Anal Bioanal Chem (2014) 406:3157–3166 DOI 10.1007/s00216-014-7740-9

RESEARCH PAPER

Novel μ-membrane module for online determination of the free fatty acid content in the dispersed phase of oil-in-water emulsions Dennis Kaufhold & Janosch Fagaschewski & Daniel Sellin & Simon Strompen & Andreas Liese & Lutz Hilterhaus

Received: 2 October 2013 / Revised: 18 January 2014 / Accepted: 3 March 2014 / Published online: 25 March 2014 # Springer-Verlag Berlin Heidelberg 2014

Abstract Monitoring the dispersed phase of an oil-in-water (O–W) emulsion by means of Fourier transform infrared (FTIR) spectroscopy is a challenging task, restricted to the continuous phase that is in contact with the FTIR probe. Nonetheless, real-time measurement and kinetic analysis by FTIR, including analysis of the dispersed, often non-polar phase containing substrates and/or products, is desirable. Enzymatic hydrolysis of sunflower oil was performed in an O–W emulsion. After separation of the oil phase by use of a newly developed μ-membrane module, infrared spectra were collected using an attenuated total reflectance (ATR) cell. Different chemometric models were calibrated using the partial least squares (PLS) algorithm. Online application of a chemometric model based on the FTIR spectra enabled realtime monitoring of free fatty acid concentrations in the oil phase.

Keywords Biotechnological products . Chemometrics and statistics . Enzymes . IR and Raman spectroscopy . Process analysis . Spectroscopy instrumentation Electronic supplementary material The online version of this article (doi:10.1007/s00216-014-7740-9) contains supplementary material, which is available to authorized users. D. Kaufhold : J. Fagaschewski : D. Sellin : S. Strompen : A. Liese : L. Hilterhaus (*) Institute of Technical Biocatalysis, Hamburg University of Technology, Denickestraße 15, 21075 Hamburg, Germany e-mail: [email protected] D. Kaufhold BASF Personal Care and Nutrition GmbH, GCP/PO, 40589 Duesseldorf-Holthausen, Germany S. Strompen Universidad Nacional Autónoma de México, Instituto de Biotecnología, Cuernavaca, México

Introduction Many catalytic hydrolysis reactions are performed in oil-inwater (O–W) emulsions. Accurate analysis of such reaction systems in real time usually enables higher reproducibility, efficiency, and process control, and is therefore highly desirable. However, online analysis using Fourier transform infrared (FTIR) spectroscopy is currently restricted to the continuous phase that is in contact with the FTIR probe, and information on the dispersed phase cannot be obtained directly. Changing droplet size distributions and interface tensions are also a challenge for chemometric models and online monitoring of the continuous phase [1]. Particularly in biocatalysis, which covers the use of biological catalysts, including whole cells and natural or modified enzymes, for the synthesis or degradation of chemical compounds, reactions are frequently performed in biphasic mixtures of water-immiscible organic solvents or ionic liquids and water. This strategy is used when a low substrate or product solubility in the aqueous phase is a limitation [2, 3], and might also be used if an inhibiting effect of the substrate or product is present, reducing the reaction rate of the biocatalyst [4–6]. Moreover, some enzymes are only active at the oil–water interface. Within the two-phase reaction system, the biocatalytic reaction mostly takes place in the aqueous phase, and the mass-transport coefficient and the interfacial area determine the rate of substrate feed or product removal [7, 8]. As well as substrates dissolved in an organic phase, there is a large group of liquid substrates which can be used as a second, non-polar phase in biphasic reaction systems [9]. A common reaction of this type is the hydrolysis of different esters, especially oils, by lipases. The oil forms the non-polar phase, whereas the lipase is dissolved in the aqueous phase. During the reaction water is consumed to hydrolyze the oil, forming fatty acids and glycerol, with the intermediate products being mono and diglycerides. Whereas glycerol is dissolved in the aqueous phase, the

3158

fatty acids accumulate in the oil phase. The reaction is strongly dependent on the interfacial area of the oil and aqueous phase, because the lipase needs to adsorb at the liquid–liquid interface to be active within this reaction [10–12]. Additionally, both reactants—oil and water—must be brought into close contact. Therefore, the usual practice is to have the oil phase dispersed in the aqueous phase, for example in a stirred tank reactor [13–16]. The distribution of substrates, products, and enzymes depends on their physical properties and is not fixed, but changes over the course of hydrolysis [11]. Sampling and analysis of such emulsion-based biotransformations are laborintensive and based on wet chemical analysis and on chromatography, including high-performance liquid chromatography (HPLC) and gas chromatography (GC) [17]. Analysis of biocatalytic reactions, using Fourier transform infrared (FTIR) spectroscopy, is an attractive method of determining kinetic constants, intermediates, and side products, and the reaction progress, offline and online, in both onephase and multiphase systems [18]. By the 1990s, midinfrared FTIR and near-infrared FTIR spectroscopy were established as valuable tools for analyzing samples of oils and fatty acids regarding oxidation products, iodine value, saponification number, and free-fatty-acid content [19, 20]. Mid-infrared spectroscopy was used for the determination of the reaction progress during enzymatic hydrolysis of urea, using a CaF2 transmission cell for liquids and recording spectra after sampling [21]. Analysis of hydrolysis reactions catalyzed by an amidase was also performed using this method [22]. In these cases, D2O was used as a solvent, the specific peak heights of substrates or products were determined, and Lambert–Beer’s law was applied. In addition to the collection of FTIR spectra in transmission cells, infrared spectroscopy makes it possible to follow hydrolytic enzymatic reactions using attenuated total reflectance (ATR) probes [23, 24]. This method was used by Sills et al. [25, 26] to analyze the enzymatic hydrolysis of biomass. The spectra obtained from solid samples were analyzed using the partial least squares (PLS) algorithm to correlate the spectral data with the concentration data of the calibration. ATR-mid-infrared FTIR spectroscopy can also be used for liquid samples, as shown for, e.g., nitrile biotransformations [27], phosphatase-based processes [28], or fermentations [29], using univariate or multivariate analysis of the spectral data. Especially for realtime applications, mid-infrared spectroscopy is an attractive method of monitoring reactants within the reactor [30]. Pintar and coworkers used quantitative real-time FTIR spectroscopy on the acid-catalyzed hydrolysis of sucrose, using an ATR probe [31]. Monitoring of the enzymatic hydrolysis of starch from rye by use of near-infrared FTIR spectroscopy, using a reflectance probe immersed in the reactor, was reported by Tamburini [32]. A rapid ATR-mid-infrared spectroscopy method can be used for determination of the fatty-acid profile and peroxide

D. Kaufhold et al.

value of virgin olive oil, providing results comparable to standard procedures, and with the advantages of being less expensive and faster [33]. Thus, analyzing hydrolytic biotransformations by use of ATR-FTIR spectroscopy is an attractive alternative. However, analysis of reactions in multiphase systems—for example, enzymatic hydrolysis of oils— is restricted to the continuous phase of the reaction [34]. There is therefore a strong incentive to measure the dispersed phase (oil phase) in a stirred tank reactor to achieve the benefits of real-time measurement and kinetic analysis by FTIR, including analysis of the non-polar substrates and products. A practical approach, using membrane separation [35] combined with infrared spectroscopy in attenuated total reflection and multivariate calibration, is presented in this work. The lipasecatalyzed hydrolysis of sunflower oil was performed in an emulsion–membrane reactor [36] and the reaction components were emulsified in a stirred vessel. The emulsion was pumped through a hollow-fiber membrane [37–39], whereby the oil phase, consisting of sunflower oil and free fatty acids, was separated, enabling online analysis in a bypass (Fig. 1).

Experimental All chemicals used in this study were purchased from SigmaAldrich (Steinheim, Germany) and Carl Roth (Karlsruhe, Germany), unless indicated otherwise. Acid value, saponification value, degree of hydrolysis For calibration and validation of the chemometric model, appropriate offline analysis is required. As well as HPLC or GC, determination of the acid value, AV, offers a fast and reliable method for describing the free-fatty-acid content of an unspecified sample. The acid value gives the amount of potassium hydroxide, in milligrams, necessary to neutralize the free fatty acids of a 1 g sample, and can be calculated by use of the equation: AV ¼

c  M KOH *V msample

where AV is the acid value (mg g−1), c is the concentration of potassium hydroxide solution (mol L−1), MKOH is56.11 g mol−1, the molecular weight of potassium hydroxide, V is the volume of potassium hydroxide solution for neutralization (mL), and msample is the amount of sample used for determination of AV (g). The saponification value, SV, describes the free and bound fatty acids in a sample of oil or fat. It is defined as the amount of potassium hydroxide, in milligrams, necessary to saponify a 1 g sample. To determine SV, the sample is saponified with

Novel μ-membrane module for online determination

3159

Fig. 1 Experimental setup: A stirred tank reactor (STR) was combined with phase separation in a bypass using a μ-hollow-fiber membrane module (μ-hfm). Phase separation in the μmembrane module was achieved using a hydrophobic membrane. ATR-FTIR spectroscopic measurements in the permeate consisting of the oil phase were performed using the Bruker cell BioATR II (yellow: oil phase, blue: water phase)

an excess of potassium hydroxide. Subsequently, the remaining potassium hydroxide is determined by neutralization with hydrochloric acid. The degree of hydrolysis, H, can be calculated from the acid value and saponification value as follows: H¼

AV SV

Multivariate regression ATR-FTIR measurements provide high-resolution datasets. To extract relevant information or target variables, including concentrations or conversion, methods including, e.g., chemometric models based on the partial least squares (PLS) algorithm are used [40]. To evaluate the chemometric model, a root mean square error (RMSE) between the calibration value yi and prediction byi was calculated by use of the equation: vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2 u Xn  u yi t i¼1 yi −b RMSE ¼ n To validate the model with a test set (external validation), a root mean square error of prediction (RMSEP) was calculated. For internal cross validation (leave-one-out method), the same formula was used to calculate the root mean square error of cross validation (RMSECV) [41]. For the chemometric model, the QUANT2 package of the OPUS software (Bruker Optik GmbH, Germany) was used. μ-Hollow-fiber membrane module The μ-hollow-fiber membrane module consists of chemically inert polytetrafluoroethylene (PTFE). It has 10 individual

channels, one for each hollow fiber, with an internal channel diameter of 2 mm and a length of 75 mm (Fig. 2). One singlehollow-fiber membrane is inserted into the channel, and the fiber ends are potted into the connectors with acrylic casting resin. The membranes used for phase separation were microfiltration polypropylene (PP) hollow fibers, with an average membrane pore size dp =0.2 μm, from the module MD-020-FP-2N supplied by Microdyn-Nadir (Germany). The inner and outer fiber diameters were 0.7 and 1 mm, respectively. Experimental setup The reaction was performed in a 1000 mL stirred vessel. 250 mL sunflower oil (Oleon nv, Emmerich, Germany) were mixed with 250 mL ultrapure water. The phase-transition ratio for a sunflower oil–water mixture at 30 °C was found to be at an oil content of 60 %. The temperature was controlled and set to 30 °C. Over the whole reaction time, using a stirring rate of 400 rpm, a stable emulsion was formed. The reaction was started by the addition of enzyme (Lipase LipomodTM 34P, Biocatalysts Ltd., Cardiff, UK, 0.3 g L−1). The emulsion was continuously pumped from the reaction vessel through the μhollow-fiber membrane module at 10 mL min−1, using a peristaltic pump (Minipuls 3, Gilson, Biocatalysts Limited, USA). The dispersed organic phase was continuously separated within the membrane module, with a permeate flow of 300 μL min−1. The degree of hydrolysis was measured online in the permeate, using a flow cell installed in a Bruker Vertex 70 FTIR-spectrometer (Fig. 1). The tubes, connections, and fittings were purchased from Upchurch Scientific (Oak Harbor, USA). For ATR-FTIR measurements, a Bruker Vertex 70 FTIRspectrometer equipped with a liquid-nitrogen-cooled mercury cadmium telluride (MCT) detector was used. The spectrometer chamber was continuously purged with pure nitrogen to

3160

D. Kaufhold et al.

Fig. 2 (a) 3D-CAD-Model of one hollow-fiber channel and potting connection (b) μ-hollowfiber membrane module

remove water vapor and carbon dioxide. For spectra collection, the oil phase was pumped through the sample compartment of a BioATR II cell (Bruker, penetration depth: 6–8 μm for water). Spectra collection was performed using OPUS software. All spectra were collected in the range 600–4000 cm−1, at a resolution of 4 cm−1. For each spectrum 110 scans were accumulated. To identify chemometric methods of high accuracy, loweffort sample preparation, and robustness in terms of longtime measurements, three different chemometric models were developed and compared with regard to a small prediction error over the whole reaction course. For PLS calibration and data analysis, the QUANT2 OPUS software package was used. For every chemometric model, the spectra were subjected to the same preprocessing. Because the vibration band at 1462 cm−1 does not change over the course of the reaction, the region 1448–1560 cm−1 was set as an internal standard (Fig. 3). As a result, all spectra were normalized to the intensity of the band at 1462 cm−1. Additionally, the range of wavenumbers was limited to the region 1275–2002 cm−1. No further manipulation was applied to the spectra. For calibration of the first chemometric model, 21 mixtures of free fatty acid (FFA) from previous hydrolysis and of sunflower oil—indicating different degrees of hydrolysis— were prepared in duplicates. Each sample was mixed with water and pumped through the membrane. The permeate was directed to the BioATR II cell, and a spectrum was collected. Fig. 3 Calibration spectra of sunflower oil for different contents of free fatty acid (solid lines) and for different degrees of hydrolysis as determined by acidvalue titration (dash-dotted lines)

In the hydrolysis experiment a 1:1 (v/v) mixture of sunflower oil and water was prepared in a stirred vessel. The reaction was started by the addition of enzyme. Samples were taken from the reactor at several time points and the acid value was determined. The reaction mixture was separated by the membrane, and the dispersed phase pumped through the BioATR II cell. A spectrum was collected every 30 s. The acid-value samples taken from the reactor were used to validate the online prediction of the degree of hydrolysis of the first chemometric model, and for calibration of the second model. Five and two latent variables (PLS components) were chosen for the first and second chemometric model, respectively. Figure 3 shows the preprocessed spectra of sunflower oil for different concentrations of free fatty acids and for different degrees of hydrolysis, as determined by acid-value titration. The acid values determined from the hydrolysis experiment were fitted with a kinetic equation. Every spectrum from the hydrolysis experiment was used for the third chemometric model. The spectra were correlated with the corresponding fitted acid values for calibration. Nine PLS components were chosen for the third chemometric model.

Results and discussion The concentration measurement in the dispersed, non-polar phase of an oil–water emulsion was realized by membrane

Novel μ-membrane module for online determination

separation: the phases were separated by a novel μ-membrane module, and the biocatalyst was retained in the aqueous phase. The concentration of free fatty acids in the oil phase was determined by means of infrared spectroscopy in attenuated total reflection. Afterwards, three possibilities for model development and application were analyzed: a) based on calibration samples, b) based on reaction spectra, and c) based on kinetic simulation. First a chemometric model based on calibration samples was developed and tested. The samples were mixed with water, and pumped through the membrane and then through the BioATR II cell to collect a spectrum. The amount of FFA in the samples, in combination with the corresponding ATRFTIR spectra, was used for the calibration of the model. Use of five latent variables (PLS components) obtained the best results, with an RMSEP of 2.27 % and a bias of −0.761 %, and was therefore selected for this model. The relatively high penetration depth of the BioATR II cell (6–8 μm according to the manufacturer) and the presence of solvent-free compounds yields spectra of high signal-to-noise ratio. The water band, which can be observed as a small shoulder at ~1652 cm−1, is constantly low, irrespective of the mixing ratios of FFAs and sunflower oil and throughout the hydrolysis reaction, proving the consistent separation performance of the membrane. Therefore, the potential error that may arise from changing intensity over the reaction course, caused by changing interfacial tension accompanying changing droplet sizes, is eliminated [1, 11]. The spectra of calibration mixtures and the reaction spectra of the permeate strongly resemble each other. Contributions of mono and diglycerides, which normally display shoulders at ~1718 cm −1 (mono) and ~1730 cm −1 (di), are almost completely eliminated by the membrane separation. The peaks of the reaction spectra for ester and carboxylic acid vibration, at ~1740 and ~1710 cm−1, respectively, are just slightly broader for degrees of hydrolysis in the region of 40 %. Also, the reaction spectra display some vibration around 3235 cm−1 (−OH) and some spectral differences to the calibration mixtures in the region of 1160 cm−1 (not shown). This gives rise to the assumption that mono and diglycerides, which are formed as intermediates during the reaction, are mostly retained by the membrane. To verify this model, the spectra from the hydrolysis experiment were added as test spectra and the degree of hydrolysis was predicted. Figure 4a shows the predicted degree of hydrolysis H compared with H determined from acid values, and the predicted values from the FTIR calibration samples are presented in the electronic supplementary material (Table S1) This external validation of the model reveals good accuracy over the major course of the reaction. The prediction at t=0 starts with a degree of hydrolysis H=6.0 %, and adapts to the degree of hydrolysis as determined by acid-value titration after

3161

10 min or 20 spectra. The inaccurate prediction of the initial conversion values results from the deviations between calibration and reaction spectra at degrees of hydrolysis in the region of 40 %, and is discussed in detail below. After 4.17 h a final degree of hydrolysis of 100.0 % is predicted. The approach is comparatively high effort, as a result of calibration design, preparation of calibration samples, and measurement of the calibration spectra. The second approach was based on spectra, which were collected over the course of the reaction itself for use in model development, instead of using defined samples for calibration as used in the first model. This simplifies the process of model development. Indeed, the model is based on fewer data points, because only 13 acid values were determined in this experiment. Using two PLS components, an RMSECVof 1.6 % and a bias of 0.061 % were obtained. Figure 4b shows the predicted degree of hydrolysis for this model and the already known course of reaction as determined by titration. The prediction starts at 0.6 % degree of hydrolysis at t=0, indicating that the model is more precise in the initial phase than the previous model based on external calibration. The course of the reaction is well predicted. The highest predicted degree of hydrolysis is 100.2 %, which is acceptable considering the prediction error of this model. Because the model is based on only 13 spectra from the hydrolysis experiment, and the corresponding acid values, it is susceptible to disturbances and strongly depends on the quality of the acid-value titration. Pronk et al. and Schroën published a mathematical model to simulate lipase-catalyzed reactions under specific reaction conditions [36, 42]. The fatty acid production rate rFFA (molFFA h−1 mol−1), based on the molar amount of the organic phase in an emulsion reactor, can be described by:   −ΔEH 1 1  n rFFA ¼ a1 ⋅a⋅E⋅ 1−a2 ⋅X Gly ⋅e R ⋅ð T − 298 Þ ⋅ X eq FFA −X FFA where a is enzyme activity (U mg−1), E is enzyme concentration (mg mL−1), XGly is glycerol mole fraction (molFFA mol−1), XFFA is free fatty acid mole fraction (molFFA mol−1), Xeq FFA is equilibrium mole fraction (molFFA mol−1), ΔEH is activation energy of the hydrolysis reaction (12.3 kJ mol−1), R is the gas constant (kJ mol−1 K−1), T is reaction temperature (K), a1 is a fitting variable (L U−1 h−1), a2 is second fitting variable (–), and n is reaction order (–). To integrate this kinetic equation, the fourth and fifth order Runge–Kutta algorithm is used. By using the enzyme activity of the lipase LipomodTM 34P, of 52 U mg−1, the kinetic model is adjusted to the experimental data. Thus, the fit parameters and reaction are obtained for the hydrolysis of sunflower oil: a1 =0.26 L U−1 h−1, a2 =2.55, and n=1. An additional enzyme deactivation term is neglected in the kinetic model because of the short batch time. Using this simulation, implementation of all collected reaction spectra is possible, because offline data can be

3162

D. Kaufhold et al.

Fig. 4 Degree of hydrolysis H vs. time for lipase-catalyzed hydrolysis: (a) calculated from acid values (empty circles) and predicted with a chemometric model based on calibration samples (filled circles); (b) calculated from acid values (empty circles) and predicted with a chemometric model based on spectra collected over the course of the reaction

(filled circles); (c) calculated from acid values (empty circles) and simulated with a model for lipase-catalyzed hydrolysis [42] (filled circles); (d) calculated from acid values (empty circles) and predicted with a chemometric model based on simulation of reaction progress (filled circles)

calculated for all time points and correlated to the corresponding spectra. Figure 4c compares the simulation of our validation experiment and the degree of hydrolysis determined by acid-value titration. The fit of the experimentally determined acid values by the simulation starts at H=0, and approaches 100 % without exceeding this value. The third chemometric model is based on 500 spectra collected over the course of the reaction, and the corresponding acid values from the simulation. Figure 4d illustrates the comparison with the degree of hydrolysis determined by acid

values. The prediction starts at 1.2 % and does not exceed 100 %, with a maximum degree of hydrolysis of 99.2 %. The middle part, in the range 0.5–3.5 h, is also predicted with good accuracy. RMSECV is 0.21 % with nine principal components, and the bias is 0.002 %. Also, the development of the chemometric model is straightforward. After validation of the simulation, collection of offline data and preparation of calibration samples is not necessary. The different models are summarized in Table 1 and compared with regards to the overall error of prediction and the

Table 1 Summary of characteristic variables of different models for predicting the composition of the oil phase Model based on:

No. of calibration spectra

Rank

RMSECV / RMSEP (%)

Predicted H at t=0 (%)

Maximum degree of hydrolysis (%)

Comment

Calibration Reaction spectra Simulation

21 duplicates 13 500

5 2 9

2.27 (RMSEP) 1.61 (RMSECV) 0.21 (RMSECV)

6.0 0.6 1.2

100.0 100.2 99.2

Best at maximum Best at minimum Smallest overall error

Novel μ-membrane module for online determination

3163

performance of the model at very low and very high values of H. Of the three models tested, the model based on the reaction spectra provided the best overall prediction, with small deviations from initial and final degree of hydrolysis, and a relatively low RMSECV. The residual plots for the chemometric models based on calibration samples, based on the spectra collected over the course of the reaction, and based on the simulation of the reaction progress are presented in Fig. 5. The highest differences between true and predicted values are from the first chemometric model (Fig. 5a), which is based on the calibration mixtures. Interestingly, the second chemometric model (Fig. 5b), which is based on the reaction spectra with only two PLS components, displays a similar pattern of deviations, but the differences are smaller. The differences in the third chemometric model (based on the simulation curve) are the smallest. This was expected, because the amount of data used for the calibration was the highest. For all three models, the deviations are highest for small degrees of hydrolysis, where the prediction is too low. The corresponding true vs. prediction plots are presented in the electronic supplementary material (Figs. S1–S3). In all three residual plots in Fig. 5, a sinusoidal pattern in the model residuals indicates the presence of systematic variation in the data which is not accounted for by the model. The cause of this variation is unknown and needs further investigation. However, the residuals are small, so that there is little effect on the overall performance of the analysis. As well as process control, FTIR measurements are often used to determine enzyme properties, e.g. activity, deactivation constant, and half-life [43]. Initial reaction rates contain important information that can be used to derive these properties. Figure 6 shows the degree of hydrolysis during the first

Fig. 5 Residuals vs. degree of hydrolysis H: (a) chemometric modeling based on calibration samples; (b) chemometric modeling based on spectra collected over the course of the reaction; (c) chemometric modeling based on simulation of reaction progress Fig. 6 Comparison of different chemometric models (lines) and degree of hydrolysis calculated from acid values (circles)

3164

hour of the reaction time as predicted by the three developed chemometric models, compared with the offline results from acid-value determination. At reaction times above 30 min the predicted degree of hydrolysis is nearly the same for all models. As mentioned above, the prediction based on calibration samples starts at 6.0 % hydrolysis at t=0. This is caused by spectral deviations between the calibration and reaction spectra at medium degrees of hydrolysis, i.e. in the range 20–70 %. The largest deviations are at H=38 %, because the percentage of intermediates (mono and diglycerides) in the mixture is here at its maximum. The spectral contributions of these intermediates cannot be totally eliminated using the membrane module. Therefore, the accuracy of prediction at low degrees of hydrolysis was sacrificed for the sake of better prediction at medium degrees of hydrolysis. This can be achieved by increasing the number of PLS components. When choosing only three principal components, instead of five, the prediction of degrees of hydrolysis in the range 20–70 % is too low (14 % too low at H=38 %) but is better at t=0, with H=0.6 % (electronic supplementary material (Figs. S4 and S5)). With five principal components, deviations in H are reduced for medium conversions, but a deviation remains at t=0. The second model, based on 13 reaction spectra, is strongly dependent on the determined acid values. This model has a good correlation with the real experiment. As mentioned above, the membrane must be wetted with oil to ensure phase separation. Thus, a small additional amount of pure oil remains in the reactor periphery. Additionally, there is a small time gap between recorded FTIR spectra and correlated acid values, determined offline from samples taken from the reactor, which corresponds to the residence time in the tubes and membrane module on the way to the BioATR II cell. Only the second model is able to predict this time gap. The third model provides the best results with respect to an ideal course of reaction without any hold-up volume and time gap. This is not surprising, because these effects are not implemented in the model. They are not implemented in the previous model either, but in the third model the high number of calibration spectra and the use of nine PLS components results in their total elimination. The use of nine PLS components was possible because the simulated data does not contain any noise. However, noise is introduced into the calibration with the 500 calibration spectra. The noise is reduced with the number of samples in the pffiffiffi order of magnitude proportional to n . Therefore, the noise in this model is the lowest of all models presented in this work. This means even the higher PLS components still contribute to an improvement of the prediction, which otherwise would contribute to the residual matrix of the X data. In contrast is the second calibration model, which is

D. Kaufhold et al.

based on only 13 reaction spectra. Small data sets usually have more rigid, irregular, and noisy structures. Therefore, the data set of 13 reaction spectra can barely support more than two PLS components [44]. The plot of the RMSECV vs. the number of PLS components for the second and third model reveal a minimum of two and ten PLS components, respectively (Electronic Supplementary Material, Figs. S6 and S7). Using ATR-FTIR spectroscopy, combined with μ-hollowfiber membranes as a tool for process control, we suggest the second approach using real experimental spectra. This model best implements the true experimental conditions. For the determination of activity or half-life times the third approach delivers the best results, because it is close to the ideal and theoretical course of the reaction.

Conclusion Online determination of the degree of hydrolysis of the enzymatic hydrolysis of sunflower oil in a biphasic system with oil content below 60 % was demonstrated. The concentration measurement in the non-polar phase of an oil–water emulsion was realized by a simple and effective method. First, the phases were separated by a novel μ-membrane module. The biocatalyst was retained in the aqueous phase. Second, the concentration of free fatty acids in the oil phase was determined by means of infrared spectroscopy in attenuated total reflection. The phase-transition ratio for a sunflower-oil–water mixture at 30 °C was found to be at an oil content of 60 % [36]. In this work, emulsions with 50 % oil content were monitored. When determining concentrations in emulsions, ATR probes are limited to the continuous phase. This limitation was overcome by phase separation using a novel μ-membrane module. Both phases were recycled to the stirring vessel after concentration determination. The constantly low water content in the permeate leads to infrared spectra of high and consistent quality throughout the process. Another major advantage of this method is that the reaction spectra of the permeate strongly resemble those of the mixtures of FFA and sunflower oil (compare Fig. 3) in the monitored range of wavenumbers (1275–2002 cm−1). This means that any contributions of di and monoglycerides, which are formed during the reaction, are almost completely separated with the aqueous phase by the membrane. This circumstance makes the μ-membrane particularly well suited to enabling the monitoring of hydrolysis reactions of plant oils. Acknowledgements The authors thank Johanna Rodriguez for technical assistance and Oleon nv, Emmerich, Germany for kindly providing the sunflower oil.

Novel μ-membrane module for online determination

References 1. Reis MM, Araújo PH, Sayer C, Giudici R (2007) Spectroscopic online monitoring of reactions in dispersed medium: Chemometric challenges. Anal Chim Acta 595(1–2):257–265 2. Liese A, Zelinski T, Kula M, Kierkels H, Karutz M, Kragl U, Wandrey C (1998) A novel reactor concept for the enzymatic reduction of poorly soluble ketones. J Mol Catal B Enzym 4(1–2):91– 99 3. Kohlmann C, Leuchs S, Greiner L, Leitner W (2011) Continuous biocatalytic synthesis of (R)-2-octanol with integrated product separation. Green Chem 13(6):1430 4. Lyagin E, Drews A, Bhattacharya S, Ansorge-Schumacher MB, Kraume M (2010) Continuous screening system for inhibited enzyme catalysis: A membrane reactor approach. Biotechnol J 5(8): 813–821 5. Götz K, Liese A, Ansorge-Schumacher M, Hilterhaus L (2013) A chemo-enzymatic route to synthesize (S)-γ-valerolactone from levulinic acid. Appl Microbiol Biotechnol 97:3865–3873 6. Whittall J, Sutton P (2009) Practical Methods for Biocatalysis and Biotransformations 7. Brenna E, Gatti FG, Monti D, Parmeggiani F, Sacchetti A (2011) Productivity enhancement of C = C bioreductions by coupling the in situ substrate feeding product removal technology with isolated enzymes. Chem Commun 48(1):79 8. Fagaschewski J, Bohne S, Kaufhold D, Müller J, Hilterhaus L (2012) Modular Micro Reaction Engineering for Carboligation Catalyzed by Benzoylformate Decarboxylase. Green Processing and Synthesis(1): 337–344 9. Lilly M (1982) Two-liquid-phase biocatalytic reactions. J Chem Technol Biotechnol 32(1):162–169 10. Saktaweewong S, Phinyocheep P, Ulmer C, Marie E, Durand A, Inprakhon P (2011) Lipase activity in biphasic media: Why interfacial area is a significant parameter? J Mol Catal B Enzym 70(1–2):8– 16 11. Reis P, Holmberg K, Watzke H, Leser M, Miller R (2009) Lipases at interfaces: A review. Adv Colloid Interf Sci 147– 148:237–250 12. Straathof AJ (2003) Enzymatic catalysis via liquid-liquid interfaces. Biotechnol Bioeng 83(4):371–375 13. Al-Zuhair S, Ramachandran K, Hasan M (2008) Effect of enzyme molecules covering of oil–water interfacial area on the kinetic of oil hydrolysis. Chem Eng J 139(3):540–548 14. Woodley JM, Cunnah PJ, Lilly MD (1991) Stirred Tank Two-Liquid Phase Biocatalytic Reactor Studies: Kinetics, Evaluation and Modelling of Substrate Mass Transfer. Biocatal Biotransformation 5(1):1–12 15. van Sonsbeek H, Beeftink H, Tramper J (1993) Two-liquid-phase bioreactors. Enzym Microb Technol 15(9):722–729 16. Puthli MS, Rathod VK, Pandit AB (2006) Enzymatic hydrolysis of castor oil: Process intensification studies. Biochem Eng J 31(1):31– 41 17. Hamilton R (1998) Lipid Analysis in Oils and Fats 18. Müller JJ, Baum S, Hilterhaus L, Eckstein M, Thum O, Liese A (2011) Simultaneous Determination of Mono-, Di-, and Triglycerides in Multiphase Systems by Online Fourier Transform Infrared Spectroscopy. Anal Chem 83(24):9321–9327 19. Man YBC, Moh MH (1998) Determination of free fatty acids in palm oil by near-infrared reflectance spectroscopy. J Amer Oil Chem Soc 75(5):557–562 20. van de Voort F, Ismail A, Sedman J, Emo G (1994) Monitoring the Oxidation of Edible Oils by Fourier Transform Infrared Spectroscopy. J Am Oil Chem Soc 71(3):243–253 21. Karmali K, Karmali A, Teixeira A, Marcelo Curto M (2004) The use of Fourier transform infrared spectroscopy to assay for urease from

3165 Pseudomonas aeruginosa and Canavalia ensiformis. Anal Biochem 331(1):115–121 22. Pacheco R, Karmali A, Serralheiro M, Haris P (2005) Application of Fourier transform infrared spectroscopy for monitoring hydrolysis and synthesis reactions catalyzed by a recombinant amidase. Anal Biochem 346(1):49–58 23. Minnich CB, Buskens P, Steffens HC, Baeuerlein PS, Butvina LN, Kuepper L, Leitner W, Liauw MA, Greiner L (2007) Highly flexible fibre-optic ATR-IR probe for inline reaction monitoring. Org Process Res Dev 11(1):94–97 24. Gebhard J, Sellin D, Hilterhaus L, Liese A (2013) Online Analysis of Enzymatic Polycondensation Reactions in Bubble Column Reactors by Means of ATR-FTIR Spectroscopy. Chemie Ingenieur Technik 85(7):1016–1022 25. Sills DL, Gossett JM (2012) Using FTIR spectroscopy to model alkaline pretreatment and enzymatic saccharification of six lignocellulosic biomasses. Biotechnol Bioeng 109(4): 894–903 26. Sills DL, Gossett JM (2012) Using FTIR to predict saccharification from enzymatic hydrolysis of alkali-pretreated biomasses. Biotechnol Bioeng 109(2):353–362 27. Dadd MR, Sharp DC, Pettman AJ, Knowles CJ (2000) Real-time monitoring of nitrile biotransformations by mid-infrared spectroscopy. J Microbiol Methods 41(1):69–75 28. Gäb J, Melzer M, Kehe K, Richardt A, Blum M (2009) Quantification of hydrolysis of toxic organophosphates and organophosphonates by diisopropyl fluorophosphatase from Loligo vulgaris by in situ Fourier transform infrared spectroscopy. Anal Biochem 385(2):187–193 29. Roychoudhury P, McNeil B, Harvey LM (2007) Simultaneous determination of glycerol and clavulanic acid in an antibiotic bioprocess using attenuated total reflectance mid infrared spectroscopy. Anal Chim Acta 585(2):246–252 30. Natalello A, Sasso F, Secundo F (2013) Enzymatic transesterification monitored by an easy-to-use Fourier transform infrared spectroscopy method. Biotechnology Journal 8(1, SI) 31. Pintar A, Batista J, Levec J (2002) In situ Fourier transform infrared spectroscopy as an efficient tool for determination of reaction kinetics. Analyst 127(11):1535–1540 32. Tamburini E, Castaldelli G, Bernardi T (2011) Fourier transform-near infrared spectroscopy in-line monitoring of the enzymatic hydrolysis of starch in rye:water mashes for first-generation bioethanol production. J Near Infrared Spectrosc 19(3):181 33. Maggio RM, Kaufman TS, Carlo MD, Cerretani L, Bendini A, Cichelli A, Compagnone D (2009) Monitoring of fatty acid composition in virgin olive oil by Fourier transformed infrared spectroscopy coupled with partial least squares. Food Chem 114(4):1549–1554 34. Müller JJ, Neumann M, Scholl P, Hilterhaus L, Eckstein M, Thum O, Liese A (2010) Online Monitoring of Biotransformations in High Viscous Multiphase Systems by Means of FT-IR and Chemometrics. Anal Chem 82(14):6008–6014 35. Dupuy A, Athes V, Schenk J, Jenelten U, Souchon I (2011) Solvent extraction of highly valuable oxygenated terpenes from lemon essential oil using a polypropylene membrane contactor: potential and limitations. Flavour Fragr J 26(3): 192–203 36. Schroën K (1995) Physical chemistry and process engineering of an emulsion/membrane bioreactor. Dissertation, University of Wageningen 37. Kopf F, Schlüter M, Kaufhold D, Hilterhaus L, Liese A, Wolff C, Beutel S, Scheper T (2011) Laminar Mixing in Miniature HollowFibre Membrane Reactors by using Secondary Flows (Part 1). Chemie Ingenieur Technik 83(7, SI):1066–1073 38. Kaufhold D, Kopf F, Wolff C, Beutel S, Hilterhaus L, Hoffmann M, Scheper T, Schlüter M, Liese A (2012) Generation of Dean vortices

3166 and enhancement of oxygen transfer rates in membrane contactors for different hollow fiber geometries. Journal of Membrane Science 423424:342–347 39. Kaufhold D, Kopf F, Wolff C, Beutel S, Hilterhaus L, Hoffmann M, Scheper T, Schlüter M, Liese A (2013) Chemical Absorption of CO2 in Helically Wound Hollow Fiber Membrane Contactors. Chemie Ingenieur Technik 85(4, SI):476–483 40. Wold H (1974) Causal flows with latent variables. Eur Econ Rev 5(1):67–86

D. Kaufhold et al. 41. Kessler W (2007) Multivariate Datenanalyse. Für die Pharma-, Biound Prozessanalytik. WILEY-VCH, Weinheim 42. Pronk W, Boswinkel G, van't Riet K (1992) Parameters influencing hydrolysis kinetics of lipase in a hydrophilic membrane bioreactor. Enzym Microb Technol 14(3):214–220 43. Kumar S, Barth A (2010) Following Enzyme Activity with Infrared Spectroscopy. Sensors 10(4):2626–2637 44. Esbensen KH, Guyot D, Westad F, Houmoller LP (2002) Multivariate Data Analysis – in practice. CAMO ASA, Trondheim