Fuel 319 (2022) 123794 Contents lists available at ScienceDirect Fuel journal homepage: www.elsevier.com/locate/fuel Full Length Article Near Infrared Spectroscopy: A useful technique for inline monitoring of the enzyme catalyzed biosynthesis of third-generation biodiesel from waste cooking oil Josu López-Fernández a, Desirèe Moya b, María Dolors Benaiges a, Francisco Valero a, *, Manel Alcalà b a Department of Chemical, Biological and Environmental Engineering, School of Engineering, Universitat Autònoma de Barcelona, Bellaterra, 08193 Barcelona, Spain b Department of Chemistry, Faculty of Sciences, Universitat Autònoma de Barcelona, Bellaterra, 08193 Barcelona, Spain A R T I C L E I N F O A B S T R A C T Keywords: Near Infrared (NIR) spectroscopy was used for inline monitoring the enzymatic production of third-generation Biodiesel biodiesel from waste cooking oil (WCO) —with fatty acid profile similar to olive oil— under the principles of Biocatalysis Continued Process Verification (CPV) and Process Analytical Technology (PAT). For this purpose, covalently Near infrared spectroscopy immobilized mature Rhizopus oryzae lipase (rROL) was used to transesterify WCO with methanol and ethanol, Lipase Komagataella phaffii firstly in 10 mL vials under orbital stirring and then it was successfully scaled up to a mechanically stirred 50 mL Inline monitoring laboratory reactor specially designed for use of a NIR probe. Biocatalyst half-life and productivity after ten re­ action cycles with methanol and ethanol were assessed. Slightly higher operational stability with methanol (337.5 h, 54 batches) vs with ethanol (146.7 h, 44 batches) was observed, but decreased productivity as a result of the increased reaction times used with the former (219 vs 327 µmol min− 1). The NIR spectroscopy results were highly correlated with those of gas chromatography (GC) used as reference. Thus, the root mean-square standard error of prediction (RMSEP) was 2.0% for methanol and 2.1% for ethanol. Therefore, NIR spectroscopy, which allows data acquisition in real time, is suitable for inline monitoring of enzymatic production of biodiesel. 1. Introduction non-edible oils such as radish and apricot kernel oils [4] while third- generation biodiesel is produced from substrates like waste cooking oil Circular economy is gaining momentum for reducing the environ­ (WCO) in a manner that complies with the principles of circular econ­ mental footprint of mankind and overcoming the widely adopted system omy [5,6]. grounded on linear resource use and waste generation. Using alternative At present, biodiesel is industrially produced mainly by chemical and renewable sources of energy such as biogas, 2,5-dimethylfuran, transesterification of first-generation substrates in the presence of a hydrogen or biodiesel for transportation has been proposed to alle­ homogeneous basic catalyst. Using second- or third-generation sub­ viate the current heavy dependence on fossil fuels [1–3]. The latter is a strates requires their pre-treatment to remove free fatty acids (FFAs), mixture of mono-alkyl esters of long-chain fatty acids that can be which they typically contain in high proportions [7,8]. In fact, unless directly used by existing combustion engines and is arousing increasing properly neutralized, FFAs favor saponification and give biodiesel of interest by virtue of its being renewable and biodegradable, accelerating decreased quality requiring further downstream work. Enzymatic the carbon cycle and reducing net emissions of greenhouse gases. Bio­ transesterification with lipases (glycerol ester hydrolases, E.C.3.1.1.3) diesel is obtained largely by transesterifying triacylglycerides with has emerged as a solution to avoid unwanted side reactions such as short-chain alcohols (ethanol and methanol, mainly). The resulting saponification and hence additional downstream processing. Although product is classified as first-, second- or third-generation biodiesel ac­ alternatives such as using acids and heterogeneous solid catalysts have cording to the oil substrate used. Due to the ethical dilemma of food vs been proposed, most are unfeasible or result in low reaction rates [9,10]. fuel, public institutions are taking steps to shift production from first- to Using lipases for transesterification has additional advantages, such as second- and third-generation biodiesel. The former is produced from more modest energy requirements from milder reaction conditions * Corresponding author. E-mail address: francisco.valero@uab.cat (F. Valero). https://doi.org/10.1016/j.fuel.2022.123794 Received 22 December 2021; Received in revised form 4 March 2022; Accepted 5 March 2022 Available online 12 March 2022 0016-2361/© 2022 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/). J. López-Fernández et al. Fuel 319 (2022) 123794 —lower reaction temperatures, close to 30 ◦ C, are required instead of the principles of CPV and PAT while avoiding the environmental and the higher temperatures, usually over 50 ◦ C, needed for chemical cata­ economic costs of withdrawing samples during the process. Unlike lysts [11]—, a simplified flowsheet for purification [12] and the need for previous works on the NIR based monitoring of biodiesel production, the no organic solvents, which makes processes compliant with the princi­ fact of performing a stepwise addition of alcohol and the employment of ples of green chemistry [13,14]. immobilized enzyme in this case might generate background noise in the Rhizopus oryzae lipase (ROL) has been extensively studied as a bio­ spectra, making the NIR monitoring more challenging. Besides, the catalyst for biodiesel production on the grounds of its 1,3-regiospecific­ biocatalyst operational stability and mechanical strength were evalu­ ity preventing formation of glycerol during the transesterification ated and compared with previously reported data. reaction [15] —this alcohol is deemed a residue owing to its worldwide large scale production [16]. Instead, ROL leads to the formation of 2- 2. Materials and methods monoacylglycerol, which is a by-product with a high added value by virtue of its capacity to increase biodiesel quality and its being useful for 2.1. Materials the pharmaceutical and food industries [17–19]. Although ROL is a stable biocatalyst, it is generally used in immobilized form to improve its Waste cooking oil (WCO) was obtained from a local public waste industrial suitability. Immobilizing the enzyme allows it to be reused management company and centrifuged prior to use. Polymethacrylate and usually increases its stability. ROL for biodiesel production has so matrix support D6308 was kindly donated by Purolite® (King of Prussia, far been immobilized in various ways including covalent binding, PA, USA). The colorimetric kit (11821729) for the enzymatic assay was adsorption or the use of whole cells [15]. supplied by Roche (Mannheim, Deutschland). Heptane, ethanol and During chemically or enzymatically catalyzed transesterification for methanol were purchased from Panreac (Barcelona, Spain). Standards of biodiesel production, the reaction medium contains a mixture of fatty methyl/ethyl palmitate, methyl/ethyl stearate, methyl/ethyl oleate, acid methyl and ethyl esters (FAM/EE), FFA and tri-, di- and mono­ methyl/ethyl linoleate and methyl linolenate, and all unstated reagents, acylglycerols. Production performance is usually assessed by gas chro­ were obtained from Sigma–Aldrich (St. Louis, MO, USA). matography (GC), using an offline and sluggish analytical procedure [20,21]. Moreover, low volatile substances can damage capillary col­ 2.2. Lipase heterologous production umns and other elements of the chromatographic system, so a number of alternative techniques including high-performance liquid chromatog­ The mature sequence of Rhizopus oryzae lipase (rROL) was heterol­ raphy (HPLC) [22] and proton nuclear magnetic resonance (1H NMR) ogously produced in the methylotrophic yeast Komagataella phaffii by spectroscopy [23] are being increasingly used instead. Chromatographic the Bioprocess Engineering and Applied Biocatalysis Group of the Uni­ techniques, however, are expensive and time-consuming; also, they versitat Autònoma de Barcelona (Barcelona, Spain) as described else­ cannot operate inline, which prevents timely decision-making during where [33]. After fermentation, the culture broth was centrifuged, the process and can increase production costs by requiring stops —or microfiltered, ultrafiltered and freeze-dried to remove biomass and even batch reprocessing if the target quality is not achieved [24]. The concentrate the enzyme [34]. principles of Continued Process Verification (CPV) have made bio­ process automatization and Process Analytical Technology (PAT) 2.3. Lipase activity measurement increasingly attractive; also, they have promoted the use of analytical methods allowing real-time quality assessment, adoption of corrective Enzymatic activity was determined by using the Roche lipase measures and a better understanding of bioprocesses [25,26]. Such colorimetric kit in 200 mM Tris-HCl buffer at pH 7.25 at 30 ◦ C on a Cary methods require in- or online monitoring of the process, and hence 300 spectrophotometer from Varian (Mulgrave, VIC, Australia) oper­ using, for instance, spectroscopic techniques (Raman, fluorescence, ating at 580 nm. All measurements were made in triplicate following an UV–Vis, IR, NIR) in combination with multivariate calibration [27]. already published procedure in which one unit of lipolytic activity was NIR spectroscopy is an extensively studied analytical technique defined as the amount of lipase necessary to hydrolyze 1 µmol of ester based on the interaction of matter and light radiation in the wavelength bond per minute under assay conditions [35]. region from 780 to 2500 nm that affords multicomponent, fast, reliable, inexpensive and non-destructive analysis [20]. Besides, it avoids the 2.4. Support functionalization and lipase immobilization need for sample withdrawal when used inline, waste production, and the need for complex pre-treatments of samples with solvents or other Polymethacrylate-matrix Purolite® D6308 with epoxide and octa­ chemicals, all of which makes it a safe, clean, energy-saving choice fully decyl surface groups (EO) was functionalized as described elsewhere compliant with the principles of green chemistry [28]. NIR spectra are [10]. Briefly, epoxide functional groups were pre-treated by incubating complex and possess broad overlapping bands that require special 1 g of support with 1 M ethylenediamine solution at pH 10 under orbital mathematical procedures to accurately interpret spectra and understand stirring at 60 ◦ C for 4 h. Then, the support was rinsed with distilled water the results, such as principal component analysis (PCA) or partial least- by vacuum filtration and incubated in a 2.5% w/v glutaraldehyde so­ squares (PLS) regression [29]. NIR spectroscopy has so far been suc­ lution in phosphate buffer at pH 8 on a roller at room temperature for 2 cessfully used by the biodiesel industry to assess the quality or properties h. Finally, the glutaraldehyde-treated support was rinsed for incubation of biofuel/diesel blends [29,30], and also for inline monitoring of with a 3500 AU mL− 1 lipase solution at 4 ◦ C for 42 h. The biocatalyst chemically catalyzed transesterification reactions [20,31,32]. thus obtained (EO-rROL) was dried on silica gel and stored at –20 ◦ C In this work, the mature sequence of ROL (rROL) heterologously until use. produced in the methylotrophic yeast Komagataella phaffii (Pichia pas­ The specific activity of the biocatalyst was calculated as the differ­ toris) was immobilized onto a suitable support in order to catalyze the ence between those in the final blank and supernatant solutions divided transesterification of waste cooking oil (WCO) in a solvent-free system by the final dry weight of enzyme. with ethanol or methanol as acyl-acceptor. The main novelty of this work is the scale up of the transesterification reaction to a laboratory- 2.5. Transesterification reactions scale reactor specially designed for use of an inline near infrared spec­ troscopy probe for monitoring the enzyme catalyzed transesterification Transesterification was done in closed 10 mL vials containing 8 g of reaction. The results were compared with those obtained by gas chro­ WCO at 30 ◦ C under orbital stirring at 350 rpm [14]. The acyl-acceptor matography as reference in order to confirm the suitability of the NIR was added in 5 (ethanol) or 10 pulses (methanol) by splitting the stoi­ technique for accurate real-time monitoring of transesterification under chiometric volume of alcohol (2:1 alcohol:oil mixture) during the 2 J. López-Fernández et al. Fuel 319 (2022) 123794 Table 1 the spectral resolution was 2 nm and the optical path length was 1 mm. Percent fatty acid composition profile for sunflower, olive and waste cooking oil Raw absorbance spectra were exported to NSAS file format by using the (WCO). software Vision 2.51, then transformed to Matlab file format with The Acid Sunflower oil Olive oil WCO Unscrambler 10.3 (Camo Analytics, Norway) and finally processed for Stearic 2.8 2.3 2.46 ± 0.04 viewing, exploration and multivariate modeling with the software Solo Oleic 28.0 66.4 70.78 ± 0.38 (Eigenvector Research Inc., Wenatchee, WA, USA). Linoleic 62.2 16.4 10.21 ± 0.01 First- and second-derivative spectra were obtained by using the Linolenic 0.16 1.6 2.45 ± 0.02 Savitzky–Golay algorithm with a 15-point moving window and a second-order polynomial. Mean-centered data (X-matrix for spectra) were subjected to principal component analysis (PCA). Partial least- reaction in order to reach the theoretical maximum yield (200 min re­ squares (PLS) calibration models for mean-centered data (X-matrix for action for ethanol and 375 min for methanol) [17]. All substrates and spectra and Y-matrix for reaction yield) were constructed by cross- biocatalysts were previously equilibrated for water activity by using a validation, using the leave-one-out method. Samples for inclusion in saturated KOH solution (aw = 0.093) for a minimum of 16 h. The re­ the calibration and prediction sets were selected by using the Ken­ action was scaled up to a 50 mL HME-R mini reactor from Scharlab nard–Stone method [41]. The quality of the PLS models and their pre­ (Sentmenat, Barcelona, Spain) as described above but using a final oil dictive ability were assessed in terms of the root mean square error of mass of 40 g and mechanical stirring at 700 rpm. calibration (RMSEC) and prediction (RMSEP), defined as:. The minireactor lid and tank were specially designed in collabora­ √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ tion with Scharlab (Supplementary Material 1). Thus, the lid was ∑n ( pred )2 custom-made to include an inlet that allowed the NIR probe to be i=1 Yi − Yref RMSE = i (3) inserted in the reaction medium, and a second, slightly curved inlet for n the stirrer, so that both could reach the reaction medium without where n is the number of samples used, Yref the reaction yield provided colliding. Besides, the stirrer was modified by using an additional pro­ by the reference method and Ypred that estimated by the NIR model. peller —the two were “marine propellers”— to facilitate generation of a The number of PLS factors required to define each model was chosen turbulent regime in the reactor. from the minimum of a plot of RMSEP vs number of factors. The pre­ Operational stability was assessed by allowing the biocatalyst to dictive ability was also assessed by statistical evaluation of the least- settle in the bottom of the reactor. Then, depleted reaction medium was squares regression line between the reference GC reaction yield and removed and all components were prepared for the next run [17]. the NIR-predicted value. Relative yields of consecutive transesterification reactions cycles were The aim of quantitative multivariate modeling is reducing prediction used as described elsewhere to fit the results to a first-order exponential errors by using the simplest possible model (i.e., that with the fewest decay equation (Eq. (1)) and a two-component first-order exponential factors). PLS models were evaluated over wide or narrow spectral in­ decay equation (Eq. (2)), using the software Sigma Plot v 14 [36–38]. tervals selected from different calculations, namely: regression co­ Y(%)t = 100e− kt (1) efficients, X-loadings, XY correlation vector, variable importance in projection and selectivity ratio. These tools allow simple numerical Y(%)t = 100e− k1 t + ce− k2 t (2) assessment of the usefulness of each X-variable in a regression model. The higher were the results obtained with these calculations over the where k, k1 and k2 are deactivation coefficients. significant threshold, the greater was the usefulness of the variables for regression and prediction. Selection of the spectral interval was com­ 2.6. Gas chromatography (GC) analysis bined with different spectral processing methods, first in the absorbance mode, and then as their first and second derivatives, and also upon Fatty acid ethyl esters (FAEEs) and fatty acid methyl esters (FAMEs) standard normal variate SNV-based scaling. The combination of SNV were quantified on a model 7890A gas chromatograph from Agilent followed by derivatization was also tested. (Santa Clara, CA, USA) equipped with a 19095N-123 capillary column and an autosampler [39]. Relative standard deviations (RSD) never 3. Results and discussion exceeded 3%. 3.1. Fatty acid composition of the waste cooking oil 2.7. Fatty acid composition Local public waste management companies typically collect and mix The fatty acid composition of WCO was established by adapting a large amounts of waste cooking oil (WCO). As a result, WCO composi­ previously reported procedure of chemically catalyzed trans­ tion can vary widely depending on consumption patterns and regional or esterification [40]. Briefly, 40 g of oil was heated in an HME-R mini­ local cooking traditions. In Spain, most oil based cooking is done with reactor commercially available from Scharlab (i.e., with the original olive or sunflower oil, which leave vast amounts of oleaginous waste tank and lid) at 65 ◦ C under mechanical stirring at 350 rpm. Then, KOH [42]. (1 % w/w) and methanol (6:1 alcohol/oil mixture) were added and the Because it was obtained from a public waste management company reaction was allowed to develop for 6 h with periodic sample with­ not implementing traceability, the oil used here was of unknown origin. drawal for GC analysis and calculation of the relative fatty acid Analysis of its fatty acid composition was performed as described in composition. Section 2.6 to compare it with published profile of olive and sunflower oil and results exhibited that the fatty acid profile for the WCO was 2.8. Acquisition and processing of near infrared (NIR) spectra similar to that for olive oil (Table 1) [43]. The acidity of WCO, which is another relevant factor having influence on the performance of the The transesterification reactions were monitored by recording NIR biocatalyst during transesterification, has already been analyzed in a spectra at 5 min intervals in each reaction cycle [28]. Spectra were ac­ previous publication exhibiting a low acidity value (0.77%) suggesting quired in the transflectance mode, using a Model 5000 spectropho­ that oil was not extensively cooked [14]. tometer from FOSS NIRSystems (Silver Spring, MD, USA) equipped with an immersive optical probe. A reference spectrum for air was obtained before the reaction. The wavelength range scanned was 1100–2498 nm, 3 J. López-Fernández et al. Fuel 319 (2022) 123794 Fig. 1. Yield profile for WCO transesterification with 5 pulses of ethanol (A) and 10 of methanol (B) in the presence of EO-rROL. Relative yield of consecutive transesterification cycles with 5 pulses of ethanol (C) and 10 of methanol (D). White bars and points correspond to the reaction in 10 mL vials and grey bars and points to that in the laboratory minireactor. Relative yields were calculated against that of the first cycle (100%). Fig. 2. (A) Customized reactor for use of the NIR probe. (B) The EO-rROL biocatalyst after 10 consecutive reaction cycles as seen under a binocular loupe. 3.2. Transesterification reaction: scale-up, productivity and half-life addition strategy for biodiesel production from WCO were employed [14], and as can be seen from Fig. 1A and B, the yield profiles for the first The transesterification reactions initially conducted in 10 mL vials reaction batch were almost identical with ethanol and methanol in both under orbital stirring were successfully scaled up to a 50 mL laboratory the vial and the reactor —because of the 1,3-regioespecificity of ROL, reactor under mechanical stirring (mimicking industrial large-scale the theoretical maximum yield in the absence of acyl-migration phe­ conditions), which was specially designed for use with a NIR probe. nomena is 66%, resulting in the formation of the corresponding alkyl Already optimized conditions, regarding the biocatalyst and the alcohol esters and 2-monoacylglycerol [44]. The results therefore suggest that 4 J. López-Fernández et al. Fuel 319 (2022) 123794 Table 2 EO-rROL productivity and half-life in biodiesel production with ethanol and methanol as calculated by fitting to a first-order exponential decay (Eq. (1)) and two-component first-order exponential decay (Eq. (2)) deactivation model. Acyl Model R2 Half-life (h) Productivity (μmol acceptor equation * min− 1) Ethanol 1 0.87 170 [51] 327 2 0.96 146.7 [44] Methanol 1 0.87 225 [36] 219 2 0.95 337.5 [54] * The numbers in square brackets represent the number of cycles for those of reaction hours. Fig. 4. Second-derivative NIR spectra obtained in the first reaction batch with (A) ethanol and (B) methanol as acyl-acceptor. The arrows represent reac­ tion time. Table 3 Dimensions of the X-dataset and Y-dataset. Reaction with ethanol Reaction with methanol Reaction time (min) 200 375 X-matrix (1100–2350 nm) 322 rows and 626 columns 684 rows and 626 columns Y-matrix (Reaction yield %) 123 rows and 1 column 243 rows and 1 column negative impact on the operational stability of the biocatalyst— but can be useful with others such as methanol with severe influence on enzyme stability [17,45]. Thus, the biocatalyst half-life in hours with methanol Fig. 3. NIR absorbance spectra recorded at the start (green) and end (red) of the reaction as compared with those for each pure alcohol (blue): ethanol (A) as acyl-acceptor was about 1.5–2.5 times higher than it was with ethanol and methanol (B). The arrows represent evolution of the reactions. as a result of the amount of methanol used being split into more pulses than that of ethanol (Table 2). However, when reaction cycles corre­ sponding to these hours were compared, half-lives (t1/2 values into mass transfer and reaction performance in the vials and the reactor were square brackets in Table 2) tended to be similar with both alcohols as a essentially identical under the employed conditions, and also that the consequence of the reaction times differing between the two. presence of the NIR probe did not detract from homogeneity in the re­ ROL has exhibited widely variable operational stability in biodiesel action medium. Thus, the 50 mL reactor with a customized lid and tank, production from various substrates. For instance, ROL covalently and a second stirrer (Fig. 2A), reproducibly echoed the results obtained immobilized onto RelizymeTM OD403 was used to transesterify olive with 10 mL vials. In fact, both systems led to similar relative yields after pomace oil with methanol and the relative yield found to decrease to 5 consecutive cycles (Fig. 1C and 1D). 60% after 26.7 h reaction (7 batches) [46]. In another study with soy­ Biodiesel productivity with each alcohol was calculated as the bean oil, the transesterification yield with glutaraldehyde cross-linked combined figure for all 10 cycles performed in the laboratory-scale whole-cell biocatalysts decreased from 84 to 65% after 2520 h reac­ reactor. As expected, FAEE productivity (ethanol as acyl-acceptor) was tion (35 batches) [47]. By contrast, the ROL catalyzed transesterification 1.5 times higher than FAME productivity (methanol as acyl-acceptor) as of WCO has been the subject of little study, so comparison with the a result of the longer reaction times required for transesterification with previous results is challenging [15]. Therefore, although a number of the latter alcohol (Table 2), even though, the final transesterification half-life values for ROL based biocatalysts used in biodiesel trans­ yields with methanol were slightly higher than those obtained with esterification reactions have been reported with several substrates, ethanol (Fig. 1A and B). In fact, splitting the amount of alcohol used into comparison with WCO would be more accurate since this substrate may 10 pulses is unproductive with alcohols such as ethanol —it has scarce 5 J. López-Fernández et al. Fuel 319 (2022) 123794 Fig. 5. PCA scores plot constructed from second-derivative spectra over the range 1100–2350 nm obtained with ethanol (A1 and A2) and methanol as acyl-acceptor (B1 and B2). The 1 and 2 subplots stand for different symbol related to cycles and reaction time, respectively. deactivate the enzyme more markedly as the likely result of its con­ spectra obtained in a reaction run with ethanol and methanol as taining detrimental components such as phenolics or varying widely in compared with those for the pure alcohols. Ethanol and methanol acidity [10,48,49]. (Fig. 3, blue plots) exhibited the following major spectral bands: The biocatalysts were imaged with a binocular loupe after the whole 1425–1475 nm (OH first overtone), 1650–1750 nm (CH first overtone), sequence of reaction with each acyl-acceptors in order to check whether 2000–2100 nm (OH combinations) and 2200–2450 nm (CH + CH they had retained their structure without fracturing by effect of me­ combinations and CC + CH combinations). Expectedly, methanol chanical stress under agitation (Fig. 2B). No substantial breaks were spectra presented less bands than that of ethanol at 1650–1750 nm and detected after the ten cycles, which confirms the strength of the support 2200–2450 nm due to the absence of CH2; the bands corresponding with and its suitability for industrial use, as indicated by the supplier CH first overtone, CH + CH and CH + CC combinations are only related (Purolite®). with the CH3 group from methanol. The spectra at the start of the re­ action (Fig. 3, green plots) exhibited the following spectral bands: high absorbance at 1650–1750 nm (CH first overtone), very low absorbance 3.3. Near infrared (NIR) spectra: A useful tool for the biodiesel industry at 1875–1950 nm (CO second overtone), low absorbance at 2000–2100 nm (OH combinations) and very high absorbance at 2200–2450 nm (CH Once the biocatalyst suitability for biodiesel production under con­ + CH and CC + CH combinations). The spectra at the end of the reaction ditions mimicking industrial work was confirmed, the development of a (Fig. 3, red plots) showed higher absorbance than that of the starting system enabling inline monitoring and compliance with CPV and PAT point almost throughout the whole studied spectral range (1100–2400 principles was sought. In this sense, NIR spectroscopy was employed by nm). The final yield after the two batches showed in Fig. 3 was 61.0% inserting a probe in a customized reactor in combination with two with ethanol and 65.9% with methanol. different calibration models. The typically extensive overlap of bands and fairly low resolution of the NIR technique required increasing resolution by using a spectral 3.3.1. Spectral data derivative treatment. The spectral ranges 1400–1700 nm and Absorbance spectra acquired across the NIR spectral range 1900–2300 nm were adequate to visualize the spectral changes during (1100–2498 nm) were visually inspected to identify gross outliers and the reaction because of the low absorbance of the WCO and high noisy spectral regions. The range from 2352 to 2498 nm exhibited little absorbance of the alcohols. Fig. 4 shows the second derivative spectra of variability and considerable noise typical of fiber optic probes, so only the first cycle with ethanol (Fig. 4A) and with methanol (Fig. 4B), in the that from 1100 to 2350 nm was used. Fig. 3 shows the first and last 6 J. López-Fernández et al. Fuel 319 (2022) 123794 Table 4 mostly explained by the PC1, accounting for 81% of the X-variance, and Number of spectra used to construct the calibration and prediction sets, figures the reaction time by the PC2 (21%). For the reaction with methanol (5B1 of merit of the calibration models and predictive ability. and 5B2), the reaction time is mostly explained by the PC1 (74%) and Reaction Reaction the reaction cycle by the PC2 (12%). Neither standard normal variate with ethanol with (SNV) scaling nor first-derivative treatment, whether individually or in methanol combination, provided better results for the reaction with ethanol —at Calibration least not on a par with the scores plots for methanol. The spectral Number of spectra 24 (4 cycles) 27 (3 cycles) variability of the reaction cycle depended on the normal laboratory PLS factors 2 2 conditions of sample and instrument. Yield range (%) 0.0–61.0 6.2–68.9 3.3.2. Calibration Cumulative Y-variance explained The Kennard–Stone method selects a subset of calibration samples Factor 1 66.9 91.4 Factor 2 98.1 99.6 which provide a very uniformly distributed network of selected points over the dataset and includes samples on its boundary. Table 4 shows the NIR vs GC regression number of spectra used for calibration and prediction. The calibration Slope 0.98 ± 0.06 0.99 ± 0.03 set for the ethanol reaction included 24 samples from reaction cycles 1, Intercept 0.6 ± 2.1 0.2 ± 1.2 2, 3 and 6, which spanned an FAEE yield range of 0.0–61.0%, while the Correlation coefficient (r) 0.990 0.998 set for methanol comprised 27 samples from reaction cycles 2, 3 and 10, RMSEC (%) 2.4 1.3 with an FAME yield of 6.2–68.9%. No spectrum for a near-zero yield was available for methanol owing to instrumental issues arising at the Prediction beginning of the reaction. All other spectra were included in the pre­ Number of spectra (total) 298 (8 657 (9 cycles) cycles) diction set. Number of spectra (with GC reaction yield value) 99 203 Reaction yields were quantified by using PLS calibration models Number of spectra (without GC reaction yield 199 454 constructed from independent calibration and prediction sets of NIR value) spectra. Models spanning a narrow spectral range failed to reduce cali­ Yield range (%) 0.0–59.4 6.3–68.8 bration or prediction errors relative to the whole spectral range. Also, narrow ranges failed to reduce the number of factors required by each NIR vs GC regression model. Second-derivative spectra proved the best choice in any case. Slope 0.98 ± 0.03 0.99 ± 0.01 Table 4 summarizes the calibration and prediction results obtained Intercept 1.0 ± 1.1 –0.3 ± 0.6 Correlation coefficient (r) 0.989 0.996 with the PLS model for each alcohol. Both were constructed with two RMSEP (%) 2.1 2.0 factors and explained a cumulative Y-variance higher than 98%. The only difference between the two was the Y-variance captured by the first factor, which was 66.9% with ethanol and 91.4% with methanol. This wavelength range between 2200 and 2275 nm. As can be seen, the bands was a result of the distribution of PCA scores and the two sources of were highly ordered in terms of yield and reaction time. This result is variability observed (reaction time and reaction cycle; Fig. 5). The Y- also observed throughout the entire wavelength range of 1100–2350 variable (yield) was explained mainly by PC1 in the reaction with nm. methanol, and by both PC1 and PC2 in the reaction with ethanol. The Table 3 summarizes the dimensions of the X- and Y- datasets. The model using PC1 alone was more robust for methanol than it was for number of rows coincided with that of measured spectra and the number ethanol. With PC1 and PC2 jointly, however, the two models captured of columns with that of variables (wavelengths and reaction yield). The almost the same Y-variance (98.1% for ethanol and 99.6% for meth­ reaction with methanol took almost twice as long as that with ethanol anol). The upper range of biodiesel yield for the reaction with methanol (375 min vs 200 min). Since the spectrum acquisition frequency was (68.9%) was higher by effect of the increased experimental reaction identical with both alcohols (1 spectrum every 5 min), the available X- yield. Calibration errors (RMSEC), expressed in the same units as yield, matrix for the reaction with methanol contained many more rows than were 2.4% for ethanol and 1.3% for methanol and the correlation co­ that for ethanol (684 vs 322). The sampling frequency for GC analysis efficients were higher than 0.99. was approximately 1 sample every 15 min. As a result, the number of The prediction results of Table 4 were obtained from pure indepen­ rows in the Y-matrix for which a yield value was available was roughly dent sets which were not employed during calibration. Regression lines one-third of all (viz., 123 out of 322 for ethanol and 243 out of 684 for for NIR yield versus GC yield plots were evaluated through different tests methanol). and analysis. In this sense, the correlation coefficients were high (viz., Data from second-derivative spectra spanning the range 1100–2350 0.989 and 0.996 for ethanol and methanol, respectively). The residuals nm were initially explored through PCA method. One PCA for each were randomly distributed around 0 and their normal distribution was alcohol was calculated by using the whole X-matrix. The scores plot of demonstrated with the Anderson-Darling, Ryan-Joiner and Kolmogorov- the second principal component against the first (PC2 vs PC1) explained Smirnov tests (95% of significance). The Mandel test for linearity was approximately 95% of the variance in X (Fig. 5). Fig. 5A were calculated calculated for the two regressions and the F-tests (95% of significance) from the reactions with ethanol, and 5B with methanol. There were two concluded that quadratic models do not performe better than linear main sources of spectral variability, namely: reaction time and reaction ones. The absence of significant differences between the two techniques cycle. For the reaction with ethanol (5A1 and 5A2), the reaction cycle is was demonstrated through the t-tests (95% of significance) of the slope Table 5 RMSEP (%) value for each reaction cycle. RMSEP Cycle Overall 1 2 3 4 5 6 7 8 9 10 Reaction with ethanol 2.4 2.0 2.7 2.1 1.9 2.1 1.6 N/A N/A 1.6 2.1 Reaction with methanol 2.0 1.0 1.9 2.1 2.1 N/A 2.5 1.9 2.0 2.3 2.0 N/A: not available. 7 J. López-Fernández et al. Fuel 319 (2022) 123794 Fig. 6. NIR prediction of reaction yields (solid line) with ethanol as acyl-acceptor. Dots represent reference GC yield values. and intercept of the NIR versus GC regressions. In fact, the slopes and batch, and peaked at the beginning and end. This result is consistent intercepts were not significantly different from 1 and 0, respectively, with the typical distribution of interpolation errors observed with least- which highlights the accuracy of the NIR method. Finally, the satisfac­ squares regression. The PCA scores plot (Fig. 5) affected the larger tory predictive ability was also assessed with the low RMSEP prediction estimated error for the reaction with ethanol compared with the reaction errors (2.1% for ethanol and 2.0% for methanol). As can be seen in with methanol. The Y-variable (yield) in the reaction with ethanol was Table 5, RMSEP for the individual reaction cycles ranged from 1.0 to explained by both PC1 and PC2. This led to a PLS model for the reaction 2.7%. Although RMSEC for the methanol model was lower than for the of ethanol with a Y-explained variance of the first factor of 66.9% ethanol model (1.3% vs 2.4%), there were virtually no differences in (Table 4). The yield in the reaction with methanol was explained mostly predictive ability between the two on an individual batch basis. by PC1 and the Y-explained variance of the PLS model was 91.4%. The PLS for both reactions used two factors, achieving 98.1% and 99.6% of 3.3.3. Validation the Y-explained variance. The previous results testify to the robustness Figures 6 and 7 show the NIR profiles of yield vs time used to validate of NIR spectroscopy for monitoring biodiesel production by enzymatic the results of the reaction cycles with the two alcohols. The figures transesterification with ethanol or methanol as acyl-acceptor. represent the NIR predictions from the whole set of prediction spectra (with or without a matching reaction yield obtained by GC), namely: 4. Conclusions 298 spectra for ethanol and 657 for methanol. These spectra were distributed among the 8 reaction cycles examined for ethanol and the 9 Waste cooking oil (WCO), which based on its fatty acid composition for methanol. As can be seen, there was high correlation between the was probably of olive oil origin, was used to obtain third-generation NIR and GC results. Uncertainty in the prediction of each spectrum was biodiesel with immobilized Rhizopus oryzae lipase and to successfully assessed via the estimated error, which is a measure of goodness of scale up the reaction from orbitally stirred 10 mL vials to a mechanically prediction and calibration error, and was used in reaction yield units. stirred 50 mL laboratory reactor obtaining similar yields and operational This error measure is software specific (Solo) and it uses the equation 9 stability. EO-rROL exhibited a high operational stability in terms of half- of a previously reported work [50]. The 298 NIR predictions of the re­ life (t1/2), over 35 reaction batches with both ethanol and methanol. action with ethanol were subject to an estimated average error of 2.94% However, methanol-based reaction showed decreased productivity as a (min = 2.84, max = 3.13), while the 657 NIR predictions of the reaction result of the increased reaction times (219 vs 327 µmol min− 1). Besides, with methanol had an estimated average error of 1.51% (min = 1.47, transesterification reaction with both alcohols was monitored by NIR max = 1.65). The magnitude of the error was not constant throughout spectroscopy in the 50 mL reactor. The GC and NIR results were highly the reaction, however; rather, it was smallest in the middle of each correlated, with a prediction error (RMSEP) of 2.0% for methanol and 8 J. López-Fernández et al. Fuel 319 (2022) 123794 Fig. 7. NIR prediction of reaction yields (solid line) with methanol as acyl-acceptor. Dots represent GC yield values. 2.1% for ethanol. Based on these results, using immobilized enzymes Acknowledgments and adding the alcohol stepwise, which might have increased back­ ground noise in spectral measurements, was no issue here. Conse­ This work was funded by the Spanish Ministry of Science and Inno­ quently, NIR spectroscopy stands as a robust tool for monitoring vation (Project PID2019-104666GB-100). JLF was additionally awarded industrial biodiesel production in compliance with CPV and PAT prin­ a Basque Government Scholarship for the training of predoctoral re­ ciples. Inline monitoring of the reaction can help identify the most searchers (PRE_2017_1_0110). suitable time for addition of alcohol pulses and renewal of the biocata­ lyst for optimal performance. Appendix A. Supplementary data Supplementary data to this article can be found online at https://doi. CRediT authorship contribution statement org/10.1016/j.fuel.2022.123794. Josu López-Fernández: Conceptualization, Methodology, Valida­ References tion, Formal analysis, Investigation, Writing – original draft, Visualiza­ tion. Desirèe Moya: Methodology, Investigation. María Dolors [1] Connolly D, Lund H, Mathiesen BV. Smart Energy Europe: the technical and economic impact of one potential 100% renewable energy scenario for the Benaiges: Conceptualization, Methodology, Validation, Writing – re­ European Union. Renew Sustain Energy Rev 2016;60:1634–53. https://doi.org/ view & editing, Supervision, Project administration. Francisco Valero: 10.1016/j.rser.2016.02.025. Conceptualization, Methodology, Validation, Writing – review & edit­ [2] Kumar B, Verma P. Biomass-based biorefineries: an important architype towards a ing, Supervision, Project administration, Funding acquisition. Manel circular economy. Fuel 2021;288:119622. https://doi.org/10.1016/j. fuel.2020.119622. Alcalà: Conceptualization, Methodology, Software, Validation, Formal [3] Hoang AT, Nižetić S, Ölçer AI. 2,5-Dimethylfuran (DMF) as a promising biofuel for analysis, Investigation, Resources, Writing – original draft, Writing – the spark ignition engine application: a comparative analysis and review. Fuel review & editing, Visualization, Project administration. 2021;285:119140. https://doi.org/10.1016/J.FUEL.2020.119140. [4] Fadhil AB, Nayyef AW, Sedeeq SH. Valorization of mixed radish seed oil and Prunus armeniaca L. oil as a promising feedstock for biodiesel production: evaluation and analysis of biodiesels. Asia-Pacific J Chem Eng 2020;15(1). https://doi.org/ Declaration of Competing Interest 10.1002/APJ.2390. [5] Mercy Nisha Pauline J, Sivaramakrishnan R, Pugazhendhi A, Anbarasan T, Achary A. Transesterification kinetics of waste cooking oil and its diesel engine The authors declare that they have no known competing financial performance. Fuel 2021;285:119108. interests or personal relationships that could have appeared to influence the work reported in this paper. 9 J. López-Fernández et al. Fuel 319 (2022) 123794 [6] Unglert M, Bockey D, Bofinger C, Buchholz B, Fisch G, Luther R, et al. Action areas [29] Palou A, Miró A, Blanco M, Larraz R, Gómez JF, Martínez T, et al. Calibration sets and the need for research in biofuels. Fuel 2020;268:117227. https://doi.org/ selection strategy for the construction of robust PLS models for prediction of 10.1016/j.fuel.2020.117227. biodiesel/diesel blends physico-chemical properties using NIR spectroscopy. [7] Ogunkunle O, Ahmed NA. A review of global current scenario of biodiesel adoption Spectrochim Acta - Part A Mol Biomol Spectrosc 2017;180:119–26. and combustion in vehicular diesel engines. Energy Rep 2019;5:1560–79. https:// [30] Rocha W, Sheen D. Determination of physicochemical properties of petroleum doi.org/10.1016/j.egyr.2019.10.028. derivatives and biodiesel using GC/MS and chemometric methods with uncertainty [8] Günay ME, Türker L, Tapan NA. Significant parameters and technological estimation. Fuel 2019;243:413–22. https://doi.org/10.1016/j.fuel.2018.12.126. advancements in biodiesel production systems. Fuel 2019;250:27–41. https://doi. [31] Richard R, Dubreuil B, Thiebaud-Roux S, Prat L. On-line monitoring of the org/10.1016/j.fuel.2019.03.147. transesterification reaction carried out in microreactors using near infrared [9] Singh D, Sharma D, Soni SL, Sharma S, Kumar Sharma P, Jhalani A. A review on spectroscopy. Fuel 2013;104:318–25. https://doi.org/10.1016/j.fuel.2012.07.054. feedstocks, production processes, and yield for different generations of biodiesel. [32] Sales R, da Silva NC, da Silva JP, França HH, Pimentel MF, Stragevitch L. Handheld Fuel 2020;262:116553–68. https://doi.org/10.1016/j.fuel.2019.116553. near-infrared spectrometer for on-line monitoring of biodiesel production in a [10] Bonet-Ragel K, Canet A, Benaiges MD, Valero F. Synthesis of biodiesel from high continuous process. Fuel 2019;254:115680. https://doi.org/10.1016/J. FFA alperujo oil catalysed by immobilised lipase. Fuel 2015;161:12–7. https://doi. FUEL.2019.115680. org/10.1016/j.fuel.2015.08.032. [33] Arnau C, Ramon R, Casas C, Valero F. Optimization of the heterologous production [11] Biernat K, Matuszewska A, Samson-Bręk I, Owczuk M. Biological methods in of a Rhizopus oryzae lipase in Pichia pastoris system using mixed substrates on biodiesel production and their environmental impact. Appl Sci 2021;11:10946. controlled fed-batch bioprocess. Enzyme Microb Technol 2010;46:494–500. https://doi.org/10.3390/APP112210946. https://doi.org/10.1016/j.enzmictec.2010.01.005. [12] Harding KG, Dennis JS, von Blottnitz H, Harrison STL. A life-cycle comparison [34] Guillén M, Benaiges MD, Valero F. Biosynthesis of ethyl butyrate by immobilized between inorganic and biological catalysis for the production of biodiesel. J Clean recombinant Rhizopus oryzae lipase expressed in Pichia pastoris. Biochem Eng J Prod 2008;16:1368–78. https://doi.org/10.1016/J.JCLEPRO.2007.07.003. 2012;65:1–9. https://doi.org/10.1016/j.bej.2012.03.009. [13] Cavalcante FTT, Neto FS, Rafael de Aguiar Falcão I, Erick da Silva Souza J, de [35] Resina D, Serrano A, Valero F, Ferrer P. Expression of a Rhizopus oryzae lipase in Moura Junior LS, da Silva Sousa P, et al. Opportunities for improving biodiesel Pichia pastoris under control of the nitrogen source-regulated formaldehyde production via lipase catalysis. Fuel 2021;288:119577. dehydrogenase promoter. J Biotechnol 2004;109:103–13. https://doi.org/ [14] López-Fernández J, Dolors Benaiges M, Valero F. Second- and third-generation 10.1016/j.jbiotec.2003.10.029. biodiesel production with immobilised recombinant Rhizopus oryzae lipase: [36] Rodrigues J, Canet A, Rivera I, Osório NM, Sandoval G, Valero F, et al. Biodiesel influence of the support, substrate acidity and bioprocess scale-up. Bioresour production from crude Jatropha oil catalyzed by non-commercial immobilized Technol 2021;334:125233. https://doi.org/10.1016/j.biortech.2021.125233. heterologous Rhizopus oryzae and Carica papaya lipases. Bioresour Technol 2016; [15] López-Fernández J, Benaiges MD, Valero F. Rhizopus oryzae lipase, a promising 213:88–95. industrial enzyme: biochemical characteristics, production and biocatalytic [37] Aymard C, Belarbi A. Kinetics of thermal deactivation of enzymes: a simple three applications. Catalysts 2020;10:1277. https://doi.org/10.3390/catal10111277. parameters phenomenological model can describe the decay of enzyme activity, [16] Rodrigues A, Bordado JC, Dos Santos RG. Upgrading the glycerol from biodiesel irrespectively of the mechanism. Enzyme Microb Technol 2000;27:612–8. https:// production as a source of energy carriers and chemicals - A technological review doi.org/10.1016/S0141-0229(00)00258-1. for three chemical pathways. Energies 2017;10:1817–53. https://doi.org/ [38] López-Fernández J, Benaiges MD, Valero F. Constitutive expression in Komagataella 10.3390/en10111817. phaffii of mature Rhizopus oryzae lipase jointly with its truncated prosequence [17] Bonet-Ragel K, Canet A, Benaiges MD, Valero F. Effect of acyl-acceptor stepwise improves production and the biocatalyst operational stability. Catal 2021;11:1192. addition strategy using alperujo oil as a substrate in enzymatic biodiesel synthesis. https://doi.org/10.3390/CATAL11101192. J Chem Technol Biotechnol 2018;93:541–7. https://doi.org/10.1002/jctb.5399. [39] Canet A, Dolors Benaiges M, Valero F. Biodiesel synthesis in a solvent-free system [18] Dunn RO. Effects of monoacylglycerols on the cold flow properties of biodiesel. by recombinant Rhizopus oryzae lipase. Study of the catalytic reaction progress. JAOCS, J Am Oil Chem Soc 2012;89:1509–20. https://doi.org/10.1007/s11746- JAOCS, J Am Oil Chem Soc 2014;91:1499–506. https://doi.org/10.1007/s11746- 012-2045-7. 014-2498-y. [19] Feltes MMC, de Oliveira D, Block JM, Ninow JL. The production, benefits, and [40] Vicente G, Martínez M, Aracil J. Integrated biodiesel production: a comparison of applications of monoacylglycerols and diacylglycerols of nutritional interest. Food different homogeneous catalysts systems. Bioresour Technol 2004;92:297–305. Bioprocess Technol 2013;6:17–35. https://doi.org/10.1007/s11947-012-0836-3. https://doi.org/10.1016/j.biortech.2003.08.014. [20] Pinzi S, Alonso F, García Olmo J, Dorado MP. Near infrared reflectance [41] Ng W, Minasny B, Malone B, Filippi P. In search of an optimum sampling algorithm spectroscopy and multivariate analysis to monitor reaction products during for prediction of soil properties from infrared spectra. PeerJ 2018;2018.. https:// biodiesel production. Fuel 2012;92:354–9. https://doi.org/10.1016/j. doi.org/10.7717/peerj.5722. fuel.2011.07.006. [42] Consumo de aceites en España|Anierac https://anierac.org/consumo-en-espana/? [21] Tian M, Fu J, Wang Z, Miao C, Lv P, He D, et al. Enhanced activity and stability of lang=en (accessed September 17, 2021). Rhizomucor miehei lipase by mutating N-linked glycosylation site and its [43] Orsavova J, Misurcova L, Vavra Ambrozova J, Vicha R, Mlcek J. Fatty acids application in biodiesel production. Fuel 2021;304:121514. https://doi.org/ composition of vegetable oils and its contribution to dietary energy intake and 10.1016/J.FUEL.2021.121514. dependence of cardiovascular mortality on dietary intake of fatty acids. Int J Mol [22] Syed MB. Analysis of biodiesel by high performance liquid chromatography using Sci 2015;16:12871–90. https://doi.org/10.3390/ijms160612871. refractive index detector. MethodsX 2017;4:256–9. https://doi.org/10.1016/J. [44] Canet A, Benaiges MD, Valero F, Adlercreutz P. Exploring substrate specificities of MEX.2017.07.002. a recombinant Rhizopus oryzae lipase in biodiesel synthesis. N Biotechnol 2017;39: [23] Morgenstern M, Cline J, Meyer S, Cataldo S. Determination of the kinetics of 59–67. https://doi.org/10.1016/j.nbt.2017.07.003. biodiesel production using proton nuclear magnetic resonance spectroscopy (1H [45] Lotti M, Pleiss J, Valero F, Ferrer P. Effects of methanol on lipases: molecular, NMR). Energy Fuels 2006;20:1350–3. https://doi.org/10.1021/ef0503764. kinetic and process issues in the production of biodiesel. Biotechnol J 2015;10: [24] de Lima FW, Corgozinho CNC, Tauler R, Sena MM. Monitoring biodiesel and its 22–30. https://doi.org/10.1002/biot.201400158. intermediates in transesterification reactions with multivariate curve resolution [46] Bonet-Ragel K, López-Pou L, Tutusaus G, Benaiges MD, Valero F. Rice husk ash as a alternating least squares calibration models. Fuel 2021;283:119275. https://doi. potential carrier for the immobilization of lipases applied in the enzymatic org/10.1016/j.fuel.2020.119275. production of biodiesel. Biocatal Biotransformation 2018;36:151–8. https://doi. [25] Rosas JG, Blanco M, González JM, Alcalà M. Real-time determination of critical org/10.1080/10242422.2017.1308498. quality attributes using near-infrared spectroscopy: a contribution for Process [47] Sun T, Du W, Liu D. Comparative study on stability of whole cells during biodiesel Analytical Technology (PAT). Talanta 2012;97:163–70. https://doi.org/10.1016/j. production in solvent-free system. Process Biochem 2011;46:661–4. https://doi. talanta.2012.04.012. org/10.1016/j.procbio.2010.11.006. [26] Gorsky I. Process validation stage 3: continued process verification. Princ Parenter [48] Lin YH, Luo JJ, John Hwang SC, Liau PR, Lu WJ, Lee HT. The influence of free fatty Solut Valid A Pract Lifecycle Approach 2020:209–32. https://doi.org/10.1016/ acid intermediate on biodiesel production from soybean oil by whole cell B978-0-12-809412-9.00008-3. biocatalyst. Biomass Bioenergy 2011;35:2217–23. https://doi.org/10.1016/j. [27] De Lima SM, Silva BFA, Pontes DV, Pereira CF, Stragevitch L, Pimentel MF. In-line biombioe.2011.02.039. monitoring of the transesterification reactions for biodiesel production using NIR [49] Tan Y, Chang SKC, Zhang Y. Comparison of α-amylase, α-glucosidase and lipase spectroscopy. Fuel 2014;115:46–53. https://doi.org/10.1016/j.fuel.2013.06.057. inhibitory activity of the phenolic substances in two black legumes of different [28] Blanco M, Alcalá M, González JM, Torras E. A process analytical technology genera. Food Chem 2017;214:259–68. https://doi.org/10.1016/j. approach based on near infrared spectroscopy: tablet hardness, content uniformity, foodchem.2016.06.100. and dissolution test measurements of intact tablets. J Pharm Sci 2006;95:2137–44. [50] Faber NM, Bro R. Standard error of prediction for multiway PLS: 1. Background https://doi.org/10.1002/jps.20653. and a simulation study. Chemom Intell Lab Syst 2002;61:133–49. https://doi.org/ 10.1016/S0169-7439(01)00204-0. 10