Quantitative in situ monitoring of an elevated temperature reaction using a water-cooled mid-infrared fiber-optic probe

Article abstract of DOI:10.1021/ac9502965

A novel water-cooled mid-infrared fiber-optic probe is described which is heatable to 230°C. The probe has chakogenide fibers and a ZnSe internal reflection element and is compact and fully flexible, allowing access to a wide range of standard laboratory reaction vessels and fume cupboard arrangements. Performance is demonstrated via the in situ analysis of an acid-catalyzed esterification reaction in toluene at 110°C, and the results are compared with those from a conventional extractive sampling loop flow cell arrangement Particular emphasis is given to the quantitative interpretation of the spectro-scopic data, using gas chromatographic reference data. Calibration data are presented for univariate and partial least squares models, with an emphasis on procedures for improving the quality of interpreparation calibration and prediction through the use of focused reference analysis regimes. Subset univariate procedures are presented that yield relative errors of <5%, and bias-corrected partial least squares procedures are described that result in relative errors of interpreparation calibration and prediction consistently <3%. This paper demonstrates the considerable power of fiber-optic mid-IR spectroscopy combined with bias correction partial least squares procedures for the efficient in situ quantitative analysis of laboratory scale reactions.

Full text of DOI:10.1021/ac9502965

Anal. Chem. 1996, 68, 1116-1123
Quantitative in Situ Monitoring of an Elevated
Temperature Reaction Using a Water-Cooled
Mid-Infrared Fiber-Optic Probe
Paul MacLaurin* and Nicholas C. Crabb
Process Technology Department, Zeneca FCMO, Leeds Road, Huddersfield HD2 1FF, U.K.
Ian Wells and Paul J. Worsfold
Department of Environmental Sciences, University of Plymouth, Drake Circus, Plymouth PL4 8AA, U.K.
David Coombs
Graseby Specac Limited, River House, 97 Cray Avenue, St. Mary Cray, Orpington BR5 4HE, U.K.
development. On-line analysis of laboratory-scale reactions can
markedly improve both the timeliness and quality of information
regarding reaction mechanisms and kinetics, as compared with
traditional approaches of extractive sampling and quenching
followed by liquid chromatographic analysis. In contrast to on-
plant process analyzers, the technology for on-line analysis in
process development needs to be inherently versatile. Further-
more, analytical development times need to be kept to a minimum
in order to maintain consistency with the time frame of process
development programs. Mid-infrared spectroscopy is attracting
increasing attention as a process analysis tool,2 and this paper
describes advances in the use of fiber-optic mid-IR spectroscopy
for the in situ analysis of laboratory-scale organic reactions.
Vibrational spectroscopy has long been a powerful tool in
materials characterization. Virtually all organic functional groups
exhibit fundamental vibrations at frequencies in the mid-IR region
of the electromagnetic spectrum (4000-400 cm-1). Typically, the
information content of mid-IR spectra is high, with bands that are
often narrow and well resolved, allowing direct functional group
assignment. IR absorbance spectra have traditionally been
recorded using dispersive instruments, but since 1970, Fourier
transform infrared (FT-IR) instruments have been available that
are capable of collecting high-quality spectra in a fraction of the
time previously required with enhanced signal-to-noise and wave-
number accuracy. Despite this speed of response and its potential
to directly monitor functional group vibrations, FT-IR has found
little application to date for the on-line analysis of organic reactions.
This is largely attributable to underdeveloped sample interfacing.
Absorptions in the mid-IR region tend to be strong, neces-
sitating the use of short path lengths for transmission studies.
Attenuated total reflection (ATR) techniques provide a means of
achieving path lengths of effectively only a few micrometers and
as a result have become a general solution for liquid sample
analysis.3 ATR spectroscopy utilizes the absorbances that occur
when total internal reflection takes place at the interface between
two materials. If a transparent material of high refractive index
A novel water-cooled mid-infrared fiber-optic probe is
described which is heatable to 2 3 0 °C. The probe has
chalcogenide fibers and a ZnSe internal reflection element
and is compact and fully flexible, allowing access to a wide
range of standard laboratory reaction vessels and fume
cupboard arrangements. P erformance is demonstrated
via the in situ analysis of an acid-catalyzed esterification
reaction in toluene at 1 1 0 °C, and the results are
compared with those from a conventional extractive
sampling loop flow cell arrangement. P articular emphasis
is given to the quantitative interpretation of the spectro-
scopic data, using gas chromatographic reference data.
Calibration data are presented for univariate and partial
least squares models, with an emphasis on procedures
for improving the quality of interpreparation calibration
and prediction through the use of focused reference
analysis regimes. Subset univariate procedures are pre-
sented that yield relative errors of <5 %, and bias-cor-
rected partial least squares procedures are described that
result in relative errors of interpreparation calibration and
prediction consistently <3 %. This paper demonstrates
the considerable power of fiber-optic mid-IR spectroscopy
combined with bias correction partial least squares pro-
cedures for the efficient in situ quantitative analysis of
laboratory scale reactions.
From its roots in commodity chemicals manufacturing, process
analytical chemistry is now impacting on virtually all sectors of
chemical processing. One example is the fine chemicals industry,
where typically batch processing, often in a nondedicated plant,
has required the development of a multifaceted approach to
process analysis.1 In addition to producing a diverse range of
chemical entities, fine chemicals manufacturing generally involves
the frequent introduction of new processes, which require
extensive and time-consuming process research and development,
and the desire for greater efficiency has led to the process
analytical chemistry philosophy being applied during process
(2) Crabb, N. C.; King, P. W. B. Molecular Spectroscopy. In Process Analytical
Chemistry; McLennan, F., Kowalski, B. R., Eds.; Blackie: London, 1995.
(3) Harrick, N. J. Internal Reflection Spectroscopy; Wiley: New York, 1967.
(1) Crabb, N. C.; McLennan, F. Process Control Qual. 1 9 9 2 , 3, 229-235.
1116 Analytical Chemistry, Vol. 68, No. 7, April 1, 1996
0003-2700/96/0368-1116$12.00/0 © 1996 American Chemical Society

is placed in direct contact with the sample, light passing through
this at angles above a critical angle will be reflected at the sample
interface. In the process of reflection, an evanescent wave
penetrates the sample medium, giving rise to partial absorption
of the beam according to the vibrational frequencies of the sample.
The missing link, therefore, for in situ FT-IR analysis is a means
of transmitting the radiation between the spectrometer and the
internal reflection element. One option is the use of optical light
guides, hollow conduits internally coated with a reflecting material
such as nickel or gold.4 “Light pipes” have been used in
conjunction with ATR probes for reaction monitoring, but their
versatility is hampered by their rigid construction. Alternatively,
reaction vessels can be constructed with integral internal reflection
elements in dedicated spectrometer systems, but again, these are
somewhat restrictive.5
Figure 1. Acid-catalyzed reaction of but-2-enoic acid and butan-
2-ol, to give the corresponding ester, sec-butyl but-2-enoate.
Table 1. Spectroscopic Variables for the Fiber Probe
and Flow Cell Analysis
spectral spectral
range,
resoln, J-stop, B-stop,
throughput,
accessory
cm-1
cm-1
mm
mm gain %
fiber probe 4000-1000
flow cell 4000-700
4
4
11.00 15.00
11.00 8.00
2
1
∼13
∼48
Process development activity tends to involve the use of a wide
range of reactor designs and capacities, and to gain access to a
significant proportion of these reactors, a flexible approach to in
situ IR sampling is required, something that is now possible due
to recent developments in fiber-optic technology. In the simplest
terms, optical fibers are a medium for transmitting radiant power
and consist of a core and cladding. Radiation traveling in the core
is confined to the core by reflection with the cladding that has a
lower index of refraction. Fiber links are now commonplace for
applications in visible and near-IR spectroscopy, but their perfor-
mance decreases markedly in the mid-IR because the vast majority
of materials have fundamental absorptions in this region. This
problem is exacerbated by the low throughput of available fibers,
the inherently low energy of mid-IR sources, and the relative
insensitivity of mid-IR detectors. Recently, however, chalcogenide
fibers have become commercially available and are beginning to
be used in this context.
apparent IR activity, and minimal safety hazards.7 The reaction
scheme is given in Figure 1. A series of three separate reactions
was monitored using the fiber probe, and a further series of three
reactions was monitored using an overhead ATR flow cell with a
recirculating loop. Particular emphasis is given to the quantifica-
tion of the reaction product (sec-butyl but-2-enoate) and the
approach taken to improve its quantitative prediction with a
minimum of analytical development.
EXPERIMENTAL SECTION
Instrumentation. All spectra were recorded using a Perkin-
Elmer System 2000 spectrometer fitted with a liquid nitrogen-
cooled narrow band mercury-cadmium-telluride (MCT) detec-
tor. The spectrometer was operated in the infrared data manager
(IRDM) environment using the time-resolved facility (TRIR).
Spectra were stored as the mean of 64 co-added scans collected
at 10 min intervals and were ratioed against an air background
recorded immediately prior to analysis. The spectroscopic
variables used for the fiber probe and the flow cell are summarized
in Table 1.
Chalcogenide glasses are vitreous materials made from group
IV metals As, Ge, and Sb, together with the chalcogen elements
S, Se, and Te. Chalcogenide transmission is good over the range
4000-1000 cm-1, apart from lattice vibrations and absorption due
to hydrogen-containing impurities at 2250-2100 cm-1, although
high-attenuation characteristics have restricted their practical use
to lengths of 2 m. Nevertheless, 10 m chalcogenide fibers have
been reported6 for environmental analysis over a restricted spectral
range.
The cooled mid-IR fiber probe consisted of two single-strand
750 µm, 1.5 m chalcogenide fibers with a numerical aperture of
0.4, in conjunction with a two-reflection 45° ZnSe internal reflection
element (Figure 2a) and an adjustable optical interface for the
spectrometer (Figure 2b). The composition of the fiber-optic
material used was As0.4, Se0.2, Te0.4, with minimal losses of 0.001
dB m-1 at 6 µm. The glass transition temperature of this material
was 136 °C, although the fibers do change their transmission
characteristics at temperatures as low as 70 °C. The step index
fibers were polymer clad and armor plated with spirally wound
stainless steel, offering a bend radius of 15 cm. Larger diameter
fibers could have been employed, enabling greater utilization of
the available radiation, but this would have restricted the system
flexibility. The fiber was deliberately overfilled in both numerical
aperture and area, and a high-quality optical design was used to
image the radiation back to the fiber. The probe features an
integral internal water cooling system that maintains the fiber at
35 °C when the probe is immersed in a sample at 200 °C. Clearly,
this is a safe margin with respect to the 70 °C fiber transition
temperature. Cooling water was circulated via stainless steel
The analysis of elevated temperature reactions raises further
issues, however. Low glass transition temperatures, in combina-
tion with low softening temperature cladding materials, have
restricted the use of chalcogenide fibers to temperatures of ∼70
°C, thus rendering a significant proportion of typical process
development reactions inaccessible. To address this shortcoming,
a fully flexible fiber-optic probe heatable to temperatures up to
230 °C has recently been developed.
This paper describes the concept and construction of a novel
water-cooled mid-IR fiber probe and its use in the monitoring of
a laboratory-scale esterification reaction at 110 °C. The acid-
catalyzed reaction of but-2-enoic acid and butan-2-ol was selected
on account of its reaction temperature being greater than 70 °C,
(4) Doyle, W. M.; Jennings, N. A. Spectrosc. Int. 1 9 9 1 , 2, 48-52.
(5) Full, A. P.; Puig, J. E.; Gron, L. U.; Kaler, E. W.; Minter, J. R.; Mourey, T.
H.; Texter, J. Macromolecules 1 9 9 2 , 25, 5157-5164.
(6) Ewing, K. J.; Bilodeau, T.; Nau, G.; Aggarawal, I. D.; King, T.; Clark, R.;
Robitille, G. Air and Waste Management/ SPIE Conference Proceedings,
McLean, VA, November 1994.
(7) Harwood, L. M.; Moody, C. J. Experimental Organic Chemistry; Blackwell:
Oxford, 1989; pp 445-446.
Analytical Chemistry, Vol. 68, No. 7, April 1, 1996 1117

Figure 2. (a) Diagram of the water-cooled fiber probe illustrating the internal optics and cooling hardware. (b) Diagram of the instrument
optical interface (plan view).
capillaries parallel to, but not in direct contact with, the fibers.
The capillaries were housed in an air-filled tube surrounded by
an insulating polymeric material. The probe was constructed from
316 stainless steel and was 256 mm long. The maximum diameter
of the immersed probe head was 21 mm, but this could be reduced
to 19 mm by detaching the removable ATR crystal protector. The
probe was fitted with a stainless steel cone (B24/ 29) to provide a
direct coupling for general laboratory glassware. The cone could
be moved along the probe body and locked to vary the depth of
immersion. The spectrometer interface consisted of an adjustable
mirror and ZnSe lens arrangement for both the launch and
collection optics. The cooling system was fitted to a standard
laboratory tap via a 50 µm filtering device, and cooling water was
2 mL of sulfuric acid (Fisons) was added, and the reaction mixture
was heated to 110 °C. Reflux at this temperature was maintained
throughout the course of the 8 h reaction.
A sample was manually extracted at t ) 0 min and at 30 min
intervals thereafter for 8 h. The samples were used for reference
analysis by an internal standard gas chromatography procedure
for the quantification of but-2-enoic acid, butan-2-ol, and sec-butyl
but-2-enoate. The procedure used split injection and temperature
programming with undecane as internal standard. Sample solu-
tions were prepared by the dilution of ∼1.0 g of accurately weighed
reaction mixture to 100 mL with dichloromethane. All samples
solutions were analyzed in duplicate, and a standard mixture was
analyzed after every four injections.
Chemometric Data Analysis. General Multivariate Calibra-
tion. Partial least squares (PLS) regression models were devel-
oped for all constituents using the spectra from the six individual
preparations and the GC reference data. All models were
developed using the PLS-1 algorithm of the Perkin-Elmer Quant+
software. Full cross validation was used for all calibrations. Model
dimensionality was determined as the first significant local
minimum in the relative root mean square error of cross validation:
supplied at a rate of 15-20 mL min-1
.
All flow cell data were collected using a Graseby Specac
advanced overhead ATR system fitted with a 550 µL thermosta-
bilized flow-through top plate assembly, incorporating a six-
reflection 45° ZnSe internal reflection element and Kalrez gaskets.
The flow cell was maintained at 90 °C throughout all experiments.
The reaction mixture was continuously recirculated from the
reaction vessel through narrow bore stainless steel tubing via a
Gilson Minipuls 3 peristaltic pump fitted with Viton rubber pump
tubing.
I
(y - yˆi)2
Gas chromatography (GC) was carried out using a Hewlett
Packard HP 5890 Series II chromatograph with a 25 m Chrompack
Cp-Sil-5CB 0.32 mm i.d. capillary column.
∑
i
100
i)1
RRMSECV (%) )
x
jy
I
P rocedure. All reactions were carried out in a standard 1 L
flanged reaction vessel fitted with a stirrer, Dean-Stark apparatus,
thermometer, and either the fiber probe or a stainless steel
sampling tube for the flow cell experiments. Heat was supplied
via a heating mantle. The experimental setup for the fiber probe
experiments (preparations 1-3) is illustrated schematically in
Figure 3. The setup for the flow cell experiments (preparations
4-6) was identical except for the use of the stainless steel tubing
and pump arrangement in place of the fiber probe.
where y is the concentration of object i, yˆi is the predicted
i
concentration of object i, I is the total number of objects used in
the calibration, and jy is the mean analyte concentration. Because
the prediction objects are not used during calibration, no degrees
of freedom are lost.
Interpreparation Calibration and Prediction. The aim of this
exercise was not to develop a universal calibration for the long-
term determination of sec-butyl but-2-enoate in the reaction
mixture, because this would require extensive additional experi-
mental preparations, supplemented by full reference analysis by
At room temperature, ∼86 g of 98% but-2-enoic acid (Aldrich)
was added to 500 mL of toluene (Fisons) with stirring, followed
by the addition of 154 mL of butan-2-ol (Aldrich). At t ) 0 min,
1118 Analytical Chemistry, Vol. 68, No. 7, April 1, 1996

Figure 3. Experimental setup for the in situ analysis of the preparation of sec-butyl but-2-enoate using the water-cooled fiber probe.
gas chromatography. Such an approach to quantitation is in itself
time-consuming, but this would also lead to significant delays in
the process development program. Instead, procedures for
extracting the maximum information about individual preparations
with minimal reference analysis were investigated, with emphasis
on the determination of the reaction product sec-butyl but-2-enoate.
Quantitation of sec-butyl but-2-enoate was carried out using
univariate and multivariate calibration procedures, again using
PLS-1 with full cross validation for all the multivariate models.
For both univariate and multivariate procedures, the quality of
the calibration models was compared in terms of their abilities to
predict the concentration of sec-butyl but-2-enoate in other
preparations and was quantified by the relative root mean square
error of prediction (RRMSEP). The RRMSEP is calculated in the
same way as the RRMSECV and is expressed as a percentage.
Again, no degrees of freedom are lost because all the prediction
samples are independent of the calibration model.
be made of the measurement bias between the original calibration
and the preparation under investigation. The significance of the
bias was estimated using F-statistics and applied accordingly as
either an offset correction or a slope and offset correction. Bias
correction does not change the original PLS model, i.e. the original
latent variables are retained; only the newly predicted values are
corrected for the bias. It should be noted that applying bias
correction to interpreparation calibration and prediction represents
a departure from its conventional use. Bias correction is generally
used to correct for systematic differences caused by changing a
sampling accessory or transferring a calibration between instru-
ments. However, the nature of the data manipulation is very
similar, and the concept holds, providing that good-quality refer-
ence data are available.
Applying the same principle to univariate statistics leads to
subset calibration. Unlike PLS, the determination of a least
squares univariate regression equation does not involve data
reduction. It follows, therefore, that bias correction is not
necessary because the same four samples used for PLS bias
correction can be directly used with their reference data to
determine a new univariate regression equation. This equation
can then be used for intrapreparation prediction of the sec-butyl
but-2-enoate concentrations from the other spectral data collected
in that preparation.
Two calibration and prediction regimes were investigated, and
they are summarized below.
(i) Single Preparation Calibration. Calibration models (univari-
ate and PLS) were developed using the spectra from a single
preparation with full GC reference data. These models were used
to predict the concentration of sec-butyl but-2-enoate in the other
preparations from the spectra recorded using the same sampling
accessory.
(ii) Bias Correction and Subset Calibration. PLS models were
developed using the spectra from a single preparation with full
GC reference data, as for the single preparation calibration
scenario. However, prior to the prediction of the ester concentra-
tion in subsequent preparations, GC reference data and the
appropriate spectra for four samples from the preparation under
investigation were made available. This enabled an estimate to
RESULTS AND DISCUSSION
The reaction profile of preparation 1, as determined by gas
chromatography, is presented in Figure 4. As anticipated, it
clearly shows the parallel consumption of but-2-enoic acid and
butan-2-ol in tandem with the formation of sec-butyl but-2-enoate.
Very similar profiles were observed for the other preparations.
Analytical Chemistry, Vol. 68, No. 7, April 1, 1996 1119

Figure 4. Reaction profile (preparation 1) determined by gas
chromatography.
Figure 7. Spectral overlay of the C-O stretching region (preparation
1; fiber probe).
Table 2. PLS-1 Full Cross Validation Models for All
Constituents after Mean-Center Scaling
data no. of
optimal
RRMS
accessory
analyte
set objects dimensionality CV, %
fiber-optic sec-butyl
1
2
3
1
2
3
1
2
3
16
16
16
16
16
16
16
16
16
2
2
2
2
2
2
2
2
2
1.54
2.36
1.70
2.19
3.28
1.87
3.53
5.49
4.35
probe but-2-enoate
butan-2-ol
but-2-enoic acid
flow cell sec-butyl
but-2-enoate
4
5
6
4
5
6
4
5
6
15
10
14
15
10
14
15
10
14
2
2
2
2
1
2
2
1
2
2.52
2.43
2.16
2.64
1.76
2.67
5.58
3.05
4.26
Figure 5. Time-resolved spectra of the carbonyl stretching region
(preparation 1; fiber probe).
butan-2-ol
but-2-enoic acid
Figure 7. Here, the development of the multiple CsO stretching
bands associated with the formation of the R,â-unsaturated ester
are clearly visible (νASCsC(dO)sO, 1265 cm-1 and 1185 cm-1
ν
;
ASOsCsC, 1100 cm-1).
General Multivariate Calibration. Details of the PLS-1 full
cross validation models developed with the data from the six
individual preparations is presented in Table 2. Preliminary
modeling indicated that most of the correlated spectral variance
was in the spectral region 2000-1000 cm-1, and all models were
therefore developed using this range. Various approaches to
spectral preprocessing were considered, but only mean-center
scaling had a beneficial impact on the RRMSECV. Model
dimensionality was generally estimated at 2, returning RRMSECV
values of less than 5% in all but two cases.
Figure 6. Spectral overlay of the carbonyl stretching region
(preparation 1; fiber probe).
Figure 5 shows the high-quality time-resolved spectra of the
carbonyl stretching region, collected during preparation 1 using
the fiber probe. In accordance with the consumption of but-2-
enoic acid and concurrent formation of sec-butyl but-2-enoate, the
absorbance at 1700 cm-1 (νSCdO, R,â-unsaturated carboxylic acid)
clearly decreases as the band at 1720 cm-1 (νSCdO, R,â-
unsaturated ester) evolves. Overlaying a series of spectra taken
at 30 min intervals during preparation 1 (Figure 6) illustrates this
transition more clearly and verifies the high resolution achievable
with liquid phase mid-IR spectroscopy. A spectral overlay of the
CsO stretching region, again from preparation 1, is shown in
Figure 8 shows the coefficient weightings plot for sec-butyl but-
2-enoate from the preparation 1 data set. The areas of correlation
are consistent with those of high spectral variance identified in
Figures 5-7, i.e., the CdO and CsO stretching regions. Similar
correlation patterns were observed for the sec-butyl but-2-enoate
models for the data from the other preparations. It is interesting
to note the negative peak at 1700 cm-1, consistent with acid
consumption, next to the large positive peak at 1720 cm-1
,
consistent with ester formation. Examination of the coefficient
1120 Analytical Chemistry, Vol. 68, No. 7, April 1, 1996

Table 4. Univariate Calibrations and Predictions for
sec-Butyl But-2-enoate (Single Preparation Calibration)
calibrn
set
correln predn RRMSEP,
accessory
slope
intercept coeff
set
%
fiber-optic
probe
1
2
3
4
0.006 57 0.0911
0.006 38 0.0931
0.006 00 0.1176
0.020 94 0.2331
0.9966
0.9882
0.9976
0.9983
2
3
1
3
1
2
5
6
2.86
7.61
3.53
8.93
8.25
9.32
53.86
5.89
flow cell
5
6
0.016 35 0.5610
0.019 54 0.2921
0.9978
0.9976
4
6
4
5
48.71
44.24
6.06
9.32
Figure 8. PLS-1 coefficient weightings plot for sec-butyl but-2-
enoate model from preparation 1.
Table 5. PLS-1 Predictions for sec-Butyl But-2-enoate
(Single Preparation Calibration)
Table 3. PLS-1 Full Cross Validation Models of the
Combined Data Sets After Mean-Center Scaling
accessory
calibrn set
1
predn set
RRMSEP, %
no. of
optimal
RRMSECV,
%
fiber-optic probe
2
3
1
3
1
2
6.19
5.22
accessory
analyte
objects dimensionality
2
3
5.24
fiber-optic sec-butyl
48
3
2.26
13.69
13.93
21.11
probe
but-2-enoate
butan-2-ol
but-2-enoic acid
48
48
3
3
3.18
5.18
flow cell
4
5
6
5
6
4
6
4
5
42.76
43.00
48.99
136.0
4.73
flow cell
sec-butyl
but-2-enoate
butan-2-ol
39
2
18.83
39
39
2
2
13.08
21.05
but-2-enoic acid
58.67
weightings for butan-2-ol and but-2-enoic acid revealed a very
similar correlation structure but in mirror image to that for the
sec-butyl but-2-enoate. This is rationalized by the fact that there
is no mechanism for controlling the calibration design during
reaction monitoring; as the acid and alcohol are consumed, the
ester is produced concurrently, and therefore no discrimination
can be made in terms of covariance.
Calibration data are presented in Table 4, together with the
relative errors of prediction for the other preparations. The slope
and intercept values are quite similar for the three fiber probe
models, and they yield RRMSEP values that are consistently <10%
for the other preparations. The variability of the slope and
intercept values for the univariate flow cell models is consistent
with the background differences described above. These differ-
ences have manifested in gross errors in the prediction of other
preparations, e.g., a relative error of 54% using the preparation 4
model to predict preparation 5.
The results of interpreparation predictions based on the PLS
models from Table 2 are presented in Table 5. This again reveals
large relative errors for the flow cell preparation data sets, but
the fiber probe preparation models also resulted in prediction
errors of up to 21%.
While some of the univariate interpreparation calibration and
prediction results are quite encouraging, the general standard of
prediction is rather poor, and in the context of a process
development/ process optimization excercise, predictions with
relative errors of 40-50% are of little diagnostic value.
Bias Correction and Subset Calibration. Univariate subset
calibrations were undertaken using four selected samples from
each preparation. For preparations 1-4 and 6, the samples taken
at t ) 60, 180, 300, and 420 min were used. The samples taken
at t ) 60, 150, 240, and 330 min were used from preparation 5.
Calibration details and prediction errors for each preparation are
presented in Table 6. The univariate subset approach to intra-
preparation calibration and prediction has resulted in relative
errors of <3% for the fiber probe data and <5% for the flow cell
data. This represents a marked improvement on the values of
In addition, all of the data obtained from the fiber probe were
combined to form a larger data set for calibration, as were all of
the flow cell data. The resulting model details are presented in
Table 3. It can be seen that the models for all constituents
developed from the fiber probe data yielded relative errors of the
same magnitude as the individual preparation models (Table 2)
but that a third dimension was required to model the interprepa-
ration variance. However, the relative errors of the flow cell
models were significantly higher (>20% for the but-2-enoic acid),
and examination of the model parameters revealed significant
differences between the spectral data sets. The scores of the first
two latent variables revealed three distinct groupings according
to the individual preparation (4-6), although the GC reaction
profiles were very similar. This was explained by differences in
the background measurements collected prior to each preparation.
Interpreparation Calibration and P rediction. Single Prepa-
ration Calibration. Resolution in the mid-IR allows the direct
selection of frequencies for univariate calibration. Preliminary
investigations indicated that the CsO stretching region was more
stable than the carbonyl region, and the absorbance at 1185 cm-1
(νASCsC(dO)sO) was therefore used for calibration. The ab-
sorbance at 1055 cm-1 was subtracted from this to account for
temporal drift in the background spectrum.
Analytical Chemistry, Vol. 68, No. 7, April 1, 1996 1121

to mutual interferences. In such cases, the numerical selectivity
offered by reduced dimension multivariate calibration will be
required for accurate predictions. This is supported by the sec-
butyl but-2-enoate prediction results, which reveal that the bias
correction PLS procedures have accommodated the interprepa-
ration variability more comfortably than the univariate routines.
In the context of a process development environment, the
concept of bias correction PLS procedures for interpreparation
calibration and prediction is very powerful. A typical process
development program will involve multiple reactions, often with
only subtle differences in the conditions of these reactions. With
full reference analysis of the first preparation in such a program,
and the development of an appropriate PLS model for the
constituents of interest, a limited but well-focused regime of
reference analysis of the subsequent preparations will be sufficient
to provide fully quantitative information from in situ IR spectra
through the bias correction PLS procedure. Furthermore, there
is potential to extend this concept to series of reactions with more
radical differences in the reaction conditions, e.g., different
catalysts. It is envisaged that judicious editing of mid-IR spectra
will enable the influence of these differences to be marginalized,
while the combined resolving power of mid-IR spectroscopy and
bias correction PLS regression provides a robust quantification
of the analytes of interest.
Table 6. Univariate Subset Intrapreparation
Calibrations and Predictions for sec-Butyl
But-2-enoate
calibrn
set
correln
coeff
RRMSEP,
%
accessory
slope
intercept
fiber-optic
probe
1
2
3
0.00625
0.00643
0.00598
0.0925
0.0912
0.1184
0.9971
0.9971
0.9976
2.59
1.89
2.36
flow cell
4
5
6
0.02188
0.01606
0.01860
0.2162
0.5634
0.3068
0.9986
0.9968
0.9990
3.60
3.64
4.58
Table 7. Interpreparation Bias-Corrected PLS-1
sec-Butyl But-2-enoate Predictions
calibrn
set
predn
set
bias corrn
applied
RRMSEP,
%
accessory
fiber-optic probe
1
2
3
2
3
1
3
1
2
offset
offset
slope, offset
offset
slope, offset
offset
1.71
1.68
1.19
2.27
1.25
1.78
flow cell
4
5
6
5
6
4
6
4
5
slope, offset
slope, offset
slope, offset
offset
none
slope, offset
2.00
2.41
3.84
3.52
4.73
2.67
The adoption of these procedures goes a long way toward
improving both the quality and timeliness of analytical measure-
ments during process development. Ultimately, however, proce-
dures capable of providing accurate information without the need
for reference analysis are desired, and recent advances in the
10% and up to 54%, respectively, from the single preparation
calibration scenario, which were caused by variability in the
regression equation coefficients. Any such systematic interprepa-
ration variability will clearly have no effect on the predictive ability
of subset intrapreparation calibration and predictions.
The same four samples from each preparation were also used
to determine the bias correction for the PLS models. The
corrections used and the prediction results are given in Table 7.
For all but one of the prediction sets, either an offset or a
combined slope and offset correction was shown to have a
significant impact on the relative errors, reducing the RRMSEP
values to a mean of 1.6% for the fiber probe preparations and a
mean of 3.2% for the flow cell preparations. As with the univariate
subset predictions, this represents a significant improvement on
the gross errors recorded using PLS calibration under the single
preparation calibration scenario.
Both the univariate subset and PLS bias correction approaches
to interpreparation calibration and prediction have proved ben-
eficial in terms of the relative error of prediction under these
conditions for this series of reactions. Furthermore, this has been
possible with a minimum of reference analysis and is therefore
consistent with process development activity.
Benefits for P rocess Development. In this particular
application, the characteristic mid-IR bands of the ester provided
stable and well-resolved bands, allowing direct quantification by
the univariate procedures discussed above. However, such an
approach is unlikely to be appropriate for data from analytical
systems of lower resolution, such as UV-visible and near-IR
spectroscopy and electrochemical sensing devices. Moreover,
univariate procedures will not be adequate for interpreparation
calibration and prediction for applications in mid-IR spectroscopy,
where the spectral transitions are subtle or the bands are subject
application of multivariate curve resolution techniques8
show great
promise. Since the need for reference analysis is eliminated, no
secondary method development is required, and the potential
hazards of extractive sampling are avoided. However, curve
resolution procedures are far from trivial to implement success-
fully, and until more robust approaches are available, bias
correction PLS provides an attractive alternative.
CONCLUSIONS
Chalcogenide fibers have been shown to provide a safe and
flexible approach to mid-IR sampling. The incorporation of a water
cooling circuit in the body of a novel fiber probe has rendered
this technology amenable to elevated temperature reactions which
were previously not possible due to the instability of transmission
properties above 70 °C.
Fiber probe analysis of the esterification of but-2-enoic acid
and butan-2-ol at 110 °C produced high-quality mid-IR ATR spectra
exhibiting well-resolved spectral features, confirming the high
spectral resolution achievable with liquid phase mid-IR spectros-
copy. The resulting spectra were as good as those obtained using
a conventional flow-through overhead ATR arrangement with an
extractive sampling loop but had the advantage of being truly in
situ measurements.
Calibrations were developed using data from individual prepa-
rations from both the fiber probe and flow cell experiments, but
they failed to predict the concentration of sec-butyl but-2-enoate
in independent preparations without a high degree of error. To
minimize these differences, procedures were investigated to
provide accurate predictions with minimal time-consuming refer-
ence analysis. Using only four reference samples, subset intra-
(8) Tauler, R.; Kowalski, B. R.; Fleming, S. Anal. Chem. 1 9 9 3 , 65, 2040-2047.
1122 Analytical Chemistry, Vol. 68, No. 7, April 1, 1996

preparation univariate calibration and prediction reduced the
relative errors prediction to <3% for the fiber probe data sets and
<5% for the flow cell data sets. Combinations of offset and slope
corrections to the PLS models, determined using the same four
reference samples, reduced the relative errors of interpreparation
calibration and prediction to the order of 2% and 3% for fiber probe
and flow cell data preparations, respectively.
FCMO) with the gas chromatography. Zeneca FCMO is gratefully
acknowledged for financial support. I.W. thanks the SERC (U.K.)
for a research studentship.
Received for review March 27, 1995. Accepted July 11,
1995.X
AC9502965
ACKNOWLEDGMENT
The authors acknowledge the assistance of L. L. Day and G.
Poulter (Graseby-Specac) with the optics and J. N. Hirst (Zeneca
X Abstract published in Advance ACS Abstracts, January 1, 1996.
Analytical Chemistry, Vol. 68, No. 7, April 1, 1996 1123