Application of green analytical chemistry to a green chemistry process: Magnetic resonance and Raman spectroscopic process monitoring of continuous ethanolic fermentation

Compact 1H NMR and Raman spectrometers were used for real‐time process monitoring of alcoholic fermentation in a continuous flow reactor. Yeast cells catalyzing the sucrose conversion were immobilized in alginate beads floating in the reactor. The spectrometers proved to be robust and could be easily attached to the reaction apparatus. As environmentally friendly analysis methods, 1H NMR and Raman spectroscopy were selected to match the resource‐ and energy‐saving process. Analyses took only a few seconds to minutes compared to chromatographic procedures and were, therefore, suitable for real‐time control realized as a feedback loop. Both compact spectrometers were successfully implemented online. Raman spectroscopy allowed for faster spectral acquisition and higher quantitative precision, NMR yielded more resolved signals thus higher specificity. By using the software Matlab for automated data loading and processing, relevant parameters such as the ethanol, glycerol, and sugar content could be easily obtained. The subsequent multivariate data analysis using partial linear least‐squares regression type 2 enabled the quantitative monitoring of all reactants within a single model in real time.

Further to PLS, PLS2 allows the simultaneous quantitative prediction of several y variables, such as concentrations of different chemical components in a mixture. It is advantageous to apply PLS2 when the target variables are interdependent or correlated with each other, as is often the case with chemical reactions, and when the correlated y variables possess different error ranges. The dimensions of the analyzers should follow an ecological and economical point of view (Pinto, Pereira, Ribeiro, & Farinas, 2016;Pinto, Ribeiro, & Farinas, 2018).
These type of spectrometers often provide a sensitivity suitable for observing neat liquids or solids dissolved in neat liquids as reactants.
In this study, compact 1 H NMR and handheld Raman spectrometers in combination with automated multivariate methods are used for the real-time monitoring of a continuous biotechnological process, that is, ethanol fermentation using immobilized yeast. The data will be analyzed using chemical kinetic models and PLS2 to obtain c-t diagrams and to gain process knowledge from chemical kinetic models. Reaction monitoring is shown to obey the 12 principles of Green Analytical Chemistry.

| Process and analytical set-up
A BIOSTAT ® B fermenter I with a capacity of 2 L (Sartorius AG, Göttingen, Germany) was used for fermentation. The mounted agitator (B. Braun Biotech International GmbH,Melsungen,Germany;cf. Figure 1) was operated at a constant stirring velocity of 120 rpm. A peristaltic pump (B. Braun Biotech International GmbH) was used to flow the sucrose-containing solution from the storage vessel to the temperaturecontrolled reactor at a flow rate of 2.5 ml per minute during the entire process. From the reactor, the solution further passed a QS 0.5 mm quartz-flow cell (Hellma Analytics, Muellheim, Germany) to allow for online Raman spectral recording. The implemented MultiPurposeSampler (Gerstel, Muelheim an der Ruhr, Germany) was equipped with a stainlesssteel flow cell, VT98 tray, fast wash station, and injection valve. Through this arrangement, samples were channeled through the flow cell and transferred to the NMR spectrometer and/or pipetted into highperformance liquid chromatography (HPLC) vials (Legner, Friesen, Voigt, Horst, & Jaeger, 2016). For HPLC analysis, samples of 2 ml each were collected. Spectroscopic analyses were performed as described below.
Finally, the solution was conveyed towards the product vessel. For multivariate data analysis, two-thirds of the sample spectra were used as calibration set and one-third as validation set. The multivariate model was applied to all remaining recorded spectra, that is, spectra without corresponding HPLC analysis.
Fermentation was fed with a sugar solution. Sucrose (>99.5%, Merck KGaA), was used as obtained. A buffer solution containing 100 ml sodium hydroxide (1 mol·L −1 ; Merck KGaA) and 200 ml acetic acid (1 mol·L −1 , Merck KGaA) in 5 L water was prepared. To this solution, 50 g of calcium chloride dihydrate and sucrose were added in the appropriate amounts.
The amount of sucrose depended on the volume of the fermenter and the amount of immobilized yeast. In case of the 3 L fermenter, 17% sucrose (850 g sucrose) were added to the buffer solution.
A one-stage process was conducted by continuously conveying a 17% sucrose-containing educt solution through a 3 L bioreactor at a F I G U R E 1 Flow schematics of the fermentation (left). From the storage vessel, the sucrose-containing educt solution is fed into the temperature-controlled reactor equipped with a stirrer. For real-time spectroscopic analysis, the ethanol-containing product mixture can be analyzed by 1 H NMR and Raman spectroscopy. Both spectrometers are implemented in the process via a flow cell. Flow reactor with immobilized yeast cells (right). Sensors for pH value and temperature can be mounted to the reactor via screw caps [Color figure can be viewed at wileyonlinelibrary.com] temperature of 30°C. When steady state was reached, the process was terminated after 900 min. The process was transformed into a two-stage process using new immobilized yeast cells: a 15% sucrose solution was pumped until reaching steady state. Subsequently, a 20% sucrosecontaining solution was conveyed until a second steady state. The temperature was again kept at 30°C but a 2 L bioreactor was used (cf. Figure 1). The process was monitored over a period of 3000 min. For the separate model development of the one-stage and two-stage fermentation, samples for calibration and validation were taken in two feasibility studies carried out previously. The separate model development requires a prior execution of the two processes to be investigated to obtain the calibration by means of the spectrometers used as well as reference analysis. For single-stage fermentation, 50 samples were taken and divided into a training set and a test set (cf. Table 1). For the two-stage process, the model was developed with 40 samples (cf. Tables S1-2).

| Immobilization of the biocatalyst
Sodium alginate, 10 g (Technical grade, VWR International GmbH, Darmstadt, Germany) was dissolved in 350 ml of distilled water at ambient temperature. After cooling the alginate to 30°C, 42 g of Baker's yeast (VWR International GmbH) were added under stirring.
The homogeneous yeast-alginate mixture was added dropwise to a 10 g·L −1 calcium chloride solution (calcium chloride dihydrate, VWR International GmbH), leading to the formation of a crosslinked solution of very uniform beads.

| HPLC reference method
Concentration data for referencing were obtained using the HPLC system Smartline (Knauer GmbH, Berlin, Germany) equipped with a refractive index detector. A Eurokat H column (Knauer GmbH) of dimensions 300 × 8 mm, particle size 10 μm, was used for chromatographic separation. An isocratic flow was set to 0.8 ml·min −1 using 0.01 mol·L −1 sulfuric acid (99.999%; Merck KGaA, Darmstadt, Germany) as mobile phase. The separation of all substances involved in the process was achieved within 20 min at 60°C. The chromatograms were processed and evaluated using the software EuroChrom ® 1.57 (Knauer GmbH).

| Compact spectrometry
NMR spectra were recorded using a picoSpin80 spectrometer (Thermo Fisher Scientific, Dreieich, Germany) with a proton Larmor frequency of 82 MHz. The spectrometer was controlled by Thermo Fisher PicoSpin software 0.9.3 running on the spectrometer control board and accessed via a web interface from a laptop computer. The instrument had a flow cell with an active volume of 40 nl. The electronic chemical shift referencing corrected for magnetic field drift and allowed recording spectra of the neat liquids without the addition of deuterated solvents. A simple pulse-acquire scheme was used as an experiment. The number of accumulations was set to 16.
The pulse length was set to 60 µs corresponding to a 90°pulse. A recovery delay of 500 µs followed by a repetition or relaxation delay of 6 s was applied, which proved a sufficient compromise with respect to relaxation rates, experimental conditions, sensitivity, and precision. Spectra were recorded with 4092 acquisition points and a zero filling of 9000. The spectral width amounted to 4 kHz.
Raman spectra were recorded using the handheld Raman (h-Ram) spectrometer IDRaman mini 2.0 (Ocean Optics, Dunedin, FL). Spectra were acquired by means of a point-and-shoot adapter in combination with a QS 0.5 mm quartz-flow cell (Hellma Analytics, Muellheim, Germany). The spectrum was acquired from 400 to 2300 cm −1 and the spectral resolution was 13 cm −1 . The laser had an excitation wavelength of 785 nm ± 0.5 nm and a power of 100 mW. The software Peak 1.3.54 (Snowy Range Instruments, Laramie, WY) was used for spectral acquisition. The Multi-Mode allows any number of analyses to be carried out with a freely selectable interval. The respective raw data of the two spectroscopic methods were automatically saved into a cloud, so that data pre-processing and univariate and multivariate data analysis took place immediately afterwards.

| Data preprocessing
Spectral processing was carried out using the software MatLab R 2016b (MathWorks, Inc., Natick, MA). A home-built Matlab script waited for new raw data, loaded them, processed the data and eventually moved the file to a designated folder. The overall analysis time required for building c-t diagrams was taken from the.jdx files' time stamp. Spectral processing, Fourier transformation, and phase correction of the NMR raw data were performed using Matlab routines. After subsequent baseline correction using a Bernstein polynom fit 3rd order, a moving average over an interval of six data points for the reduction of file size and normalization to the largest peak were applied. Since the relevant spectral range used for multivariate data analysis laid between 0.3 and 5.5 ppm, the remaining data for the further analyses were rejected. The spectra were referenced to the water resonance at 4.7 ppm.
Raman raw data were extracted from text files very quickly. A subsequent baseline correction did not give better results due to T A B L E 1 Number of spectra in data subsets for calibration and validation for the one-stage fermentation and the corresponding concentration ranges and means of the relevant process media and substances  preset parameters in the control software, so this additional processing step was omitted. No standardization was performed, since the highest signal was specific for primary alcohols. The selected spectral range from 400 to 2300 cm −1 was used for multivariate analyses.

| Automated data analysis
To this purpose, the spectrometers were implemented online to the process with spectral acquisition and analysis set to a defined repetition rate. The obtained raw data were uploaded to cloud storage so that evaluations could be made independent of the laboratory. After the automated import into Matlab the data were processed, followed by the multivariate data analysis, see Figure 2.
Univariate data analysis was used to generate c-t diagrams by integrating characteristic signals. For the multivariate methods, a model was created and validated. As a result, a quantitative real-time model was obtained. An easy-to-use graphic interface allowed to obtain the relevant process data in numerical or graphical representation.

| Multivariate data analysis
For the multivariate prediction of the concentrations of the process media, a partial least-squares regression (PLS2) for more than one y variable was performed. PLS2 was performed using home-built scripts and PLS_Toolbox version 8.2.1 for MatLab (Eigenvector Research, Inc., Wenatchee, WA). As calibration or training set, 40 1 H NMR and Raman spectra out of 50, for which reference data were recorded, were used. Crossvalidation was performed applying the leave-one-out procedure. The remaining 10 spectra were used for prediction to identify the best model. The optimal number of latent variables was determined by minimizing the sum of the residual square errors of the prediction (PRESS) for the validation set (Kessler, 2006)  Thus, a multivariate model became essential such that all process media could be interpreted and quantified.
All NMR spectra displayed the solvent water resonance at 4.7 ppm (cf. Figure 3, left). The use of a spectrometer with multiple or gradient pulse capabilities thus allowing for water suppression sequences would improve sensitivity and ultimately the quality of the data analysis (Gouilleux, Charrier, Akoka, & Giraudeau, 2017). The resonance of methyl groups at 1.2 ppm stemmed from the ethanol and could be used for univariate data analysis, as this range was not superimposed by fructose, glucose and glycerol resonances. At 2.0 ppm, the methyl resonance of the acetic acid from the buffer system was clearly recognized. In the Raman spectra, the strong signal at 880 cm −1 , which originates predominantly from the C─OH vibrations from ethanol slightly superimposed by the corresponding sucrose, fructose, and glycerol vibrations (Figure 3, right and Figure S2) was visible. Further signals originating from C─C─O stretching vibration of ethanol were observed at 1065 cm −1 (Socrates, 2004). The O─H deformation vibration of primary alcohols appeared at 1460 cm −1 (Socrates, 2004).
Although representative signals of all components are a prerequisite for reaction monitoring, signals of all process media do not need to be fully resolved, specific or even assigned. Whereas true for process monitoring data after calibration and model generation, reference data require resolved, specific, hence quantifiable signals, as was the case for the HPLC reference data used in this study. Since the sugar educts did not provide intense signals, univariate analysis of the sugar transformation was not possible, requiring hence the use of multivariate methods.

| Multivariate data analysis
For the direct determination of the ethanol, fructose, glucose, and glycerol concentration, a multivariate PLS2 model was devised. Forty spectra were used for the creation of the calibration set and another ten spectra for the external validation of the model, see Table 1.
For a representative selection, every fifth spectrum was used as validation, so that no distortions occurred due to the two steady states at the beginning and end of ethanol fermentation. Figure 4 shows the results of concentrations predicted by means of PLS2 from the external validation samples and plotted versus values obtained from the HPLC data as the reference method.
Very good predictions for ethanol and fructose concentrations were achieved using external validation. This is due to the fact that fructose was degraded enzymatically more slowly than glucose.
Hence, higher concentrations were observed which can be detected more reliably (Cason, Reid, & Gatner, 1987). The same applied to the ethanol concentration, which was easily detectable due to its very pronounced bands, see Figure 3. Glycerol as a by-product of the incomplete anaerobic process was detectable in small amounts and thus led to variations in the prediction. The RMSEs are listed for both spectroscopic methods. Values were calculated using five latent variables for internal and external validation, see Table 2.
The prediction errors are well comparable for both spectroscopic methods. Only the RMSEP for the ethanol concentration was significantly higher in the NMR data based model with 0.029 as compared to the Raman spectra based model. However, the results were fully adequate for a direct real-time concentration estimation, so that a one-stage ethanolic fermentation could be monitored.

| One-stage fermentation process
The validated model was used to determine the concentration of the process constituents of unknown fermentation samples. In Figure 5, the concentration data of the calibration set (crosses) were plotted against time together with that predicted concentrations from the unknown samples (circles). The parameters of the mathematical model listed in Table 3 agree very well between the two spectroscopic methods. The data can be interpreted in terms of the mean maximum concentrations M 1 , the rate constants k 1 and the inflection points t 1 of the sigmoidal curves at 17% sucrose for Raman and NMR spectroscopic data.
A detailed explanation of the mathematical description and model is given in equation (3) | 2879 educt concentration. Hence, after successful monitoring of the one-stage process, a two-stage fermentation was monitored.

| Two-stage fermentation process
The resulting Raman and NMR spectra were processed and analyzed using the established PLS2 model. The detailed list of training and test samples as well as the obtained RMSEs for both spectroscopic methods can be found in Tables S1-2. Figure S1 shows the application of the predicted concentrations from the validation samples against the concentrations determined by HPLC reference analysis. The predicted concentrations were plotted in a c-t diagram (cf. Figure 6), glucose, fructose, and glycerol concentrations were predicted.
The predicted values are consistent among both spectroscopic techniques in the first stage of the fermentation, that is, 15% sucrose  Table S3.
The data derived from the multivariate method were compared against equation (3). As can be seen from Figure 6, the multivariately predicted data were well described. The corresponding parameters of the PLS2 model after computational fitting are listed in Table 4.
where M represents the mean maximum intensity observed by the corresponding spectroscopic method, t the reaction time, t H the inflection point of the sigmoidal curve, and k the rate constant. As can be recognized, equation (3) is able to describe a one-stage fermentation process using the first summand only and the twostage fermentations using both summands. In kinetic terms, M reflects the maximum concentration, and t H can be understood as modulating the reaction time and reflecting an induction period. The fit parameters of the multivariate predicted ethanol concentration are listed in Table 4. They agree well again between both spectroscopic techniques with the exception of the inflection points t 1,2 .
However, this could be attributed to the moderate scattering of the predicted concentrations from the NMR spectra at the second stage (cf. Figure 6).
The maximum ethanol yield M was comparable between both spectroscopic methods. The parameters M 1 and M 2 could be assigned T A B L E 3 Fitted parameters of predicted ethanol concentration using partial least-squares regression type 2 with handheld Raman and NMR spectroscopic data in the one-stage fermentation represented the difference between the maximum of the second stage and M 1 . Thus, a relationship of the maximum yield with the sucrose was apparent. However, there was a maximum yield with respect to educt concentration. This could be explained by substrate inhibition at too high levels of educt concentration and by the inactivation of the biocatalyst through ethanol. The rate constants k were found of the same order of magnitude for both methods.
However, values varied due to somewhat lower sensitivity and robustness of the NMR instrument due to its 40 nL flow cell, see Figure 6. Further process monitoring would need more extensive isolation and shielding of the NMR instruments. Nevertheless, the parameters k 1 and k 2 were of the same order of magnitude, which may be seen indicative for the type of yeast. In terms of process optimization, the process monitoring and analysis revealed that the biocatalyst could convert more sucrose, but at the same rates.
The application of bioreactions was aligned with the Green Chemistry principles in using immobilized biocatalysts and water as a solvent. The compact spectrometers matched the size of the small-scale reactor and could be easily coupled to the continuous bioreactor. Fast monitoring techniques enabled real time and automated concentration determination. Compared to conventional chromatographic methods, sampling was unnecessary and hence solvent waste was avoided. The automated nearline process monitoring allowed to prevent frequent physical contact with the reaction vessel, being a major advantage in terms of occupational safety. Automation, waste reduction, small-scale operation, real-time monitoring, and minimum sampling are all goals of GAC.

| CONCLUSION
Compact spectrometers proved ideal for monitoring continuous ethanolic fermentation in real time. Disadvantages of the low resolution compared to compact analyzers were well compensated by the use of multivariate data analysis. After automated data acquisition, fast laboratory independent processing of the raw data to generate the relevant reaction profiles was performed such that the current reaction progress in the bioreactor was visualized. As a result, a maximum yield of 1.4 mol·L −1 of ethanol was produced with the yeast cells having reached their optimum turnover frequency.
Higher temperatures, other immobilization support, different yeast strains, and buffer systems may increase the yields further. During condition optimization, preliminary work indicated that the present compact spectrometers were able to monitor changes in real time as well. The developed PLS2 model may further be optimized by increasing the data set for calibration. Yet, GAC principles suggest to use, the lowest possible number of samples and that spectrometry should replace chromatography with respect to consumed solvent, energy and resources. This could be achieved with the compact spectrometers as described. Conducting the fermentation as onestage and two-stage processes, obtained c-t fermentation profiles increased the process knowledge of all reactions components and rates, yields and induction periods. Based on the mathematical model developed therefrom, a control chart may be devised which might realize automated control via a feedback loop.

R. Legner is very grateful for a grant from the German Academic
Scholarship Foundation. The authors are grateful to HSNR for financial support and Thermo Fisher Scientific for providing the compact picoSpin80 1 H NMR spectrometer.