Automated data evaluation and modelling of simultaneous 19F-1H medium-resolution NMR spectra for online reaction monitoring

Article abstract of DOI:10.1002/mrc.4216

Medium-resolution nuclear magnetic resonance spectroscopy (MR-NMR) currently develops to an important analytical tool for both quality control and process monitoring. In contrast to high-resolution online NMR (HR-NMR), MR-NMR can be operated under rough environmental conditions. A continuous re-circulating stream of reaction mixture from the reaction vessel to the NMR spectrometer enables a non-invasive, volume integrating online analysis of reactants and products. Here, we investigate the esterification of 2,2,2-trifluoroethanol with acetic acid to 2,2,2-trifluoroethyl acetate both by ¹H HR-NMR (500 MHz) and ¹H and ¹⁹F MR-NMR (43 MHz) as a model system. The parallel online measurement is realised by splitting the flow, which allows the adjustment of quantitative and independent flow rates, both in the HR-NMR probe as well as in the MR-NMR probe, in addition to a fast bypass line back to the reactor. One of the fundamental acceptance criteria for online MR-MNR spectroscopy is a robust data treatment and evaluation strategy with the potential for automation. The MR-NMR spectra are treated by an automated baseline and phase correction using the minimum entropy method. The evaluation strategies comprise (i) direct integration, (ii) automated line fitting, (iii) indirect hard modelling (IHM) and (iv) partial least squares regression (PLS-R). To assess the potential of these evaluation strategies for MR-NMR, prediction results are compared with the line fitting data derived from the quantitative HR-NMR spectroscopy. Although, superior results are obtained from both IHM and PLS-R for ¹H MR-NMR, especially the latter demands for elaborate data pretreatment, whereas IHM models needed no previous alignment.

Full text of DOI:10.1002/mrc.4216

Special issue research article
Received: 22 August 2014
Revised: 31 October 2014
Accepted: 19 January 2015
Published online in Wiley Online Library
(wileyonlinelibrary.com) DOI 10.1002/mrc.4216
Automated data evaluation and modelling of
19 1
simultaneous F– H medium-resolution NMR
spectra for online reaction monitoring
Nicolai Zientek,a* Clément Laurain,a,b Klas Meyer,a Andrea Paul,a Dirk Engel,c
Gisela Guthausen,d Matthias Kraumee and Michael Maiwalda
Medium-resolution nuclear magnetic resonance spectroscopy (MR-NMR) currently develops to an important analytical tool for
both quality control and process monitoring. In contrast to high-resolution online NMR (HR-NMR), MR-NMR can be operated under
rough environmental conditions. A continuous re-circulating stream of reaction mixture from the reaction vessel to the NMR spec-
trometer enables a non-invasive, volume integrating online analysis of reactants and products. Here, we investigate the esterifi-
cation of 2,2,2-trifluoroethanol with acetic acid to 2,2,2-trifluoroethyl acetate both by 1H HR-NMR (500 MHz) and 1H and 19F MR-
NMR (43 MHz) as a model system. The parallel online measurement is realised by splitting the flow, which allows the adjustment of
quantitative and independent flow rates, both in the HR-NMR probe as well as in the MR-NMR probe, in addition to a fast bypass
line back to the reactor. One of the fundamental acceptance criteria for online MR-MNR spectroscopy is a robust data treatment
and evaluation strategy with the potential for automation. The MR-NMR spectra are treated by an automated baseline and phase
correction using the minimum entropy method. The evaluation strategies comprise (i) direct integration, (ii) automated line
fitting, (iii) indirect hard modelling (IHM) and (iv) partial least squares regression (PLS-R). To assess the potential of these evalua-
tion strategies for MR-NMR, prediction results are compared with the line fitting data derived from the quantitative HR-NMR spec-
1
troscopy. Although, superior results are obtained from both IHM and PLS-R for H MR-NMR, especially the latter demands for
elaborate data pretreatment, whereas IHM models needed no previous alignment. Copyright © 2015 John Wiley & Sons, Ltd.
1
Keywords: NMR; H; 19F; medium-resolution NMR; online NMR; quantitative NMR; reaction monitoring; data processing; automation;
process analytical technology; IHM; PLS
strengths are usually below 1.88T (80 MHz proton frequency) for
Introduction
the observation of 1H. Currently 19F and 31P nuclei are (also) observ-
The perfect process analytical method would be based on a robust
and easy to handle customised technique operating in real time
that come without any need for calibration, is an absolute method,
has a professional support, if needed at all, and is compliant to in-
creasing regulatory requirements. Quantitative high-resolution on-
line NMR spectroscopy meets most of these requirements and
has thus been used to gain process understanding in pilot plant
and research environments because it works non-invasively, with-
out sample preparation and without elaborate calibration.[1–3] Be-
cause NMR spectroscopy exhibits strict linearity between absolute
signal area according to the number of spins and sample concen-
tration, it is an absolute analytical comparison method independent
of the matrix and therefore ideal for the investigation of complex
fluid mixtures. Various theoretical and experimental papers have
dealt with quantitative NMR spectroscopy,[4,5] e.g. with respect to
continuous flow[6] in agriculture,[7] for trace components,[8] process
engineering or pharmaceutical applications.[9–14]
Because of stray fields, magnet maintenance and cost, high-
resolution NMR (HR-NMR) spectrometers are not applicable in pilot
plant or production environments, whereas medium-resolution
NMR (MR-NMR) instruments with permanent magnets and small
stray fields are typically compatible for these ‘in-field’ applications.
Initially, Skloss et al.[15] and Nordon et al.[17] demonstrated applica-
tions of MR-NMR spectroscopy for process monitoring.[16–18] Re-
cently, overviews on the current technology were given[19,20] Field
able with MR-NMR instruments with acceptable sensitivity – and
even 13C.
Here, a simple 1/16″ fluorinated ethylene propylene tube with an
ID of 0.04″ (1.02 mm) was used as flow cell. The thermal isolation be-
tween tube and NMR probe was realised with a simple 5 mm glass
Dewar tube inserted into a benchtop MR-NMR spectrometer
(43.32 MHz 1H NMR and 40.68 MHz 19F). The set-up has been
described previously in ref. 21. In the current system using a
*
Correspondence to: Nicolai Zientek, BAM Federal Institute for Materials Research
and Testing, Richard-Willstaetter-Straße 11, 12489 Berlin, Germany. E-mail:
nicolai.zeintek@bam.de
a BAM Federal Institute for Materials Research and Testing, Richard-Willstaetter-
Straße 11, 12489, Berlin, Germany
b École Nationale Supérieure de Chimie de Lille, Avenue Mendeleiev CS 90108,
59652, Villeneuve D’ascq Cedex, France
c
Pro2NMR, Institute of Mechanical Process Engineering and Mechanics and
Institute of Biological Interfaces, KIT, Adenauerring 20b, 76131, Karlsruhe,
Germany
d S-PACT GmbH, Turmstraße 46, 52064, Aachen, Germany
e Department of Chemical Engineering, Technische Universität Berlin, Straße des 17.
Juni 136, MA 5-7, 10623, Berlin, Germany
Magn. Reson. Chem. (2015)
Copyright © 2015 John Wiley & Sons, Ltd.

N. Zientek et al.
two-channel probe, it is not required to change the tuning of the
probe or the shim to measure fluorine and protons consecutively.
Because of its favourable nuclear properties and high abundance,
19F NMR measurements are almost equally fast as 1H NMR experi-
ments (84% relative sensitivity compared with protons). 19F is prob-
ably the third most studied nucleus and is contained in a large
number of organic chemicals, polymers and pharmaceutical prod-
ucts. Furthermore, 19F NMR spectra can also be beneficially used
in reaction monitoring when proton spectra show overlapping be-
cause the range of the chemical shift extends from approx. 0 to
ꢀ200 ppm compared with 0 to 10 ppm for protons and is thus less
likely for signal overlap to occur.
In this study, calibration mixtures in the relevant concentration
1
range of all components were measured by H and 19F MR-NMR
as well as 1H HR-NMR spectroscopy (cf. Fig. 2). Based on the calibra-
tion data, IHM and PLS-R models were built and concentration axes
were scaled. In the next step, reactions were performed and re-
corded by the previously mentioned NMR techniques. The evalua-
tion of spectra included classical techniques as integration and
line fitting as well as chemometric approaches.
In general, the line fit data retrieved from 1H HR-NMR served as
reference data for the validation of the predicted concentrations
of MR-NMR from both IHM and PLS.
In this study, we focus on automated data analysis of the MR-
NMR spectra because process monitoring produces an increasing
quantity of data with demand for both automated signal process-
ing and automated data evaluation. Furthermore, user-introduced
drawbacks of manual data preparation are circumvented by auto-
mated data processing. Even though the vendors of NMR instru-
ment software and of manufacturer-independent spectral data
processing software recognise the need for implementation of au-
tomatic preprocessing procedures, the results are not yet satisfying.
This is particularly true to stand-alone solutions for a series of
spectra.
Experimental
Esterification
In the present work, the 2,2,2-trifluoroethanol (TFE) esterification
with acetic acid (AA) to produce 2,2,2-trifluoroethyl acetate (TFEA)
according to Eqn. (1) was studied. Details are described in Zientek
et al.[21].
CF3CH2OH þ HCOOCH3 ⇌ CF3CH2OOCH3 þ H2O (1)
The most obvious and classically applied approach for spectral
data analysis is integration of lines that can be assigned to the cor-
responding nuclei of the reactants. Another common approach to
reveal signal area of even overlapping signals is line fitting that is
also known as deconvolution or hard modelling. Here, NMR signals
are fitted by analytical functions such as Gaussian, Lorentzian or
Voigt profiles.
Esterification reactions represent favourable model reactions for
kinetic studies on organic reaction systems that are easily influ-
enced by temperature and/or the amount of catalyst.[26]
Here, the reaction was conducted at 35 °C and started by the ad-
dition of 2 ml of sulfuric acid as homogeneous catalyst to a 50ml
freshly prepared solution of TFE and AA of equimolar ratio
(7.54 mollꢀ1).
A more distinguished modification of the hard modelling is indi-
rect hard modelling (IHM).[22,23] IHM considers each line separately
but allows restraint of the manifold. MR-NMR spectroscopy is appli-
cation for IHM.
Online NMR set-up
The flow scheme of the set-up is depicted in Fig. 1. A dosing pump
P1 (HPD Multitherm 200, Bischoff Chromatography, Leonberg,
Germany) was used to transport the sample from the thermostated
reactor C1 (50ml volume) to the NMR probe of both HR and MR de-
vices. Temperature control of the system, i.e. reactor, pump head
and all tubing (PEEK, 0.02″ ID for NMR bypasses and 0.03″ for the re-
cycle line) is maintained by a thermostat (Petit Fleur, Peter Huber
Kältmaschinenbau GmbH, Offenburg, Germany). To prevent dust
and solid precipitate from entering the system, a 10μm filter (F1)
was used. Before entering the NMR spectrometers, the flow was
Because spectral dispersion of MR-NMR spectra is reduced and
the multiplet structure becomes more dominant because of the
comparably low-field strengths, evaluation by chemometric models
may be useful for a multi-component multi-parameter correlation,
which is state of the art for the analysis of optical spectra, in partic-
ular near-infrared (NIR) spectra and also in liquid-state NMR. The
prediction performance of partial least squares regression (PLS-R)
in comparison to classical analysis approaches to both high-field
and MR-NMR spectroscopic data analysis has been tested
previously.[24,25]
Direct line integration and line fitting methods do not require
calibration data, if peak area fractions are analysed and related to
molar fractions. To obtain substance concentrations or mass con-
centrations, peak areas have to be transformed into concentrations
by comparison to reference standards or via calibration. In contrast,
PLS-R and IHM with few exceptions (refer to the succeeding texts)
require calibration data for model building. The IHM method relies
on prior knowledge on reactants and products, i.e. the pure compo-
nent spectra. If only component fractions are derived and the signal
of the analytical method is directly proportional to the amount of
substance, it can likewise be used without calibration. PLS-R models
are built on calibration spectra containing reactants and products
with known composition. The calibration samples should cover
both the relevant concentration range and the matrix of the ex-
pected reaction to provide stable PLS-R models. Changes of reac-
tion conditions such as temperature, pH etc. that may affect line
shapes or peak positions should be included in the stage of model
building, otherwise, the prediction powers of PLS-R is hampered.
Figure 1. Flow scheme. The reaction mixture was pumped through
thermostated PEEK lines and distributed via a four-way valve to the
HR-NMR and MR-NMR flow cells (right and bottom) and a recycle line
(left). Refer to ref. 1,21 for more details.
wileyonlinelibrary.com/journal/mrc
Copyright © 2015 John Wiley & Sons, Ltd.
Magn. Reson. Chem. (2015)

Simultaneous online 19F–1H medium-resolution NMR and data evaluation
split. This allowed quantitative and independent flow rates, both in
the HR-NMR probe (0.1–1.5 ml minꢀ1) as well as in the MR-NMR
probe (0.1–0.5 ml minꢀ1), in addition to a fast bypass line back to
the reactor. The latter was introduced to achieve a rapid sample
transfer from the reactor to the four-way split in the vicinity of the
NMR instruments. The flow rates of the recycling bypass line, the
HR-NMR spectrometer and the MR-NMR spectrometer were ad-
justed by variable back pressure regulators V1 and V2 (P-446, IDEX
Health & Science, Oak Harbor, WA, USA) or a needle valve V3 (P-788,
IDEX Health & Science), respectively.
quantitative flow rate had to be achieved, these slight time differ-
ences between the spectrometers were acceptable compared with
the relatively long reaction time. All figures and calculations were
corrected according to the delay time differences.
Data preparation
Because of the large amount of data, automatic spectral pretreat-
ment is a crucial step in data analysis. After Fourier transformation
of the FID, the spectral baseline and phase have to be adjusted. In
contrast to manual data pretreatment, custom-designed MATLAB
(The MathWorks, Inc., Natick, Massachusetts, USA) scripts provide
the capability for batch processing. Therefore, the baseline of each
spectrum is fitted by a third-order polynomial and subsequently
subtracted from the spectrum. This approach is satisfying because
the proton spectrum usually includes features in the range of 0 to
10ppm, whereby the overall spectral range for fluorine is
115 ppm. Thus, the featureless spectral areas include enough grid
points to estimate the baseline. For automated phase correction,
the entropy minimisation approach from Chen et al.[28] was imple-
mented in a MATLAB module. Results have been presented and are
discussed later.[21]
NMR spectroscopy
The MR-NMR benchtop instrument (Magritek Spinsolve SPA 102,
43 MHz 1H NMR and 40 MHz 19F)[27] is based on a Halbach magnet
design for standard 5 mm NMR tubes, with a magnet temperature
of 28.5°C without active sample temperature control. The system
is operated via a desktop computer and can be triggered and oper-
ated by shell commands (e.g. from LabVIEW™). Spectra are col-
lected locally in a proprietary (common) NMR file format or
transferred via interface.
HR-NMR spectra were acquired using a 500 MHz NMR spectrom-
eter, (Varian) Agilent with a medium pressure broad band flow
probe: 1H/15N–31P, 95 μl, max. 3.0MPa, max. 130 °C having an inner
diameter of 3.4mm.
Calibration
For monitoring the reaction, 60 spectra per hour were collected
over a total reaction time of 3 h. Each spectrum used a single tran-
sient, π/2 excitation pulse (7 μs) and a spectral width of 5000 Hz
with 32k (64k zero-filling) data points. The fluorine π/2 excitation
pulse was 35 μs at 20 000 Hz spectral width. HR-NMR spectra were
acquired with a π/2 excitation pulse of 6.62μs and a spectral width
of 8000 Hz with 32 k (64 k zero-filling) data points.
As the term calibration is often used in different contexts, we start
with an explanation of its meaning in our context. Calibration in
chemometrics refers to the quantitative conversion between two
scales. In quantitative NMR spectroscopy, this typically means con-
version of the analytical signal to the molar concentrations (mol lꢀ1
)
of the analyte. The analytical signal is the peak area of the analyte,
which is obtained from the NMR signal by means of a mathematical
model.
Because NMR spectroscopy is an absolute method, i.e. the abso-
lute peak area is strictly linear to the number of spins, calibration is
classically left out. Instead, molar ratios are calculated from peak
area ratios. These ratios can then be converted to molar concentra-
tions by comparison with internal or external standards[5] or scaled
to 100%. Scaling only applies for closed systems and when all
analytes are observed.
Because the true molar concentrations of the technical mixtures
should be investigated – independent of any possible side reaction
or whether all signals are fully observed in NMR – we preferred the
previously mentioned calibration procedure using the calibration
mixtures in Fig. 2. Because these mixtures do not contain the cata-
lyst sulfuric acid, calibration and process spectra will show different
line shapes and peak positions. These spectral changes are ex-
pected to pose a serious challenge for the applied chemometric
methods. However, for reactive systems, it would be of great advan-
tage to extrapolate experimental data from a limited set of external
calibration samples, which must not contain the catalyst because it
would immediately start the reaction.
MR-NMR flow cell
The thermal isolation of fluid and probe was achieved by the design
of a simple flow cell arrangement conducting a 1/16″ fluorinated
ethylene propylene line with an ID of 0.04″ (1.02 mm) through an
evacuated Dewar tube (HWS, Mainz, Germany) for passive temper-
ature control having an outer diameter of 5.0 mm and a total length
of 400 mm. The polymer tubing provides a seamless and replace-
able line through the whole instrument with high quality and safety
attributes. Performance and operation of the MR-NMR instrument
as well as the flow probe and its quantitative dynamic flow behav-
iour and limits of detection are in detail described in ref. [21].
Residence times[1]
The registration of a change in the reactor by the NMR spectrome-
ter is delayed because of the bypass set-up. The delay parameters
were experimentally determined from the residence time distribu-
tion function that was achieved via a step tracer experiment.
The transfer time ttrans is the time after which the first molecules
of fresh sample cover the distance from the reactor to the active re-
gion and enter it. The delay time tdelay is found as the time after
which the detected tracer concentration has risen from zero to its
stationary value. The time span between ttrans and tdelay is called
dwell time tdwell. The delay time tdelay was determined to 2.5 min
with a dwell time tdwell of 1 min at 0.8mlminꢀ1 flow rate for the
HR-NMR spectrometer. In contrast, the delay time tdelay for the
MR-NMR spectrometer was determined to 6.5min with a dwell time
tdwell of 4 min at 0.3ml minꢀ1 flow rate. Because an optimal
Data analysis methods
The focus of this paper is a survey of four data analysis methods in-
cluding (i) direct integration (DI), (ii) automated line fitting, (iii) IHM
and (iv) PLS-R (cf. Table S3). For DI, the integration limits have to be
set to the 64-fold of the peaks FWHM. Line fitting requires appropri-
ate starting values for each Gaussian–Lorentzian line. PLS-R and
IHM are supervised multivariate methods, which need previous
model building.
Magn. Reson. Chem. (2015)
Copyright © 2015 John Wiley & Sons, Ltd.
wileyonlinelibrary.com/journal/mrc

N. Zientek et al.
The PLS-R models were validated with respect to explained
calibration and validation variance, with the latter obtained after
cross-validation, performed with mean centring inside the valida-
tion segments. The PLS-R models were in addition evaluated with
respect to root-mean-square error (RMSE), a measure in the same
unit as the concentration.[32] PLS-R was carried out with The
Unscrambler 10.3 (CAMO, Norway).
Results
Esterification reaction
The esterification of TFE with AA to TFEA was monitored over a time
span of 3 h. In Fig. 3, spectra of the reacting mixture at time intervals
of 10min are depicted. The proton spectra on the left side are dom-
inated by overlapping peaks of the CH3 groups resulting from the
AA and the resulting ester at 2.19 and 2.25ppm, respectively. The
quartet at 4.08 ppm is assigned to the CH2 group of the alcohol
and shows an overlap with the quartet at 4.63 ppm arising from
the CH2 group of the ester. The singlet at 8 ppm is related to the
OH/H and was not used for the analysis. In contrast to the proton
spectra, no spectral overlap is observed in the fluorine spectra.
The triplet of the CF3 group of the fluorinated alcohol appears at
ꢀ75 ppm while the triplet for the CF3 group for the ester can be
found at ꢀ71.7 ppm.
Figure 2. Representation of calibration mixtures in the system 2,2,2-
trifluoroethanol–acetic acid–2,2,2-trifluoroethyl acetate. Molar ratios and
associated concentrations can be found in the supplementary material.
Indirect hard modelling (IHM)[29]
IHM uses a mechanistic model (hard model) of the measured spec-
trum based on first principles of molecular spectroscopy, i.e. the
model resembles the signal using parameters that correspond to
physical properties like peak positions and concentrations. The full
spectral model of a mixture is a weighted sum of pure component
models, which again are sums of parameterised peak functions
(Gaussian–Lorentzian). While the component weight parameters
are directly related to component concentrations, the peak param-
eters are used to describe the shape of the spectrum. Disturbing
shape variations in the measured spectrum, e.g. because of shifting
peaks, can be accounted for by an adjustment of the corresponding
peak parameters, thus reducing error propagation to the relevant
component weights. With the component weights being (a) robust
against shape changes and (b) highly selective for a certain compo-
nent, the model is capable for extrapolation. For instance, compo-
nent weights can be calibrated using only a few non-reacting
binary mixtures while prediction of concentrations of the full reac-
tive mixture is possible.
Automated data pretreatment and data analysis
The automated minimum entropy phasing method turned out to
be very efficient correcting the phase of a big set of even noisy
NMR spectra in short time. In comparison to the auto-phased spec-
tra using vendor software, an improvement in phase trueness was
1
found. However, H MR-NMR spectra came up with about 2-fold
smaller phase drifts or phase jumps during the course of the reac-
tion compared with the 19F MR-NMR spectra.[21]
Direct integration
Reasonable results are already obtained by DI that is the classic
method of analysis; although especially for broader signals, this
method leads to a signal line width-dependent underestimation
as a result of signal truncation. Because of the overlapping quartets
of the methylene group in the 1H MR-NMR spectra, only the non-
overlapping parts of each quartet were integrated. In addition, be-
cause of overlap of methylene singlet signals for all 1H spectra inte-
gration with 64-fold, the half width was not possible. Thus, the
signals were integrated to the minimum between both. As a conse-
quence, integration of the smaller peak will include peak area of the
bigger one and results in overestimates for the smaller product sig-
nals (cf. Fig. 4).
Creation of the hard model as well as fitting the hard model to
mixture spectra in order to adjust peak parameters and obtain
component weights are fully automated[22,30] and implemented
in software PEAXACT.
Partial least squares regression (PLS-R)
PLS-R uses a linear empirical model that is entirely data-driven (soft
model). The multivariate calibration (regression) correlates NMR
spectra (X) to concentrations of the analytes (Y). PLS-R conducts
data reduction for functionally correlated data by finding compo-
nents from X that are also relevant for Y. It searches for a set of com-
ponents (called latent vectors) that performs a simultaneous
decomposition of X and Y with the constraint that these compo-
nents explain as much as possible of the covariance between X
and Y. Because PLS-R is a linear method, it typically requires addi-
tional latent vectors to model nonlinear such as peak shifts or devi-
ations of line shape that are typical for NMR spectra. Extrapolation
should be avoided when deploying PLS-R because the established
relationship between X and Y is only supported in a region close to
the calibration points.[31]
Line fitting
Alternatively to integration line fitting that is also known as
deconvolution or hard modelling of NMR signals by analytical func-
tions such as a Gaussian–Lorentzian, function or a set of functions
was used (Figs 4–6). In general, NMR processing software provides
line fitting tools with graphical user interfaces for setting the start
values, for single spectra or batch treatment of serial spectra. Prior
knowledge can be exploited, for example, in the form of multiplet
structure and chemical shifts in order to make fitting procedures ro-
bust even in case of overlapping lines.
wileyonlinelibrary.com/journal/mrc
Copyright © 2015 John Wiley & Sons, Ltd.
Magn. Reson. Chem. (2015)

Simultaneous online 19F–1H medium-resolution NMR and data evaluation
Figure 3. 1H and 19F NMR single-scan spectra from an esterification of 2,2,2-trifluoroethanol to 2,2,2-trifluorethyl acetate with acetic acid, acquired with an
43 MHz MR-NMR instrument.[21]
Figure 4. Course of the reaction monitored by MR-NMR and HR-NMR
spectroscopy and data analysis by the direct integration (DI) method after
calibrating signal areas to molar concentrations. The time axis has been
1
19
corrected to the transfer times. (○ H HR-NMR; 1H MR-NMR; □ F MR-
Figure 5. Comparison of MR-NMR spectroscopy analysed by direct
integration after calibrating signal areas to molar concentrations with HR-
◆
NMR; TFE, 2,2,2-trifluoroethanol (blue); AA, acetic acid (green); TFEA, 2,2,2-
1
NMR spectroscopy analysed by line fitting. The time axis has been
trifluoroethyl acetate (red). For comparison: filled ○ H HR-NMR line fit
corrected to the transfer times. ( 1H MR-NMR; □ F MR-NMR; TFE, 2,2,2-
19
data). Balance, summarised signal for each NMR method without OH/H
signal (black); OH/H, summarised signal for each NMR method including
OH/H signal (magenta).
◆
trifluoroethanol (blue); AA, acetic acid (green); TFEA, 2,2,2-trifluoroethyl
acetate (red). For comparison: ○1H HR-NMR line fit data).
Magn. Reson. Chem. (2015)
Copyright © 2015 John Wiley & Sons, Ltd.
wileyonlinelibrary.com/journal/mrc

N. Zientek et al.
Figure 6. Comparison of MR-NMR spectroscopy analysed by line fitting
after calibrating signal areas to molar concentrations with HR-NMR
Figure 7. Comparison of MR-NMR spectroscopy analysed by individual
indirect hard modelling (IHM) in the 1H and 19F domain with HR-NMR
spectroscopy analysed by line fitting. Signal areas were modelled to molar
concentrations within PEAXACT. The time axis has been corrected to the
spectroscopy analysed by line fitting. The time axis has been corrected to
19
the transfer times. ( 1H MR-NMR; □ F MR-NMR; TFE, 2,2,2-trifluoroethanol
◆
1
19
(blue); AA, acetic acid (green); TFEA, 2,2,2-trifluoroethyl acetate (red). For
comparison: ○ H HR-NMR line fit data).
transfer times. (● H MR-NMR; ■ F MR-NMR; TFE, 2,2,2-trifluoroethanol
1
(blue); AA, acetic acid (green); TFEA, 2,2,2-trifluoroethyl acetate (red). For
1
comparison: ○ H HR-NMR line fit data).
During the esterification reaction, TFE and AA-related signals de-
crease, whereas those of TFEA increase to the same extent. Because
water is formed at equimolar ratios during the reaction but not
plotted, the proton balance signals go down because of concentra-
tion changes by water formation on the molar concentration axis.
However, when the OH/H band is included into the balance, a
straight horizontal line is obtained. Any deviations from linearity
would indicate the formation of undetected by-products or alterna-
tively temperature, density and partial excess volume effects. Initia-
tion of the reaction with sulfuric acid can be observed as a decrease
for the reactant and balance signals for all substances present at
this time.
Most remarkable is the low extent of scattering of concentration
data along the reaction propagation for MR-NMR proton, which ex-
hibits only slightly more scattering than HR-NMR signals. This is
considered as a good result for single-scan spectra taken from the
small sample volume. Compared with 1H, scattering in 19F MR-
NMR data is about 2-fold larger that we believe is because of the
special construction of the probe in addition to the physical reasons
mentioned in the preceding texts, including probe tuning and re-
duced sensitivity.
Similar observations are made in the case of line fitting. It is ex-
pected that line fitting would reduce the influence of line overlap,
especially for MR-NMR. The comparison of DI (Fig. 5) and line fit
data (Fig. 6) shows that the scatter of the 19F data is clearly reduced
by line fitting, which might be because of the smaller influence of
phase artefacts in fitting routines compared with integration. How-
ever, an overestimation of the TFE concentration is observed,
whereas that of TFEA determined by the MR-NMR methods coin-
cides within the experimental error. A slight systematic deviation
of MR-NMR from HR-NMR results for TFEA is observed with respect
to the high-field values, which is not found with DI. Altogether, a
strong influence of data processing is observed when it comes to
quantification. In course of this paper, we decided to use the line
fit-derived 1H HR-NMR data as reference for IHM and PLS-R evalua-
tion results.
Figure 8. Comparison of MR-NMR spectroscopy analysed by combined 1H
and 19F indirect hard modelling (IHM) with HR-NMR spectroscopy analysed
by line fitting. Signal areas were modelled to molar concentrations within
PEAXACT. The time axis has been corrected to the transfer times.
(
combined 1H MR-NMR and 19F MR-NMR (in case of AA only 1H MR-
◆
NMR); TFE, 2,2,2-trifluoroethanol (blue); AA, acetic acid (green); TFEA, 2,2,2-
1
trifluoroethyl acetate (red). For comparison: ○ H HR-NMR line fit data).
Indirect hard modelling (IHM)
Comparing the separate IHM and combined model (cf., Figs 7 and 8),
a clear improvement of data scattering is observed with the
latter. This is because of the redundancy of data in the com-
bined model where several signals of each analyte are ‘com-
bined to mean values’ that result from independent physical
methods. The TFEA data compares well to the HR-NMR line fits,
whereas the AA and TFE data exhibit the same offsets as found
with the other methods as well. While the separate IHM
models where created for each block (1H and 19F NMR spectra)
separately, the combined model where created with
a
wileyonlinelibrary.com/journal/mrc
Copyright © 2015 John Wiley & Sons, Ltd.
Magn. Reson. Chem. (2015)

Simultaneous online 19F–1H medium-resolution NMR and data evaluation
horizontal concatenation of the two blocks, without further
scaling. Detailed information of the individual IHM models is
presented in the supplementary materials in Figures S9–S16.
19F MR-NMR allows for an estimation of AA, although no AA-related
lines are present in the spectra at all. The prediction in that case is sim-
ply based on a negative correlation to TFE and TFEA-related lines, as
can be deduced from the weighted correlation coefficients in Fig. S7.
Partial least squares regression (PLS-R)
PLS-R models were calculated after selecting suitable parameter
Discussion
1
1
ranges referred to as region of interest from H HR-NMR, H MR-
NMR and 19F MR-NMR measurements. In one case, a single calibra-
tion sample, which completely deteriorated the PLS-R result, had to
be excluded (cf. Table S2). Reduction of factor number and RMSE of
cross-validation, but also homogeneous scores distribution, coinci-
dence of calibration and validation lines, served as criteria for PLS-R
model optimisation. Regression parameters are summarised in
Table S3, and the individual PLS-R models are provided in the sup-
plementary material in Figures S17–S26. In Fig. 9, the prediction of
MR-NMR data based on of PLS-R is depicted. In the course of model
evaluation, the prediction power of PLS-R models compared with
line fit 1H HR-NMR data could be improved by the inclusion of the
first three-process spectra using the known starting concentration
of the reactants (7.535 mollꢀ1) together with the assumption of
0 mol lꢀ1 for the ester. The observed improvement obviously results
from the fact that now the spectral differences between calibration
and process samples are included into the models. All in all, PLS-R
After appropriate data preparation, such as baseline correction and
phasing, the DI method seems to be applicable to MR-NMR data,
whenever NMR signals have a good dispersion and lines can be in-
tegrated separately. Because of the nature of 1H MR-NMR spectra,
i.e. small dispersion and large multiplet splittings on the ppm scale,
integration of signals in close vicinity with their 64-fold half width as
necessary for a quantitative survey cannot be performed. Therefore,
overestimates must be taken into account for weaker signals close
to stronger ones. Because the integration results strongly depend
on the line width, small uncertainties have to be accepted because
of changes of signal line widths either in the course of a reaction or
between calibration set and technical mixtures. In the case of the
ester reaction, DI yields underestimates for the product TFEA,
whereas the reactants are slightly overestimated.
As long as the underlying line shape function is known, the line
fitting method presents an alternative easy-to-use method for data
analysis of serial MR-NMR spectra. Various software solutions with
line fitting options are commercially available. For a given applica-
tion, macros with all necessary steps can be programmed. Limita-
tions occur for line fitting, for example, when NMR signals fall
completely together and are hence not resolvable for this method.
In case of weak or non-existent product signals, which arise during
the course of the reaction, line fitting suffers from uncertainties
when it comes to low concentrations. An increase in uncertainty
is also observed for NMR lines that differ from their ideal
Gaussian–Lorentzian line shape. However, depending on the nu-
merical optimisation routine used in that case, line fits are iterated
such that residues are reduced for the final fit. Because line distor-
tions, e.g. from improper shimming, are equal for all NMR signals
in a spectrum, these residues produce a larger uncertainty to the re-
sult but do not falsify it severely. In cases of more extended signal
distortions, a set of interconnected curves presents a practical solu-
tion. Software packages like MATLAB allow a relatively flexible ap-
proach in which special functions are available and where prior
knowledge about the reaction can be considered. These ap-
proaches allow a strong reduction of uncertainty and therefore also
of the necessary signal-to-noise ratio required for a stable line fit.
IHM came out with reliable results for the observed reaction,
both with low scattering results – especially for 19F data – as well
as accurate numbers. This is because of the model’s flexibility to ad-
just for peak changes as well as because of the exploitation of re-
dundant information from the combination of all lines of an
individual component, which is the major difference to DI and line
fitting. This effect is even more obvious, when 1H and 19F data are
combined into a single IHM model (cf. Fig. 8). Compared with the
separate IHM model, results are distributed more homogeneously
while accuracy and robustness are improved. All in all, IHM is supe-
rior to simple line fits of individual compounds.
1
of H MR-NMR data provides reliable models, however, only after
elaborated peak alignment of both, calibration and process spectra.
Figure S26D illustrates the low prediction power found with non-
aligned spectra. This could be overcome by including an increasing
number of process spectra into the calibration set as can be seen in
Figures S26A–C. Consequently, in order to prevent time-consuming
alignment procedures, for PLS-R of NMR data, calibration strategies
should use offline samples measured with independent methods,
rather than employ external calibration samples measured under
non-identical conditions.
In contrast to the 1H MR-NMR predictions, evaluation of 19F
MR-NMR spectra yielded less accurate results and overestimation
in prediction of the reactants, whereas the prediction of TFEA pro-
1
vided results comparable to H MR-NMR. It should be noted that
For the observed esterification, PLS-R requires a correct line align-
ment of the calibration as well as of the process data for establish-
ing a robust model. Under this precondition, a reliable prediction of
all compounds of interest was possible by 1H MR-NMR (cf. Table 1).
It should be noted, however, that variations of experimental param-
eters like temperature or pH shifts as well as different matrices,
Figure 9. Course of the reaction based on predictions based on
multivariate data analysis (PLS-R) of MR-NMR 1H and 19F spectra. Signal
areas were modelled to molar concentrations within UNSCRAMBLER. The
1
19
time axis has been corrected to the transfer times. (● H MR-NMR and ■ F
MR-NMR; TFE, 2,2,2-trifluoroethanol (red); AA, acetic acid (black); TFEA,
1
2,2,2-trifluoroethyl acetate (blue). For comparison: ○H HR-NMR line fit data).
Magn. Reson. Chem. (2015)
Copyright © 2015 John Wiley & Sons, Ltd.
wileyonlinelibrary.com/journal/mrc

N. Zientek et al.
[2] M. Maiwald, O. Steinhof, C. J. Sleigh, M. A. Bernstein, H. Hasse, in NMR
Spectroscopy in Pharmaceutical Analysis (Eds: I. Wawer, B. Diehl, U.
Holzgrabe), Elsevier, Amsterdam, 2008, pp. 471–491. DOI: 10.1016/
B978-0-444-53173-5.00020-2
[3] L. F. Gladden. Chem. Eng. J. 1995, 56, 149–158. doi:10.1016/B978-0-444-
53173-5.00020-2.
Table 1. Comparison of root-mean-square errors (RMSEs) calculated by
using line fit-derived data from 1H HR-NMR as a reference for predicted
data by the tested approaches
1H MR-NMR
19F MR-NMR
TFE
AA
TFEA
TFE
TFEA
[4] G. F. Pauli, T. Gödecke, B. U. Jaki, D. C. Lankin. J. Nat. Prod. 2012, 75(4),
834–851. doi:10.1021/np200993k.
(mol lꢀ1
)
(mol lꢀ1
)
(mol lꢀ1
)
(mol lꢀ1
)
(mol lꢀ1
)
[5] S. K. Bharti, R. Roy. Trends Anal. Chem. 2012, 35, 5–26. doi:10.1016/j.
trac.2012.02.007.
[6] J. F. Haw, T. E. Glass, H. C. Dorn. J. Magn. Reson. 1982, 49, 22–31.
doi:10.1016/0022-2364(82)90292-X.
[7] G. Maniara, K. Rajamoorthi, S. Rajan, G. W. Stockton. Anal. Chem. 1998,
70, 4921–4928. doi:10.1021/ac980573i.
DI
0.190
0.153
0.155
0.058
0.143
0.105
0.075
0.061
0.183
0.264
0.170
0.114
0.286
0.622
0.517
0.702
0.171
0.249
5.350
5.543
LF
IHM
PLS-R
[8] T. Nakayama, Y. Fujiwara. Anal. Chem. 1982, 54, 25–28. doi:10.1021/
ac00238a010.
[9] A. M. Wolters, D. A. Jayawickrama, C. K. Larive, J. V. Sweedler. Anal.
Chem. 2002, 74, 4191–4197. doi:10.1021/ac025585p.
[10] M. Maiwald, H. H. Fischer, Y.-K. Kim, H. Hasse. Anal. Bioanal. Chem. 2003,
375, 1111–1115. doi:10.1007/s00216-002-1723-y.
[11] M. A. Bernstein, M. Stefinovic, C. J. Sleigh. Magn. Res. Chem. 2007, 45,
564–571. doi:10.1021/ac100110m.
which affect line form and position, have to be included into the
models by suitable calibration measurements. When these precon-
ditions are not met, prediction results may suffer from high uncer-
tainties. A typical approach to circumvent this problem is binning.[23]
[12] M. Khajeh, A. Botana, M. A. Bernstein, M. Nilsson, G. A. Morris. Anal.
Chem. 2010, 82, 2102–2108. doi:10.1021/ac100110m.
[13] D. A. Foley, C. W. Doecke, J. Y. Buser, J. M. Merritt, L. Murphy, M. Kissane,
S. G. Collins, A. R. Maguire, A. Kaerner. J. Org. Chem. 2011, 76,
9630–9640. doi:10.1021/jo201212p.
[14] J. M. Merritt, J. Y. Buser, A. N. Campbell, J. W. Fennell, N. J. Kallman,
T. M. Koenig, H. Moursy, M. A. Pietz, N. Scully, U. K. Singh. Org. Process
Res. Dev. 2013, 18, 246–256. doi:10.1021/op400226m.
[15] T. W. Skloss, A. J. Kim, J. F. Haw. Anal. Chem. 1994, 66, 536–542.
doi:10.1021/ac00076a019.
[16] C. A. McGill, A. Nordon, D. Littlejohn. Analyst 2002, 127, 287–292.
doi:10.1039/B106889J.
[17] A. Nordon, C. Meunier, R. H. Carr, P. J. Gemperline, D. Littlejohn. Anal.
Chim. Acta 2002, 472, 133–140. DOI: 10.1016/S0003-2670(02)00939-X
[18] A. Nordon, C. A. McGill, D. Littlejohn. Appl. Spectrosc. 2002, 56, 75–82.
doi:10.1366/0003702021954241.
[19] F. Dalitz, M. Cudaj, M. Maiwald, G. Guthausen. Prog. Nucl. Magn. Reson.
Spectrosc. 2012, 60, 52–70. doi:10.1016/j.pnmrs.2011.11.003.
[20] S. S. Zalesskiy, E. Danieli, B. Blümich, V. P. Ananikov. Chem. Rev. 2014,
114, 5641–5694. doi:10.1021/cr400063g.
[21] N. Zientek, C. Laurain, K. Meyer, M. Kraume, G. Guthausen, M. Maiwald.
J. Magn. Reson. 2014, in press. DOI: 10.1016/j.jmr.2014.10.007
[22] E. Kriesten, F. Alsmeyer, A. Bardow, W. Marquardt. Chemometr. Intell.
Lab. 2008, 91, 181–193. doi:10.1016/j.chemolab.2007.11.004.
[23] E. Kriesten, D. Mayer, F. Alsmeyer, C. B. Minnich, L. Greiner,
W. Marquardt. Chemometr. Intell. Lab. 2008, 93, 108–119. doi:10.1016/
j.chemolab.2008.05.002.
Conclusion and outlook
Diverse data processing strategies are known in NMR. Four out of
this variety of mathematical approaches were applied on data ob-
tained during a sample esterification reaction by HR-NMR and
MR-NMR spectroscopy. Similar results in terms of concentrations
were obtained that reflect the reaction progress consistently. All
data analysis methods in general and especially chemometric
models require a constant signal quality over the duration of the ex-
periment, i.e. line shape, symmetry and ideally Gaussian–Lorentzian
shape. Finally, it turns out that the instrumental stability and robust-
ness against environmental influences are important prerequisites
for a successful data analysis and even more important than the
highest potential dispersion, especially in the case of PLS-R. In the
case of IHM, potential developments include the implementation
of prior knowledge of the multiplet structure, flexible line shape
functions and an improvement of the optimisation algorithm. Sim-
ilar arguments hold for the line fitting approach. Line shifts pose
problems and are addressed by suitable line alignment algorithms.
Overlap of lines poses problems especially to DI and to a smaller ex-
tent on line fitting, in which cases IHM and PLS-R can successfully
be applied. A unified data structure, not only for the raw data but
also for switching between modular data preparation and data
analysis tools, is preferable for future developments because it
smoothes the way for automated applications.
[24] T. M. Alam, M. K. Alam. Ann. Rep. NMR Spectrosc. 2005, 54, 41–80.
doi:10.1016/S0066-4103(04)54002-4.
[25] G. Guthausen, A. Garnier, R. Reimert. Appl. Spectrosc. 2009, 63,
1121–1127. doi:10.1366/000370209789553156.
[26] F. Dalitz, L. Kreckel, M. Maiwald, G. Guthausen. Appl. Magn. Reson. 2014,
45, 411–425. doi:10.1007/s00723-014-0522-x.
[27] http://www.magritek.com/products/spinsolve, accessed 23.07.2014
[28] L. Chen, Z. Weng, L. Goh, M. Garland. J. Magn. Reson. 2002, 158,
164–168. doi:10.1016/S1090-7807(02)00069-1.
[29] F. Alsmeyer, H.-J. Koß, W. Marquardt. J. Appl. Spectrosc. 2004, 58(8),
975–986. doi:10.1366/0003702041655368.
[30] F. Alsmeyer, W. Marquardt. Appl. Spectrosc. 2004, 58(8), 986–994.
doi:10.1366/0003702041655421.
Acknowledgements
The authors acknowledge support from Deutsche Forschungs-
gemeinschaft by projects MA 2292/1-1 and GU1123/2-1. Helpful
discussions with Clemens Minnich are also gratefully acknowl-
edged. The authors thank Magritek (Germany and New Zeeland)
for the straightforward support. Further, helpful discussions with
Franz Dalitz are gratefully acknowledged. MM wants to thank Ulrich
Panne for supporting this project. GG also thanks the Deutsche
Forschungsgemeinschaft for financial support of the instrumental
facility Pro2NMR.
[31] N. M. Faber, R. Rajkó. Anal. Chim. Acta 2007, 595, 98–106. doi:10.1016/j.
aca.2007.05.030.
[32] W. Kessler, Multivariate Datenanalyse: für die Pharma-, Bio- und
Prozessanalytik,
Wiley-VCH,
Weinheim,
2007.
doi:10.1002/
9783527610037.
Supporting information
References
Additional supporting information may be found in the online ver-
sion of this article at the publisher’s web site.
[1] M. Maiwald, H. H. Fischer, Y. K. Kim, K. Albert, H. Hasse. J. Magn. Reson.
2004, 166, 135–146. doi:10.1016/j.jmr.2003.09.003.
wileyonlinelibrary.com/journal/mrc
Copyright © 2015 John Wiley & Sons, Ltd.
Magn. Reson. Chem. (2015)