aguchi, H. Uenoyaa, M. Yagi,
and D. Xiaoming, US Patent 5,753,449
(1998).
H. Yamam oto, H. Uenoyama, X. Dou,
Y. X. Wang, and K. Shimada, US Patent
5,754,288 (1998).
J. B. Cooper, P. E. Flecher, Jr., and W. T.
Welch, US Patent 5,684,580 (1997).
J. B. Cooper, K. L. Wise, W. T. Welch,
and M. B. Sumner, US Patent 5,596,196
(1997).
D. C. Alsm eyer, M . J. Pearce, and V. A.
Nicely, US Patent 5,652,653 (1997).
S. Sollinger and M. Diamantoglou, R.
Raman Spectrosc. 28, 811 (1997).
100. G. Kister, G. Cassanas, and M. Vert,

Polymer 39, 335 (1998).
101. H. Yamamoto , H. Uenoyama, K. Kirai,
X. Dou, and Y. Ozaki, Appl. Opt. 37,
2640 (1998).
102. A. Al-khansashi, M. Dhamdhere, and M.
Hansen, Appl. Spectrosc. Rev. 33, 115
(1998).
103. A. Geze, I. Chourpa, F. Boury, J. Benoit,
and P. Dubois, Analyst (Cam bridge,
U.K.) 124, 37 (1999).
104. H. Swierenga, A. P. De Weijer, and L.
M . C. Buydens, J. Chemom. 13, 237
(1999).
105. H. Sadeghi-Jorabchi, P. J. Hendra, R. H.
Wilson, and P. S. Belton, J. Am. Oil
Chem. Soc. 67, 483 (1990).
106. B. Sandner, S. Kam mer, and S. Wartewig, Polymer 37, 4705 (1996).
107. M. Shimoyama, H. M aeda, K. Matsukawa, H. Inoue, T. Ninomyia, and Y.
Ozaki, Vib. Spectrosc. 14, 253 (1997).
108. X. Dou, Y. Yamaguchi, H. Yamamoto ,
S. Doi, and Y. Ozaki, Vib. Spectrosc. 13,
83 (1996).
109. V. Vacque, N. Dupuy, B. Sombret, J. P.
Huvenne, and P. Legrand, J. Mol. Struct.
410, 555 (1997).
110. A. J. Berger, T. Koo, I. Itzkan, G. Horowitz, and M. S. Feld, Appl. Opt. 38,
2916 (1999).
111. A. J. Berger, Y. Wang, and M. S. Feld,
Appl. Opt. 35, 209 (1996).
112. A. J. Berger, I. Itzkan, and M. S. Feld,
(1997).
113. A. J. Berger, Y. Wang, D. M. Sammeth ,
I. Itzkan, K. Kneipp, and M. S. Feld,
Appl. Spectrosc. 49, 1164 (1995).
114. B. Chu, G. Fytas, and G. Zalczer, Macromolecules 14, 395 (1981).
115. D. H. Rank, R. W. Scott, and M. R. Fenske, Anal. Chem. 14, 816 (1942).*
116. E. J. Rosenbaum , C. C. Martin, and J.
L. Lauer, Anal. Chem. 18, 731 (1946).*
117. M. R. Fenske, W. G. Braun, R. V. Wiegand, D. Quiggle, R. H. M cCormick,
and D. H. Rank, Anal. Chem. 19, 700
(1947).*
118. R. F. Stamm, Anal. Chem. 17, 318
(1945).
119. R. Szostak and S. Mazurek, Analyst
(Cambridge, U.K.) 127, 144 (2002).
120. S. Auguste, H. G. M. Edwards, A. F.
Johnson, Z. G. Meszene, and P. Nicol,
Polymer 37, 3665 (1996).
121. E. W. Nelson and A. B. Scranton, J. Polym. Sci., Part A: Polym. Chem. 34, 403
(1996).
122. D. N. Phillips, T. M . Suckling, and W.
van Bronswijk, Talanta 43, 221 (1996).
123. C. G. Kontoyannis, M . G. Orkoula, and
P. G. Koutsoukos, Analyst (Cambridge,
U.K.) 122, 33 (1997).
124. S. G. Skoulika, C. A. Georgiou, and M.
G. Polissiou, Talanta 51, 599 (2000).
125. O. Svensson, M. Josefson, and F. W.
Langkilde, Eur. J. Pharm. Sci. 11, 141
(2000).
126. O. Svensson, J. Josefson, and F. W.
127.
128.
129.
130.
131.
132.
133.
134.
135.
136.
137.
138.
139.
140.
141.
Langkilde, Chemom. Intell. Lab. Syst.

49, 49 (1999).
F. W. Langkilde, J. Sjoblom, L. Tekenbergs-Hjelte, and J. Mrak, J. Pharm. Bio.
Anal. 15, 687 (1997).
C. G. Kontoyannis, J. Pharm. Biomed.
Anal. 13, 73 (1995).
M. Johnck, L. M uller, A. Neyer, J. W.
Hofstraat, Polymer 40, 3631 (1999).
C. G. Kontoyannis, M. Orkoula, Talanta
41, 1981 (1994).
T. M. Hancewicz and C. Pety, Spectrochim. Acta, Part A 51, 2193 (1995).
P. Marteau, N. Zanier-Szydlowski, A.
Aou , G. Hotier, and F. Cansell, Vib.
Spectrosc. 9, 101 (1995).
F. Estienne, D. L. Massart, N. ZanierSzydlowski, and P. M arteau , Anal.
Chim. Acta 424, 185 (2000).
P. M arteau and N. Zanier, Spectroscopy
10, 26 (1995).
P. Marteau, A. Aou , and N. Zanier,
An Application of Raman Spectroscopy in the Petroleum Industry, in Recent
Developments in Scienti c Imaging, M.
Bonner Denton, R. E. Fields, and Q. S.
Hanley, Eds. (The Royal Society of
Chemistry, Cam bridge, U.K., 1996).
A. G. M iller and J. W. Macklin, Anal.
Chem. 55, 684 (1983).
W. Cai, L. Wang, Z. Pan, J. Zuo, C. Xu,
and X. Shao, J. Ram an Spectrosc. 32,
207 (2001).
T. Jawhari, P. J. Hendra, H. A. Willis,
and M. Judkins, Spectrochim. Acta, Part
A 46, 161 (1990).
T. Kaiser, C. Vosmerbaumer, and G.
Schweiger, Ber. Bunsen-Ges. Phys.
Chem. 96, 976 (1992).
W. M. Sears and J. L. Hunt, J. Chem.
Phys. 74, 1599 (1981).
F. Dywan, B. Hartmann, S. Klauer, M.
D. Lechner, R. A. Rupp, and M . Wohlecke, M akrom ol. Chem . 194, 1527
(1993).
142. M. J. Pelletier, K. L. Davis, and R. A.

Carpio, Electrochem. Soc. Proc. 95(2),
282 (1995).
143. J. A. Frankland, H. G. M. Edwards, A.
F. Johnson, I. R. Lewis, and S. Poshyachinda, Spectrochim. Acta, Part A 47,
1511 (1991).
144. J. W. Schoppelrei, M. L. Kieke, and T.
B. Brill, J. Phys. Chem. 100, 7463
(1996).
145. K. H. Fung and I. N. Tang, J. Colloid
Interface Sci. 130, 219 (1989).
146. D. E. Irish and J. D. Riddell, Appl. Spectrosc. 28, 481 (1974).
147. D. E. Irish and H. Chen, Appl. Spectrosc. 25, 1 (1971).
148. E. Gulari, K. McKeigue, and K. Y. S.
Ng, Macromolecules 17, 1822 (1984).
149. H. Abderrazak, M. Dachraoui, M . J. A.
Canada, and B. Lendl, Appl. Spectrosc.
54, 1610 (2000).
150. B. Wopenka and J. D. Pasteris, Appl.
Spectrosc. 40, 144 (1986).
151. D. P. Schweinsberg and Y. D. West,
Spectrochim. Acta, Part A 53, 25 (1997).
152. D. R. Lombardi, C. K. Mann, and T. J.
Vickers, Appl. Spectrosc. 49, 220
(1995).
153. L. Black, G. C. Allen, and P. C. Frost,
Appl. Spectrosc. 49, 1299 (1995).
154. I. M. Clegg, N. J. Everall, B. King, H.
M elvin, and C. Norton, Appl. Spectrosc.
55, 1138 (2001).
155. N. Wen and M . H. Brooker, J. Phys.
Chem. 99, 359 (1995).
156. M. L. Scheepers, R. J. M eier, L. Markwort, J. M . Gelan, D. J. Vanderzande,
and B. J. Kip, Vib. Spectrosc. 9, 139
(1995).
157. C. G. Kontoyannis, N. Ch. Bouropoulos,
and P. G. Koutsoukos, Analyst (Cam bridge, U.K.) 120, 347 (1995).
158. E. van Overbeke, J. Devaux, R. Legras,
J. T. Carter, P. T. M cGrail, and V. Carlier,
Appl. Spectrosc. 55, 1514 (2001).
159. J. D. Womack, T. J. Vickers, and C. K.

160. T. B. Shope, T. J. Vickers, and C. K.
161. B. De Clercq, T. Smellinckx, C. Hugelier, N. Maes, and F. Verpoort, Appl.
Spectrosc. 55, 1564 (2001).
162. X. Zheng, W. Fu, S. Albin, K. L. Wise,
A. Javey, and J. B. Cooper, Appl. Spectrosc. 55, 382 (2001).
163. S. G. Skoulika and C. A. Georgiou,
164. P. L. Wancheck and L. E. Wolfram,
165. S. G. Skoulika and C. A. Georgiou,
Appl. Spectrosc. 55, 1259 (2001).
166. J. P. Wicksted, R. J. Erckens, M. Motam edi, and W. F. March, Appl. Spectrosc.
49, 987 (1995).
167. J. Y. Qu and L. Shao, Appl. Spectrosc.
55, 1414 (2001).
168. E. D. Lipp and R. L. Grosse, Appl.
Spectrosc. 52, 42 (1998).
169. D. D. Archibald, S. E. Kays, D. S. Him m elsbach, and F. E. Barton II, Appl.
Spectrosc. 52, 22 (1998).
170. Y. Melendez, K. F. Schrum, and D. BenAm otz, Appl. Spectrosc. 51, 1176
(1997).
171. E. Gantner and D. Steinert, Fresenius J.
Anal. Chem. 338, 2 (1990).
172. C. Wang, T. J. Vickers, and C. K. Mann,
173. J. Stokr, B. Schneider, A. Frydrychova,
and J. Coupek, J. Appl. Polym. Sci. 23,
3553 (1979).
174. B. Chu and D. Lee, Macromolecules 17,
926 (1984).
175. M. van Eck, J. P. J. van Dalen, and L.
de Galan, Analyst (Cambridge, U.K.)
108, 485 (1983).
* The journal Analytical Chemistry was actually called Industrial and Engineering
Chemistry, Analytical Section at this time.
39A
focal point
Application
Polybutadiene structural content
Polybutadiene microstructure
Butadiene microstructure
Polybutadiene structural content
C5C during polybutadiene hydrogenation
Butadiene polymerization
Hydrogenated polybutadiene microstructure
Styrene polymerization
Styrene polymerization in benzene
Styrene emulsion polymerization
Epoxy cure
Epoxy cure
Epoxy cure
Epoxy cure
Epoxy cure
Epoxy cure
Methyl methacrylate polymerization
Methyl methacrylate emulsion polymerization
Poly(methylmethacrylate) residual monomer
Emulsion polymerization of methylethacrylate
Polyethylene cr ystallinity (density)
Poly(aryl ether ether ketone) cr ystallinity
Poly(aryl ether ether ketone) cr ystallinity
Polyethylene density (cr ystallinity)
PET yarn shrinkage
Polyethylene density (cr ystallinity)
Polyurethane ex modulus
Butylacr ylate polymerization
Acr ylamide polymerization kinetics
Diacetylene polymerization kinetics
Melam ine, melam ine cyanurate in nylon
Polyimide curing
Polyimide curing
trans-Polyactylene synthesis
Photopolymerization of divinyl ethers
Poly(aryl ether sulphone) sulfonation
Poly(ethylene glycol) in chloroform
Polyester synthesis
Vinyl acetate emulsion polymerization
Vinyl acetate emulsion polymerization
Photocuring of halogenated acr ylates
Norbornene polymerization
Styrene, butadiene, methylethacrylate terpolymer composition
Styrene monomer in styrene/butadiene
Styrene-divinylbenzene C5C content
Styrene-divinylbenzene polymer content
C5C in styrene-ethylene dimethacr ylate
Nitrile in nitrile-butadiene rubber
PES/PEES co-polymer composition
Emulsion polymerization (3 monomers)
Sulphone co-polymer composition
Co-polymer butadiene M MA C5C type
Monomer in melamine-formaldehyde
Epoxy-methacrylate copolymerization
Styrene-butylacrylate polymerization
Lactide-glycolide copolymer composition
Styrene-ethylene dimethacr ylate copolymer content
Styrene-butadiene polymerization
Content of epoxy-thermoplastic blends
Content of epoxy-thermoplastic blends
Styrene-ethylene dimethacr ylate copolymer content
Melam ine-formaldehyde copolymer composition
40A
ID
Metric
11
11
11
11
11
11
11
12
12
12
12
12
12
12
13
13
13
13
13
13
14
14
14
14
14
14
14
15
15
15
15
15
15
16
16
16
16
16
16
16
16
16
16
16
16
16
16
16
16
17
17
17
17
17
17
17
17
17
17
17
17
17
17
17
17
17
17
17
17
H
A
A
A
A
A
A
A
H
A
H
A
H
A
H
?
A
H
PCA
PCA
A
H
H
A
CLS
A
H
A
PLS
SHIFT
PLS
PLS
PLS
PLS
H
A
H
CLS
H
PCA
A
H
A
A
PLS
H
H
H
H
H
H
H
H
H
A
H
H
H
A
A
A
H
A
H
A
PLS
PLS
H
A
Norm
S SEG
S SEG
S SEG
S SEG
IN STD
IN STD
S SEG
IN STD
IN STD
NO NE
NO NE
IN STD
FRACT
IN STD
IN STD
FRACT
FRACT
IN STD
IN STD
S SPEC
NO NE
IN STD
NO NE
IN STD
IN STD
IN STD
NO NE
IN STD
IN STD
NO NE
S SPEC
NO NE
MSC
IN STD
NO NE
NO NE
IN STD
IN STD
IN STD
IN STD
IN STD
IN STD
IN STD
IN STD
IN STD
FRACT
FRACT
IN STD
IN STD
S SEG
IN STD
IN STD
RATIO
IN STD
IN STD
RATIO
RATIO
RATIO
S SEG
IN STD
IN STD
IN STD
RATIO
RATIO
IN STD
IN STD
S SPEC
RATIO
RATIO
Ref
10
143
59
17
85
120
85
114
140
11
1
120
89
37
54
50
58
71
32
30
148
174
1
67
172
141
102
7
83
83
8
104
36
14
61
69
68
80
79
79
13
121
99
170
98
89
90
129
161
12
164
173
173
173
2
54
53
51
72
156
106
102
100
173
37
158
158
173
18
Application
ID
Metric
Norm
Ref
Polybutadiene methyl methacr ylate copolymer structural cont

Ethylene/vinyl acetate copolymer composition
Ethylene/vinyl acetate copolymer composition
Acr ylonitrile-butadiene rubber hydrogenation
Industrial C 8 aromatic separation
C 8 aromatic m ixtures
C 8 aromatic m ixtures
Xylene isomer content
Mixtures of meta and para xylene
Xylenes in hydrocarbon mixtures
Isomer composition of xylenes
Isomer composition of xylene
Isomer composition of xylene
Hydrocarbon mixtures
Total ole ns, aromatics in mixtures
Aromatic content of hydrocarbons
PNA and PONA hydrocarbon analysis
PNA and PONA hydrocarbon analysis
Fuel mixture composition
Aromatics in hydrocarbons
Cetane index of gas oil
Gasoline octane number
Fuel properties
Octane number in petroleum fuels
Octane number in gasoline
Octane number in petroleum fuels
Benzene and toluene in acetonitrile
Remote analysis for methane gas
Cyclohexanetoluene mixture composition
Cyclohexane in toluene
Oxygenate in gasoline
Glucose, other metabolites in water
Glucose and metabolites in urine
Glucose, other analytes in saline solution
Glucose in plasma and serum
Glucose and acetone in urine
Glucose and bicarbonate in whole blood
Glucose in aqueous solution
Calcium oxalate forms in urinary stones
Glucose and more in blood, serum
Glucose and more in blood, serum
Glucose, lactate, urea in aqueous solutions and aqueous humor
Glucose, acetam inophen, urea, ethanol in simulated blood serum
Total unsaturation in oils and margarines
cis-Isomer content in fats and oils
Unsaturation in fats and oils
Fat unsaturation in food
Active ingredients in pharmaceuticals
Active ingredients in pharmaceuticals
Impurities in Polyene Antibiotics
Drug in polymer m atrix
Polymorphic forms of drugs
Salicylic acid acetate in aspirin
Carbonate, glycine in antacid tablets
Vitam in A in sorbitan mono-oleate base
Drug polymorphic composition
Bucindolol in gel capsules
Bucindolol in gel capsules
Pharmaceutical active in polymer matrix
Metoprolol synthesis
Cipro oxacin in pharmaceuticals
Cipro oxacin in pharmaceuticals
17
17
17
17
21
21
21
21
21
22
22
22
22
22
22
22
22
23
23
23
23
23
23
23
23
23
23
23
24
24
24
24
24
24
25
25
25
25
25
30
30
30
30
30
30
30
30
30
30
30
30
40
40
40
40
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
A
PLS
PLS
H
A
A
A
A
PLS
H
A
PLS
PLS
PLS
PLS
PLS
PLS
H
H
H
A
A
MLR
A
MLR
PLS
PCR
PLS
PCR
A
PLS
PLS
PLS
PLS
H
A
H
A
PLS
A
H
PLS
H
A
PLS
A
H
PLS
PLS
A
PLS
A
H
A
H
H
CORR
A
A
PCR
H
H
PLS
H
PLS
PLS
A
PLS
A, H
MLR
S SEG
MSC
S SPEC
IN STD
w S BAND
w S BAND
w S BAND
w S BAND
NO NE
IN STD
RATIO
NO NE
OTHER
OTHER
NO NE
NO NE
NO NE
1 IN STD
EX STD
EX STD
EX STD
EX STD
NO NE
NO NE
NO NE
NO NE
NO NE
NO NE
NO NE
RATIO
OTHER
OTHER
NO NE
OTHER
IN STD
IN STD
RATIO
EX STD
NO NE
IN STD
IN STD
NO NE
NO NE
IN STD
NO NE
OTHER
RATIO
NO NE
NO NE
IN STD
IN STD
IN STD
IN STD
IN STD
IN STD
1 IN STD
1 IN STD
RATIO
RATIO
S SEG
RATIO
RATIO
MSC
RATIO
NO NE
NO NE
IN STD
SNV
EX STD
EX STD
17
107
107
34
134
132
135
133
133
118
138
31
31
31
98
92
93
115
116
117
6
91
44
44
20
20
20
96
5
24
28
4
86
35
3
43
139
151
97
81
108
111
21
87
112
22
29
110
63
166
167
105
60
60
66
39
39
16
49
41
130
128
131
127
38
40
103
125
165
165
41A
focal point
Application
Pharmaceutical active in tablets
Pharmaceutical active in tablets
Remote gas analysis
Remote gas analysis
Remote gas analysis
Gas phase isotope ratio
Headspace in sealed drug vials
Multiple gas analysis
Wobbe index of natural gas
Multiple component gas analysis
N 2 and O 2 in double-pane windows
Hydrogen trace level analysis
Oxygen gas in nitrogen
Components in exhaled air
Metabolic gases
Xenon in anaesthetic
Nitrite, nitrate in water
Nitrate impurity in nitrite
Sulfate concentration
Sulfate in sodium nitrate
PO 4 2 3 and HPO 4 2 equilibria
Chromium oxide lm thickness
Sulfate, nitrate in single aerosol particles
Aqueous uranium solutions
Carbonate, bicarbonate, carbamate in water
Hydroxyapatite, calcium oxalate monohydrate m ixtures
PCl3 reactor material
Nitronium ion in acid
Phosphite and hypophosphite in water
Atmospheric corrosion products of lead
Carbon dioxide dissolved in water
Metals in cation exchange resins
Nitronium ion in acid and water
SC-1 cleaning bath composition
Water in sodium nitrate
(NH 4 ) 2 CO 3 aqueous equilibria
Nitrite decomposition kinetics
Silicon in silica dust
Gypsum in sulfated m arble
Nitrile oxidation by peroxides
Chlorosilane process streams
Hydrogen peroxide in water
Phosphate in conc. phosphoric acid
Perchlorate ion in aqueous extracts
Perchlorate, nitrate in aqueous extracts
Titanium dioxide, anatase in rutile
Autooxidation of unsaturated fatty acid ethyl esters
Cyanate hydrolysis kinetics
Isomerization of Quadricyclane to norbornadiene
CH 4 and CO 2 in quartz inclusions
Azo dyes in methanol
Butyl acr ylate in solvent
Urea in aqueous solution
Phenols in water
Phenols in water
Film thickness
Speci c gravity of Ivories
Dietary ber in cereal foods
Dietary ber in cereal foods
Esteri cation of ethylene glycol and dimethylterephthalic acid
Ethyl acetate synthesis and hydrolysis
Ethyl acetate synthesis and hydrolysis
Ethanol, methanol, acetone in water
Fenthion in pesticide formulations
Adipic acid, (NH 4 ) 2 SO 4 , and NaSO 4 in caprolactam
Diazinon in pesticide formulations
Methyl-parathion in pesticide formulations
Methyl-parathion in pesticide formulations
CCl4 in ethanol
Ethanol/water solution
42A
ID
Metric
Norm
Ref
50
50
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
60
80
80
80
80
80
80
80
80
80
80
80
80
80
80
80
80
80
80
80
80
80
80
80
80
80
80
80
80
80
80
90
90
90
90
90
90
90
90
90
90
90
90
90
90
90
90
90
90
90
90
90
90
90
90
PLS
H
A
A
A
A
H
A
A
A
A
A
A
A
A
A
H
H
H
A
A
A
A
H
H
H
A
H
H
A
H
H
PLS
CLS
A
A
CLS
H
A
H
H
PLS
A
H
SHIFT
A
A
A
H
A
CLS
A
A
H
H
A
CORR
H
PLS
PLS
PLS
PLS
H
PLS
CLS
A
H
H, A
H, A
MLR
H
A
NO NE
1 IN STD
NO NE
IN STD
IN STD
RATIO
IN STD
EX STD
EX STD
EX STD
EX STD
NO NE
EX STD
NO NE
w S BAND
IN STD
IN STD
EX STD
IN STD
IN STD
NO NE
IN STD
1 IN STD
EX STD
RATIO
IN STD
IN STD
RATIO
IN STD
EX STD
IN STD
RATIO
NO NE
IN STD
EX STD
IN STD
RATIO
NO NE
NO NE
1 IN STD
RATIO
NO NE
NO NE
NO NE
NO NE
IN STD
RATIO
RATIO
IN STD
IN STD
IN STD
RATIO
IN STD
1 IN STD
1 IN STD
IN STD
IN STD
NO NE
MSC
S SPEC
SNV
IN STD
IN STD
SNV
IN STD
EX STD
RATIO
EX STD
EX STD
EX STD
NO NE
EX STD
119
119
26
26
25
23
42
55
56
57
19
45
15
70
52
95
101
33
118
147
147
146
136
9
145
171
155
88
73
77
84
153
113
74
76
142
152
144
64
122
123
109
168
94
149
27
27
154
62
78
82
150
159
118
175
47
47
75
48
169
169
93
126
126
160
46
157
124
163
163
137
162

Peletier Article

Diunggah oleh

Informasi Dokumen

Hak Cipta

Format Tersedia

Bagikan dokumen Ini

Bagikan atau Tanam Dokumen

Opsi Berbagi

Apakah menurut Anda dokumen ini bermanfaat?

Apakah konten ini tidak pantas?

Hak Cipta:

Format Tersedia

Peletier Article

Diunggah oleh

Hak Cipta:

Format Tersedia

focal point

aman spectroscopy has become m ore popular in recent

Volume 57, Number 1, 2003

preparation. Raman samples can be

three of these analytical needs for

n 5 M olecular vibrational frequency in Hertz

creased by using a stronger laser.

Volume 57, Number 1, 2003

cally present in Raman spectra can

Sample-generated noise also inclu des un w anted op tical intensity

of the detector is digitized. A lot of

a comm on source of externally generated noise. In the near-infrared

analysis is called preprocessing. Preprocessing improves the accuracy

Volume 57, Number 1, 2003

similar plots. Simply selecting wavenumber regions that correlate with

is derived from the rst few Fourier

separation between adjacent points

sp ectra availab le to th e operator.

Raman bands. Higher frequencies

Volume 57, Number 1, 2003

independently on segments of the

wide. If all Raman bands are known

sets that are becoming comm on in

as m uch variance in the set of spectra as possible. PC1 is then removed

Patterns. Spectral stripping 29 can be

Determination of Band Area.

Volume 57, Number 1, 2003

used by an expert operator to unintentionally re inforc e his incorrect

sample for the Raman scattered light,

made. Raman spectra can sometimes

quantitative analysis using absolute

travels through the sample. This can

Volume 57, Number 1, 2003

der identical m easurem ent con ditions. Unfortunately, the addition of

stro ng ban d cen tere d near 3 450

The use of several bands sum med

can all change the chemical bond

m an spectrum , but allows the band

the root m ean square error of prediction, RMSEP, given by:

The assumption of negligible error in

Volume 57, Number 1, 2003

where n is the number of samples in

lytical m odel is to insert noise into a

imately linear. Higher analyte concentration displaces the solvent from

proportional to total analyte concentration. For example, a weak organic

quantitative information about the

Volume 57, Number 1, 2003

Raman spectrum of the sample. 29

Cross-Correlation. Raman band

C(x) is the correlation function, S (t)

Indirect Classical Least Squares

does these operations simultaneously. Unlike MLR, PCR retains the

Volume 57, Number 1, 2003

K. M ann, Appl. Spectrosc. 55, 389

60. H. Abderrazak, M . Dachraoui, M. J.

87. B. J. Pelletier, p calculation program

Multiplicative scatter correction

Ref (Literature referen ces for applications)

25. D. A. Leonard, US Patent 3,723,007

F. Gaskin, and K.-N. Kruger, J. Mol.

Gervasio, Appl. Spectrosc. 47, 1115

100. G. Kister, G. Cassanas, and M. Vert,

Langkilde, Chemom. Intell. Lab. Syst.

142. M. J. Pelletier, K. L. Davis, and R. A.