Advertisement

Analytical and Bioanalytical Chemistry

, Volume 410, Issue 14, pp 3349–3360 | Cite as

Online low-field NMR spectroscopy for process control of an industrial lithiation reaction—automated data analysis

  • Simon Kern
  • Klas Meyer
  • Svetlana Guhl
  • Patrick Gräßer
  • Andrea Paul
  • Rudibert King
  • Michael Maiwald
Research Paper

Abstract

Monitoring specific chemical properties is the key to chemical process control. Today, mainly optical online methods are applied, which require time- and cost-intensive calibration effort. NMR spectroscopy, with its advantage being a direct comparison method without need for calibration, has a high potential for enabling closed-loop process control while exhibiting short set-up times. Compact NMR instruments make NMR spectroscopy accessible in industrial and rough environments for process monitoring and advanced process control strategies. We present a fully automated data analysis approach which is completely based on physically motivated spectral models as first principles information (indirect hard modeling—IHM) and applied it to a given pharmaceutical lithiation reaction in the framework of the European Union’s Horizon 2020 project CONSENS. Online low-field NMR (LF NMR) data was analyzed by IHM with low calibration effort, compared to a multivariate PLS-R (partial least squares regression) approach, and both validated using online high-field NMR (HF NMR) spectroscopy.

Graphical abstract

NMR sensor module for monitoring of the aromatic coupling of 1-fluoro-2-nitrobenzene (FNB) with aniline to 2-nitrodiphenylamine (NDPA) using lithium-bis(trimethylsilyl) amide (Li-HMDS) in continuous operation. Online 43.5 MHz low-field NMR (LF) was compared to 500 MHz high-field NMR spectroscopy (HF) as reference method

Keywords

Online NMR spectroscopy Process analytical technology Partial least squares regression Indirect hard modeling Benchtop NMR spectroscopy Smart sensors 

Introduction

Integrated process design using modular plants

Novel discoveries, developments, and concepts in the field of process engineering and in particular process intensification are currently promoted for analysis and design of innovative equipment and processing methods. This leads to substantially improved sustainability, efficiency, environmental performance, and alternative energy conversion. Intensified continuous processes of pharmaceutical products and fine chemicals are particularly in focus of current research [1, 2].

Compared to traditional batch processes, intensified continuous production gives admittance to new and difficult to produce compounds (see reaction Fig. 1 as an example), leads to better product uniformity, and reduces the consumption of raw materials and energy. Flexible (modular) chemical plants can produce various products using the same equipment with short down-times between campaigns, and quick introduction of new products to the market. Typically, such plants have smaller scale than plants for basic chemicals in batch production but still are capable to produce kilograms to tons of specialty products each day.
Fig. 1

Reaction steps during aromatic coupling of aniline (1) and 1-fluoro-2-nitrobenzene (3—o-FNB) induced by lithium bis(trimethylsilyl)amide (referred to as Li-HMDS in the text) and lithiated aniline (2) as intermediate to 2-nitrodiphenylamine (4—NDPA).; tetrahydrofurane (THF); room temperature (RT)

Consequently, full automation is a prerequisite to realize such benefits of intensified continuous plants. Due to the increasing lack of automation personnel new automation concepts such as modular automation are highly desirable in order to support these targets—which is currently addressed in Industrial Internet of Things (IIoT) or Industry 4.0 approaches. In continuous flow processes, continuous, automated measurements and tight closed-loop control of the product quality are mandatory. If these are not available, there is a certain risk of producing large amounts of out-of-spec (OOS) products.

In pharmaceutical production, the common future vision is continuous manufacturing (CM), based on real-time release (RTR), i.e., a risk-based and integrated quality control in each process unit. This will allow for flexible hook-up of smaller production facilities, production transfer towards fully automated facilities, less operator intervention, less down time, and end to end process understanding over product lifecycle, future knowledge, and faster product to market. It is also assumed to significantly reduce the quality control costs within a CM concept at the same time.

Batch and continuous flow reactors

Figure 1 represents the reaction steps for the synthesis, within two aromats are coupled using the lithium base Li-HMDS (lithium bis(trimethylsilyl)amide) [3]. This provides an essential example for a synthesis applied in the pharmaceutical industry, and at the same time, it represents a spectroscopically very difficult system due to the high overlap of (higher order) NMR signals of several chemical components at low concentration level (see Fig. 4). We investigated the aromatic coupling of aniline with 1-fluoro-2-nitrobenzene (o-FNB) forming 2-nitrodiphenylamine (NDPA). The reaction takes place as 5–10 wt% solution in tetrahydrofuran (THF). Deviations from unknown starting material and reactant concentrations together with the precipitation of LiF typically lead to severe fouling and blocking of the modules. Typically, metal organic reactants are difficult to analyze due to the sensitivity to air and moisture. Thus, this example reaction was chosen in CONSENS to develop and validate a compact NMR sensor to maintain an optimal stoichiometry during the full course of the continuous production [4] using advanced process control algorithms. On industrial scale, such a reaction is typically carried out at low reaction temperatures down to − 80 °C in order to enhance selectivity and avoid the boiling of solvents while capturing the reaction heat [5]. Here, we describe both, a classical reaction in a semi-batch reactor as well as a continuous flow reactor system.

Smart compact NMR spectroscopy in process control

Monitoring specific information (such as physico-chemical properties, chemical reactions) is the key to chemical process control [6, 7]. The challenge within the project and its given lithiation reaction was to integrate a commercially available low-field NMR spectrometer from a laboratory application to the full requirements of an automated chemical production environment including robust evaluation of NMR spectral data.

Meanwhile, NIR spectroscopy and increasingly Raman and IR spectroscopy are frequently applied methods for specific process analytical technology (PAT) applications in an industrial environment [1, 8]. The integration of UV/VIS, NIR [9], Raman [10], ATR-IR [11], or NMR spectroscopy [12] for continuous reaction monitoring has even already been demonstrated in microreactors.

Compared to conventional optical spectroscopic methods, NMR spectroscopy offers some unique features for online reaction monitoring [13]. Recently, promising benchtop NMR instruments with acceptable performance compared to higher field strengths came to market (e.g., signal-to-noise ratio = 303 (at 43 MHz) and 23,727 (at 500 MHz) for 1 wt% ethyl benzene in CDCl3 solution [14]) and process-integrated sensors developed on the basis of such laboratory instruments are on their way. Recent reviews were published by Mitchell, Gladden et al. [15], Zalesskiy, Danieli et al. [16], Dalitz, Cudaj et al. [17], Blümich [18], Singh and Blümich [19], and Meyer, Kern et al. [6]. Due to its measurement principle, it is independent from optical sample properties; it yields highly linear signal response and therefore exhibits almost no matrix effects on its linearity. At the same time, structural as well as quantitative information is provided without a necessity of previous calibration. However, a broad range of chemical reaction systems has not been studied via low-field NMR due to signal overlap in spectra.

Especially for robust low-field NMR instruments in the 40 MHz region (with adequate temperature sensitivity to the process environment), quantitative data analysis can be challenging since multiple structures and higher order effects become more dominant, which results in overlapping of multiple signals. However, instruments with field strengths of 60 to 80 MHz are severely temperature sensitive due to their permanent magnet material and are only currently improved. The signal separation issue can be solved by either applying ultrafast 2D NMR methods to separate signals or by the use of multivariate methods based on 1D spectra with the intent of deconvoluting the overlapping signals. Due to the large sensitivity loss for 2D methods [20], this study will focus on the chemometric evaluation of 1D spectra. Additionally, when it comes to spectra acquisition under industrial process conditions, the shape of signals can vary drastically due to nonlinear effects based on temperature, pH value, and the quality of field homogeneity. However, the structural and quantitative information is still present but needs to be extracted by applying appropriate predictive models.

Procedure

We present two approaches for the automated spectra analysis, including (i) indirect hard modeling based on physically motivated spectral models, which were solely derived from pure component NMR spectra and (ii) classical multivariate statistical approach based on PLS-R. The abovementioned lithiation reaction (Fig. 1) was firstly carried out in three independent semi-batch reactions based on a procedure published in Maiwald, Gräßer et al. [7]. Therefore, lithium base was stepwise added to an equimolar reactant mixture of o-FNB and aniline following the reaction coordinate from reagents to products while online NMR spectra were recorded in parallel with an online low-field (43 MHz) NMR spectrometer (with the intention to integrate it into the industrial environment) and an online high-field (500 MHz) NMR spectrometer serving as reference method. Subsequently, the same equipment was adapted to a continuous reaction set-up using a stainless-steel tubular reactor together with programmable syringe pumps. Due to the lack of thermal isolation between the low-field NMR magnet and the flow cell, the reaction temperature was controlled to 28.5 °C to avoid temperature-induced magnetic field drifts upon heat transfer from the sample line.

Low-field NMR data from the semi-batch reaction was used to build the mixture models for IHM in (i) as well as for calibration of the PLS-R model in (ii), with the thread of a strong correlation of the two reactants o-FNB and aniline which both vanish in an equimolar ratio in the course of the reaction. Therefore, slightly different ratios were used. Finally, both models were validated using the continuous reaction data.

Due to fast reaction kinetics, only the steady state of the reaction was investigated using the online NMR set-up. It allows the identification and quantification of reagents, products, and unreacted intermediates. Within this study, there was no intend of kinetic determinations.

Materials and methods

Preliminary investigations on the given reaction

To investigate the intermediates formed during aromatic coupling of aniline and o-FNB, the reaction steps were performed sequentially in small batch reaction of 1–2 mL working volume. Samples for NMR analysis were prepared in 10-mL vials, transferred into standard 5-mm NMR tubes, and measured directly afterwards. The molar ratios of the mixtures and the recorded proton spectra are depicted in the Electronic Supplementary Material (ESM) (Table S1, Fig. S1). No effects could be observed when samples were exposed for a short time to air during the preparation of NMR tubes. However, combining samples containing Li-HMDS with traces of residual water in solvents leads to quenching and formation of the corresponding siloxane species. All in all, Li-HMDS is easily manageable compared to other Li bases and not self-igniting when handled as solution in THF.

Figure 1 illustrates the reaction scheme of the complete reaction network. After adding Li-HMDS to the reactant mixture (1 and 3), a proton exchange between the primary amine and Li-HMDS takes place yielding aniline in its lithiated from 2. This is indicated by the depletion of the corresponding amine signal in the NMR spectrum (Fig. S1). In the second step, the product NDPA 4 is formed via substitution reaction of 2 and 3. Besides the abovementioned reactions, no by-products from further reactions of the lithiated product could be detected. Moreover, LiF is formed as precipitate due to its low solubility in THF [21]. Side reactions (Fig. 1, II) are analogous proton exchange reactions like the initial step between the amine group and lithium, whereas the equilibrium of the respective reaction is completely on the product side. Thus, at least a twofold excess of Li-HMDS is required for the complete conversion of the reagents due to the abovementioned unavoidable side reactions. Conventionally, a 2.15 excess of the lithium base is applied in industrial scale to compensate uncertainties in stochiometric information of the reagents as well as unknown exact concentration of the lithium base at hand due to uncertainties of the analysis certificate and/or depletion by air and moisture. The economic feasibility of the whole process is considerably reduced by this conventional approach.

Experimental set-up for semi-batch reactions

The flow scheme of the online batch reactor set-up with hyphenation to both online NMR spectrometers was adopted from previous studies [13, 14] and is depicted in Fig. 2. For the semi-batch reaction, 1 mol L−1 Li-HMDS was added stepwise up to a 2.5-fold excess to a solution of aniline and o-FNB in THF. The starting concentration of the reactants (0.75 mol L−1) was determined by sample weighting. The reaction mixture was continually circulated though the by-pass system for 10 min after each addition step (cf., Fig. 2). The reaction mixture was circulated between the reactor (50 mL) and the NMR spectrometers using a dosing pump (P2, HPD Multitherm 200, Bischoff Chromatography, Leonberg, Germany). To guarantee fully quantitative spectra acquisition while circulating the mixture, the flowrate was set to 2.5 and 0.3 mL min−1 for LF NMR (flow cell ID = 4 mm) and HF NMR (flow cell ID = 1 mm), respectively. These flow rates were determined in previous relaxation time measurements and magnetization studies [13, 14]. The flow rate to the HF NMR spectrometer was controlled by a Coriolis mass flow controller (FIC, mini Cori-Flow, Bronkhorst High-Tech B.V., Ruurlo, The Netherlands). The reactor as well as the tubing was temperature controlled by a thermostat to 28.5 °C to LF NMR magnet temperature. To prevent solid impurities from entering the tubing system, a 15-μm filter (F1) was installed (FISS-FL2-15, FITOK GmbH, Germany) at the exit of the reactor behind the pump.
Fig. 2

Scheme of the semi-batch NMR set-up including a by-pass system with 1/16″ PFA or 5-mm PTFE tubing. Pump P2 circulates the reaction mixture in a fast loop with a flow rate of 2.5 mL min1, whereas the mass flow controller (FIC) regulates the flow to 0.3 mL min−1 for quantitative measurements with high-field NMR spectroscopy as reference method

Experimental set-up for the continuous reactions

To demonstrate the robustness and flexibility of data analysis approaches for online spectra, a continuous set-up was implemented (cf., Fig. 3). The reactants aniline and o-FNB (0.75 mol L−1) were dosed by a syringe pump (P1 and P2, Gemini 88, KD Scientific, Holliston, USA), premixed and subsequently mixed with 1 mol L−1 Li-HMDS (P3, Nemesys high pressure syringe pump, Cetoni, Korbussen, Germany; with modified sealings). The overall flowrate for the tubular reactor was kept in a range between 0.9 and 1.1 mL min−1. The volume of total tubing amounted 14.3 mL yielding a mean delay time of 6.3 and 13.2 min for the measurement of LF and HF spectrometers, respectively, which were determined based on step tracer experiments. The flow cell of the LF NMR spectrometer, consisting of a simple PTFE tubing with 4-mm inner diameter, contributes with 8 mL most to the tubing volume and thus to the delay time in the NMR measurements. Various flowrates were modulated during synthesis according to Table S2 in the ESM. For stochiometric conversion, the ratio of aniline, o-FNB, and Li-HMDS was set to 1:1:2.15 for reasons mentioned above.
Fig. 3

Scheme of the validation set-up for monitoring of the continuous reaction unit with the compact NMR sensor (orange box). The lithiation reaction (Fig. 1) is continuously carried out in a thermostated 1/8″ tubular reactor using syringe pumps. HF NMR spectroscopy (upper right) served as reference

Chemicals

The chemicals aniline (Merck, ≥ 99%), 1-fluoro-2-nitrobenzene (Sigma Aldrich, 99%), 2-nitrodiphenylamine (Fluorochem, 98%,) and lithium bis(trimethylsilyl)amide (Sigma Aldrich, 1 mol L−1 in THF) substances were used without prior purification. Other sources of lithium bis(trimethylsilyl)amide can contain considerable amounts of stabilizers, such as nitrobenzene. However, the reagent used here was pure from stabilizers. Tetrahydrofurane (Chemsolute, > 99.9%) was dried over molecular sieve (Sigma Aldrich, 3A) before use.

NMR data acquisition

HR NMR spectra were acquired as reference spectra using a 500 MHz NMR spectrometer (Varian) with a medium pressure broad band flow probe: 1H/15N–31P. For these experiments, the glass flow cell was replaced by a 1 mm ID FEP tubing serving as sample cell. Online proton spectra during reaction monitoring were acquired with two pulses, 45° pulses, 5 s acquisition time, and 15 s relaxation delay.

The NMR benchtop instrument (Spinsolve Fluorine, Magritek, Aachen, Germany) operating at 43.32 MHz 1H frequency, currently built with 5 mm ID bore for standard NMR tubes at a magnet temperature of 28.5 °C without active sample temperature control. The system is operated with a desktop computer and can be triggered and operated via shell commands (e.g., from LABVIEW). Spectra are collected locally in a proprietary (common) NMR file format or transferred via a software interface. For the studies presented here, a 5 mm OD (4 mm ID) PTFE tubing served as NMR flow cell. The tubing was adopted outside the magnet to 1/16″ tubing (1 mm ID) as described above. Online proton spectra were acquired with single scans, a 90° pulse, 6.5 s acquisition time, and 15 s repetition time.

Spectra processing

The acquired proton spectra were processed in MATLAB (R2017a, The Mathworks, München, Germany). The free induction decay (FID) was zero-filled to 64 k data points and subsequently apodized by exponential multiplication with a line broadening factor of 0.5 Hz. After Fourier transformation, spectra were immediately treated by automated data preparation methods, implemented in MATLAB. These are baseline correction [22], phasing [23], and spectral alignment to a mean spectrum using the icoshift algorithm [24].

Evaluation of high-field NMR data as reference method

Despite the fact that most signals are fairly separated in the aromatic region of those spectra, deviations from mass balance to high-field NMR results were obtained in previous experiments by using simple integration method. These deviations stem from minor overlapping of signals in the aromatic regions (ESM Fig. S1). Consequently, processed high-field NMR spectra were evaluated using the indirect hard modeling (IHM) approach. IHM has already been demonstrated in literature as a suitable tool for high-field NMR data analysis [25]. Processed low-field spectra as shown in Fig. 4b were used for prior quantification of reactants using IHM as well as multivariate methods (PLS-R).
Fig. 4

a Full 43 MHz NMR spectrum after data preparation over the course of the reaction. The spectrum mainly shows solvent signals (THF) and CH3 signals from Li-HMDF. b Enlarged aromatic region as used for the data analysis via IHM and PLS-R

The NMR data (high-field and low-field) can be used in two forms as done in this study. NMR peak area fractions can either be related to amount of substance concentrations using a one-point calibration for either of the analytes (using neat analytes or samples with known concentrations) leading to a concentration to peak area conversion factor (c.f., [25]). Since NMR spectroscopy (with proper acquisitions and flow parameters) is an absolute comparison method, only the number of protons must be known. For the same reason, relative amount of substance fractions (mole fractions) can directly be derived from the peak area fractions without any concentration conversion factor. Accordingly, IHM and PLS-R model predictions can be treated in both scales.

Both scales are depicted in Fig. 5 for one of the batch reactions. However, molar concentrations are preferable for many reasons. The determination of molar concentrations enables the calculation of mass balances for the observed nuclei and monitoring sums of signals over the course of a process for plausibility checks, occurrence of unknown non-observed by-products or precipitations, dilutions [14]. Such effects cannot be detected in molar ratio plots.
Fig. 5

Comparison of “steady state” and “transient state” NMR results for NDPA synthesis in batch mode calculated in concentration (a) and molar ratios (b). Black lines and dots represent all underlying data, while colored symbols highlight ‘steady state’ values. The concave and convex curvature of the progression of concentrations and molar ratios are due to dilution effects in the batch mode

Steady-state classification

If the process being monitored is not in steady state or the transport and hold-up of fluids to both spectrometers and thus the residence time distribution functions of low-field (LF) and high-field (HF) NMR spectra are diverging, direct time-corrected comparison of these results gets challenging—as can be seen in Fig. 5. This will lead to defective calibrations of validations. Peaks in HF NMR data (black lines) after Li-HMDS injections in Fig. 5 are caused by back-mixing effects. These injections affect the reaction mixture in the batch reactor, which is then re-diluted by remaining mixture in the lines.

Therefore, all evaluated concertation values we classified in steady state and transient state. In a first step, for each timestamp of LF NMR spectra, a corresponding HF NMR result was assigned by nearest neighbor interpolation, since both spectrometers did not acquire spectra at the same time. Afterwards, the corresponding steady state could be detected for both spectrometers using the amount of substance concentration values (derived from IHM based on the abovementioned one-point calibration). Moving linear regression lines were calculated with a frame length of 27 data points for each reactant and spectrometer. An acquired spectrum was accepted as steady state if all predicted concentrations showed slopes below 0.1 mol h−1 for its linear regression lines. In Fig. 5, this data classification is depicted using colored fillings for steady states whereas transient states are black. Mostly, stable concertation values, i.e., steady states, are reached after approximately 10 min in most cases, however, also showing exceptions. Later, for the calibration or model building runs as well as the validation runs, only steady state data is used.

Indirect hard modeling

IHM makes use of first principles information within the NMR spectrum, i.e., the direct physical nature of the NMR signals (“counting nuclear spins”). In an ideal case, each transition in a pulsed NMR experiment will be represented by a Lorentzian lineshape in the frequency domain. This lineshape is broadened by instruments functions such as the homogeneity distribution of the magnetic and electro-magnetic field or coupling effects (such as dipolar coupling), which typically is described as a Gauß function, broadening the Lorentzian. Each measured spectrum of a chemical mixture can be modeled as sums of peak functions by adjusting the parameters of each peak function (peak fitting) [26]. Hereby, the experimental spectrum can be deconvoluted into its pure component spectra, i.e., Lorentz-Gauß peaks, or Voigt profiles. Subsequently to peak fitting, concentrations or molar rations of unknown samples can be disaggregated using the previously established pure component hard models. Therefore, global residuals between experimental data and the IHM component fits are minimized. IHM particularly works for overlapping NMR spectra with peak area ratios behaving proportionally to molar ratios and exhibits a calibration-free access to quantitative data analysis [25].

Indirect hard modeling was carried out on a standard personal computer using the software PEAXACT 4 (S-PACT, Aachen, Germany). During the present work, the following workflow was conducted for quantification of high- and low-field NMR spectra (see also Fig. 7).
  1. i.

    Acquisition of pure component NMR spectra: Proton NMR spectra were acquired of neat reactants of known molar concentration dissolved in THF.

     
  2. ii.

    Generation of pure component models: Peak fitting with 12–21 peaks per pure component was conducted to minimize the residuals to the appropriate limit. Peaks were added empirically and stepwise to the model until residuals were below at least one order of magnitude compared to largest peak. To avoid relatively broad peak functions, the maximum half width was constrained to 0.15 ppm.

     
  3. iii.

    Generation of mixture model: A weighted sum of each pure component model represents a mixture model, including flexible but constrained peak parameters. Additionally, a baseline function can be included in the mixture model but was not used in the case due to satisfactory performance of preceding data processing.

     
  4. iv.

    Component fitting: For the prediction of component areas in each measured mixture spectrum, the mixture model must be fitted to the spectrum by minimizing the spectral residuals. To reduce the degrees of freedom of the mixture model, fitting options and parameter constraints were defined. By this, large calculation times and the risk of overfitting are minimized. During component fitting process, weights and component shifts (shift of ppm axis for entire component) as well as entire peak parameters for typically 20 selected peaks were adjusted. Those peaks were considered by the fitting algorithm based on their influence on improving the fit. Constraints for component shift (0.03 ppm) and single peak shifts (0.015 ppm) seemed to yield sufficient quality of the fit. Figure 6 depicts the result of a component fit for an online spectrum during NDPA synthesis.

     
  5. v.

    Area conversion: The respective area for the pure components can be derived from the parameters of the peak functions (pseudo-Voigt). After normalization to number of protons for each pure component area, molar ratios can be calculated without prior calibration while calculation of concentration requires simple one-point calibration, as mentioned above [27].

     
Fig. 6

Determination of pure component areas with indirect hard modeling (PEAXACT, S-PACT). Adaption of spectral model to aromatic region of measured mixture spectrum during NDPA synthesis by three pure component models (aniline, o-FNB, NDPA). The light blue lines represent peak functions for each pure component model. The residuals represent deviation from mixture model to experimental spectrum

Multivariate data analysis

Mean-centered NMR spectra as obtained after spectral processing as described in the previous paragraph was evaluated by partial least squares regression (PLS-R) using the software The Unscrambler X (Version 10.4, CAMO Software, Oslo, Norway). When in PLS-R the response variable y is present as a matrix (Y), the method is referred to as PLS2. For the calibration and validation of the PLS-R model, 540 spectra from three independent batch reactions each with 1837 spectral variables representing the range of aromatic protons were evaluated. For model building, the unit vector transformed variables were used without further mathematical treatments. The terms RMSEC, RMSE-CV, and RMSEP denote whether the root mean square error is estimated from calibration, cross-validation, or external (test-set) validation [28].

Results and discussion

Two different data evaluation methods for the low-field NMR spectra were compared, whereby the high-field NMR data served as reference method. The major finding was that both modeling approaches—PLS-R and IHM—yield very similar results in terms of correct concentration-time profiles and prediction uncertainties (normally referred to as standard deviations of the method), even though based on completely different principles.

Figure 7 schematically shows the assignment of experimental NMR data to first calibrate and later validate the PLS-R model as well as first build the IHM pure component models, adjust the IHM mixture models according to its behavior along the reaction coordinate and finally validate these IHM models with batch data and continuous reaction data.
Fig. 7

Assignment of experimental NMR spectra from pure substances, three semi-batch reactions along the reaction coordinate, and eight continuous reaction runs were used in this work for PLS-R and IHM data analysis

Indirect hard modeling

Based on pure component spectra and solely derived from preliminary test reactions, which were measured in standard NMR tubes, quantitative results were already achieved using the indirect hard modeling approach. For both HF and LF NMR spectra, area to concentration conversion, i.e., a one-point calibration throughout initial concentrations of o-FNB was performed and used for all further data analysis. The results are depicted in a parity plot in Fig. 8 showing good agreement of both methods with root mean square errors (RSME) in the range of 13–16 mmol L−1 and conventional limits of detection in the range of 35–43 mmol L−1 for IHM (c.f., Tables 1 and 2). At this point it should be mentioned that a process analytical method is typically evaluated upon its RMSE values, which represent the short time reproducibility of the instrument along with the data evaluation method. Therefore, process control decisions and reaction control can be based on the progression and trends of numerous data points (or their floating mean values). In contrast to that, conventional limits of detection as well as limits of quantitation reflect the accuracy of a single or less frequently sampled data point using this method. Further details can be found in the ESM. Regarding the IHM results, it should be emphasized that Fig. 8 (I) highlights a comparison of LF and HF NMR data without any calibration related to each other.
Fig. 8

Parity plot of low-field (LF) NMR versus high-field (HF) NMR results in steady state during NDPA synthesis in batch mode (validation data). The gray line represents angle bisector and residuals represent deviations in mmol L−1 from the reference (HF). Histograms of concentration residuals, which are displayed at the right side of each residual plot were generated using 40 bins. KS Kennard Stone. See also Fig. S6 in the ESM

Table 1

Comparison of root mean square errors (RMSE) of cross validation (CV) and external validation for different chemometric models based on batch and continuous process data. KS Kennard Stone

Model

Samples

Validation

Aniline

o-FNB

Li-NDPA

Batch reactions: concentrations

/mmol L−1

/mmol L−1

/mmol L−1

 PLS2

One batch

Random CV

6

6

4

 PLS2

Three batches

Random CV

18

14

12

 PLS2

Three batches

Batch-wise CV

27

18

16

 PLS-R KS

Three batches

External validation

11

13

8

 IHM

Three batches

External validation

16

13

14

Continuous reactions: molar ratios

/ -

/ -

/ -

 PLS

 

External validation

0.0182

0.0182

0.0287

 IHM

 

External validation

0.0247

0.0191

0.0358

Table 2

Limit of detection (LOD) of final data evaluation models for batch reactions (concentration data). LOD was obtained from linear regression line (ax + b) and standard deviation of low-field NMR data (Sa). Sa was estimated by the standard deviation of y-intercepts of regression lines (see Fig. 8). LOD was expressed as LOD = 3 Sa / a. For IHM all batch data was used for validation and LOD determination. For PLS-R batch data was split in calibration and validation with each 200 spectra, while only the validation data was used for the LOD determination. KS Kennard Stone

Model

Aniline

o-FNB

Li-NDPA

Batch reactions: concentrations

/mmol L−1

/mmol L−1

/mmol L−1

PLS-R KS

Three batches

External validation

33

40

24

IHM

Three batches

External validation

42

35

43

Succeeding, for external validation of this method, spectra from the continuous process were evaluated as shown in Fig. 9. Here, molar ratios were calculated without prior calibration based on normalized (i.e., considering proton numbers per reactant) peak area ratios due to the lack of calibration data. Figure 9 indicates similar results for HF and NF NMR data evaluation. However, when aniline is not present in the mixture, the predictions of IHM show a biased deviation in between 0 and 30 mmol L−1, i.e., the model produces slight overestimates. This is presumably due to the applied component fitting algorithm, when aniline is deployed by the IHM model even though it is not present while minimizing the global residues upon spectral noise while allowing flexibility of the pure component models. Such an over-estimation by IHM was only found for aniline, which aromatic NMR signals completely fall together with other components. In contrary, small area fractions, which belong to NDPA, are erroneously represented by the aniline model. This issue becomes clear when examining the histograms of residual concentration for aniline and NDPA (Fig. 8). For both analytes, a deviation from expected normal distribution can be observed. For aniline, this effect became predominant for concentrations below 40 mmol L−1, whereas for o-FNB it was never problematic. Figure 6 highlights this problem since o-FNB shows peaks in a region where NDPA is not predominantly present (7.4–7.8 ppm), while aniline is mainly covered by NDPA.
Fig. 9

Prediction of continuous processes based on PLS-R and IHM models (external validation) and low-field NMR spectra. To compare the predicted concentrations of aniline, o-FNB, and Li-NDPA with the reference data, ratios of the analytes summing up to 1 were calculated. Gray areas represent points in time where pumps were not running due to refilling of syringe pumps. Residuals were calculated as difference between high-field NMR results and low-field NMR results (IHM or PLS-R)

At first glance, the subliminal coupling of overestimates of aniline together with underestimates of NDPA are incomprehensible since IHM pure component models for these components are strictly independent from each other. However, it can be explained by the closing condition of the IHM algorithm, which minimizes the global residues.

Multivariate data analysis

In a first approach, PLS2 models using a matrix of the reference data (HF NMR concentrations) for all three reactants aniline, o-FNB, and Li-NDPA as response variables were established. PLS2 of a single batch with random cross validation using 175 spectra in 20 randomly chosen segments yielded RMSEC and RMSE-CV between 4 mmol L−1 for Li-NDPA, and 6 mmol L−1 for aniline and o-FNB, in a calibration range of 0–400 mmol L−1. Including two more batch experiments, however, increased RMSE-CV. Systematic cross validation using the batches as validation segments revealed that cross-batch differences account for the majority of variance which probably originates from the effect of different shims on the spectral shape. The PLS2 approach in general does not allow for an independent modeling of the educts, i.e., predictions for aniline and o-FNB always provided similar concentrations for both analytes (not shown). This problem is caused by the experimental set-up with aniline and o-FNB concentration decreasing for the same amount after each addition step of Li-HMDS. To overcome this problem, individual PLS-R models for the three analytes were established keeping the original spectral range for Li-NDPA, but specifying a chemical shift range in the NMR spectra of 6.252–6.573 ppm for aniline and 7.396–7.738 ppm for o-FNB. Furthermore, a software implemented Kennard-Stone algorithm was used to select independent calibration and test sets, each 200 spectra, which could be used for validation. Briefly, the Kennard-Stone algorithm tries to concentrate as much of the diversity in the original data set into the calibration and validation data [29]. As a result, RMSE was diminished and became similar to the values obtained from evaluation of a single batch. Finally, these models could be successfully applied for the prediction of data from the continuous process as depicted in Fig. 9. During continuous process, contrary to semi-batch, the reactants aniline and o-FNB are varying independently. The scores, loadings, and regression coefficients of the PLS-R models used for the prediction shown in Fig. 8 are provided in the ESM in Fig. S2a–c. For all models, two factors are covering altogether 97–99% of the spectral variance.

As can be seen from Fig. 9, the predicted ratios of aniline, o-FNB, and Li-NDPA fit well with the reference data. (False) negative concentrations are only very rarely predicted, different levels of the reagents can be differentiated, and the models predict rapid changes within the mixture instantaneously and hence exhibit appropriate dynamics.

Conclusion

The online monitoring of continuous NDPA synthesis with low-field NMR spectroscopy was demonstrated in lab-scale for various process conditions. For data processing of NMR raw data, publicly available MATLAB scripts were adopted, which showed consistent results in previous studies [14, 25, 30]. Due to severely overlapping signals in the aromatic region of the spectrum two multivariate models were applied—conventional PLS-R and IHM, based on spectral models—to quantify the reactants and products. Both methods yielded consistent results while comparing to the high-field NMR method even though based on completely different principles. The determination of limits of detection (LOD) for this specific application returned values in a range of 24–40 mmol L−1 for single scan proton spectra with an acquisition time of 6.5 s (Table 1).

Whereas in the conventional PLS-R approach, a set of three batch reactions were needed for model building (i.e., calibration) and the same data sets were already used for both, building (i.e., model building and model parameterization) and validation of the IHM mixture model. In addition, a considerable number of continuous experiments were performed for validation taking account for various reaction conditions by individually adjusting the flow rates of the reagents aniline, o-FNB, and Li-HMDS. This data was used to verify the PLS-R model and to check the transferability of the IHM model. The assignment of experimental NMR data is schematically represented in Fig. 7.

Large product concentrations combined with a residual concentration of aniline below a few tens of mmol L−1 was found to be especially challenging for the data analysis due to almost complete overlapping of the spectral signals. As an implication, uncertainties of one component induce further uncertainties in the remaining components since IHM is based on minimizing global residuals while deconvoluting the whole aromatic region of the spectrum into pure components. This fact explains the slightly higher RMSE values for calibration of aniline and Li-NDPA (i.e., 16 and 14 mmol L−1) compared to PLS-R (i.e., 11 and 8 mmol L−1) while RMSE values for o-FNB are comparable. It can be concluded that IHM needs further constraints to find global minima in the mentioned challenging concentration regimes, which is a possible starting point for further improvement.

At the moment, for both PLS-R and IHM methods, input from experienced users is required. For PLS-R especially data pre-treatment (e.g., differentiation, smoothing and normalization), the selection of variables (relevant ppm range) and the selection of validation method require most work intensive user input. Regarding IHM, the peak fitting of pure component spectra as well as optimizing the peak constraints for the mixture model by model parameterization mostly affect the quality of the model. Currently, peak fitting is conducted semi-automatically, requiring the user to empirically add peaks to the spectrum. However, simulations of NMR spectra based on spin calculations can describe each peak in a spectrum, but lack in precision (e.g., due to solvent effects) for the here intended application. Additionally, the optimization of peak constrains based on reference data (i.e., concentrations) could reduce the user input. NMR specific information is currently not used for component fitting, e.g., if peaks get broader due to decreasing field homogeneity (deteriorate shimming) all peaks are affected in same manner. Consequently, this effect can be compensated for by one parameter per pure component instead of fitting each peak width individually.

Nevertheless, advantages of IHM include a minimum calibration effort (i.e., pure component spectra recorded in standard NMR tubes and one-point calibration was sufficient) and high extrapolation capabilities, e.g., for broader lines if shimming is an issue. Additionally, model adaption to new components such as byproducts, intermediates, or similar starting materials do not require complete new model development, in contrast, model constraints and common pure component models can be reused. This somehow approaches a kind of “calibration-free” method, which is highly demanded for process analytical applications. For demonstration of this concept, the continuous set-up needs to be improved for high throughput facing low delay time and narrow residence time distribution and therefore fast gain of steady states.

Notes

Acknowledgements

The authors thank Lukas Wander for his help plotting Fig. 3.

Funding information

This study was supported by the funding of CONSENS by the European Union’s Horizon 2020 research and innovation program under grant agreement No. 636942.

Compliance with ethical standards

Conflict of interest

The authors declare that they have no conflict of interest.

Supplementary material

216_2018_1020_MOESM1_ESM.pdf (977 kb)
ESM 1 (PDF 976 kb)

References

  1. 1.
    Wiles C, Watts P. Continuous flow reactors: a perspective. Green Chem. 2012;14(1):38.CrossRefGoogle Scholar
  2. 2.
    Adamo A, Beingessner RL, Behnam M, Chen J, Jamison TF, Jensen KF, et al. On-demand continuous-flow production of pharmaceuticals in a compact, reconfigurable system. Science. 2016;352(6281):61.CrossRefPubMedGoogle Scholar
  3. 3.
    Zhang P, Terefenko EA, McComas CC, Mahaney PE, Vu A, Trybulski E, et al. Synthesis and activity of novel 1- or 3-(3-amino-1-phenyl propyl)-1,3-dihydro-2H-benzimidazol-2-ones as selective norepinephrine reuptake inhibitors. Bioorg Med Chem Lett. 2008;18(23):6067.CrossRefPubMedGoogle Scholar
  4. 4.
    Bieringer T, Buchholz S, Kockmann N. Future production concepts in the chemical industry: modular–small-scale–continuous. Chem Eng Technol. 2013;36(6):900.CrossRefGoogle Scholar
  5. 5.
    Chinnusamy T, Yudha SS, Hager M, Kreitmeier P, Reiser O. Application of metal-based reagents and catalysts in microstructured flow devices. ChemSusChem. 2012;5(2):247.CrossRefPubMedGoogle Scholar
  6. 6.
    Meyer K, Kern S, Zientek N, Guthausen G, Maiwald M. Process control with compact NMR. TrAC Trends Anal Chem. 2016;83(Part A):39.CrossRefGoogle Scholar
  7. 7.
    Maiwald M, Gräßer P, Wander L, Zientek N, Guhl S, Meyer K, et al. Strangers in the night—smart process sensors in our current automation landscape. PRO. 2017;1:628.  https://doi.org/10.3390/proceedings1040628.CrossRefGoogle Scholar
  8. 8.
    Edwards JC, Giammatteo PJ. In: Bakeev KA, editor. Process Analytical Technology. Hoboken: John Wiley & Sons, Ltd; 2010.  https://doi.org/10.1002/9780470689592.ch10.CrossRefGoogle Scholar
  9. 9.
    Ferstl W, Klahn T, Schweikert W, Billeb G, Schwarzer M, Loebbecke S. Inline analysis in microreaction technology: a suitable tool for process screening and optimization. Chem Eng Technol. 2007;30(3):370.CrossRefGoogle Scholar
  10. 10.
    Leung S-A, Winkle RF, Wootton RCR, de Mello AJ. A method for rapid reaction optimisation in continuous-flow microfluidic reactors using online Raman spectroscopic detection. Analyst. 2005;130(1):46.CrossRefPubMedGoogle Scholar
  11. 11.
    Floyd TM, Schmidt MA, Jensen KF. Silicon micromixers with infrared detection for studies of liquid-phase reactions. Ind Eng Chem Res. 2005;44(8):2351.CrossRefGoogle Scholar
  12. 12.
    Markley JL. NMR analysis goes nano. Nat Biotechnol. 2007;25(7):750.CrossRefPubMedGoogle Scholar
  13. 13.
    Maiwald M, Fischer HH, Kim Y-K, Albert K, Hasse H. Quantitative high-resolution on-line NMR spectroscopy in reaction and process monitoring. J Magn Reson. 2004;166(2):135.CrossRefPubMedGoogle Scholar
  14. 14.
    Zientek N, Laurain C, Meyer K, Kraume M, Guthausen G, Maiwald M. Simultaneous 19F–1H medium resolution NMR spectroscopy for online reaction monitoring. J Magn Reson. 2014;249:53.CrossRefPubMedGoogle Scholar
  15. 15.
    Mitchell J, Gladden LF, Chandrasekera TC, Fordham EJ. Low-field permanent magnets for industrial process and quality control. Prog Nucl Magn Reson Spectrosc. 2014;76:1.CrossRefPubMedGoogle Scholar
  16. 16.
    Zalesskiy SS, Danieli E, Blümich B, Ananikov VP. Miniaturization of NMR systems: desktop spectrometers, microcoil spectroscopy, and “NMR on a Chip” for chemistry, biochemistry, and industry. Chem Rev. 2014;114(11):5641.CrossRefPubMedGoogle Scholar
  17. 17.
    Dalitz F, Cudaj M, Maiwald M, Guthausen G. Process and reaction monitoring by low-field NMR spectroscopy. Prog Nucl Magn Reson Spectrosc. 2012;60:52.CrossRefPubMedGoogle Scholar
  18. 18.
    Blümich B. Introduction to compact NMR: a review of methods. TrAC Trends Anal Chem. 2016;83(Part A):2.CrossRefGoogle Scholar
  19. 19.
    Singh K, Blümich B. NMR spectroscopy with compact instruments. TrAC Trends Anal Chem. 2016;83(Part A):12.CrossRefGoogle Scholar
  20. 20.
    Gouilleux B, Charrier B, Danieli E, Dumez J-N, Akoka S, Felpin F-X, et al. Real-time reaction monitoring by ultrafast 2D NMR on a benchtop spectrometer. Analyst. 2015;140(23):7854.CrossRefPubMedGoogle Scholar
  21. 21.
    Wynn DA, Roth MM, Pollard BD. The solubility of alkali-metal fluorides in non-aqueous solvents with and without crown ethers, as determined by flame emission spectrometry. Talanta. 1984;31(11):1036.CrossRefPubMedGoogle Scholar
  22. 22.
    Mazet V, Carteret C, Brie D, Idier J, Humbert B. Background removal from spectra by designing and minimising a non-quadratic cost function. Chemom Intell Lab Syst. 2005;76(2):121.CrossRefGoogle Scholar
  23. 23.
    Chen L, Weng ZQ, Goh LY, Garland M. An efficient algorithm for automatic phase correction of NMR spectra based on entropy minimization. J Magn Reson. 2002;158(1–2):164.CrossRefGoogle Scholar
  24. 24.
    Savorani F, Tomasi G, Engelsen SB. icoshift: a versatile tool for the rapid alignment of 1D NMR spectra. J Magn Reson. 2010;202(2):190.CrossRefPubMedGoogle Scholar
  25. 25.
    Michalik-Onichimowska A, Kern S, Riedel J, Panne U, King R, Maiwald M. “Click” analytics for “click” chemistry—a simple method for calibration–free evaluation of online NMR spectra. J Magn Reson. 2017;277:154.CrossRefPubMedGoogle Scholar
  26. 26.
    Kriesten E, Alsmeyer F, BardoW A, Marquardt W. Fully automated indirect hard modeling of mixture spectra. Chemom Intell Lab Syst. 2008;91(2):181.CrossRefGoogle Scholar
  27. 27.
    Dondoni A, Giovannini PP, Massi A. Assembling heterocycle-tethered C-glycosyl and alpha-amino acid residues via 1,3-dipolar cycloaddition reactions. Org Lett. 2004;6(17):2929.CrossRefPubMedGoogle Scholar
  28. 28.
    Kessler W. Multivariate Datenanalyse. Weinheim: Wiley-VCH Verlag GmbH & Co KGaA; 2006.  https://doi.org/10.1002/9783527610037.ch4.CrossRefGoogle Scholar
  29. 29.
    Westad F, Marini F. Validation of chemometric models—a tutorial. Anal Chim Acta. 2015;893:14.CrossRefPubMedGoogle Scholar
  30. 30.
    Zientek N, Laurain C, Meyer K, Paul A, Engel D, Guthausen G, et al. Automated data evaluation and modelling of simultaneous 19F–1H medium-resolution NMR spectra for online reaction monitoring. Magn Reson Chem. 2015;  https://doi.org/10.1002/mrc.4216.

Copyright information

© Springer-Verlag GmbH Germany, part of Springer Nature 2018

Authors and Affiliations

  • Simon Kern
    • 1
    • 2
  • Klas Meyer
    • 1
  • Svetlana Guhl
    • 1
  • Patrick Gräßer
    • 1
  • Andrea Paul
    • 1
  • Rudibert King
    • 2
  • Michael Maiwald
    • 1
  1. 1.Division Process Analytical TechnologyBundesanstalt für Materialforschung und -prüfung (BAM)BerlinGermany
  2. 2.Department Measurement and Control, Institute of Process EngineeringBerlin University of TechnologyBerlinGermany

Personalised recommendations