Direct Integration of Pervaporation as a Sample Preparation Method


Direct Integration of Pervaporation as a Sample Preparation Method...

0 downloads 74 Views 333KB Size

Anal. Chem. 2005, 77, 4927-4935

Direct Integration of Pervaporation as a Sample Preparation Method for a Dedicated “Electronic Nose” Carmen Pinheiro, Thomas Scha 1 fer,† and Joa˜o G. Crespo*

REQUIMTE-CQFB, Chemistry Department, Faculdade de Cieˆ ncias e Tecnologia, Universidade Nova de Lisboa, 2829-516 Caparica, Portugal

The present study investigates the possibility of monitoring the bioproduction of a complex aroma profile with an analytical electronic aroma-sensing technique, the socalled “electronic nose”, combined with a pervaporative sample enrichment method necessary to overcome the ethanol interference on the sensors’ response. It presents in detail the development of a direct integrated pervaporation-electronic nose unit for a simple and fast analysis, which are key criteria for this technique to be broadly implemented. The system developed was investigated using model solutions simulating the muscatel wine must fermentation. It proved to be able to evaluate different relevant aroma compounds in solutions of varying degree of complexity, and also in the presence of ethanol, which is a major interference on the sensors’ response to the aromas. The transient sensors’ response was investigated in detail, revealing information for sample discrimination and reducing the analysis time. The system developed allowed a simple, fast, and selective analysis, therefore permitting a high sample throughput over time, with the possibility of fully automation. Since the beginning of the 1980s, arrays of nonspecific chemical sensors associated with an appropriate pattern recognition technique have been developed, with the aim of emulating the human olfactory system in which semiselective olfactory receptors are combined with neural processing.1 These artificial olfactory systems, frequently found referred to as “electronic noses”-a somewhat ambitious term that nonetheless will be adopted in the following because it has become of common uses ideally produce for each complex aroma a unique response pattern, a so-called “fingerprint”, that withholds the aroma’s complexity. Electronic noses do not necessarily provide information on the quantity of the individual aroma compounds, but rather a global and qualitative analysis of the aroma profile, which in this sense resembles the human olfactory perception and which can be likewise valuable for aroma characterization. The information contained in the fingerprints is subsequently unfolded using pattern recognition techniques, such as, for example, principal * To whom correspondence should be addressed. Tel.: + 351 21 294 83 85. Fax: + 351 21 294 83 85. E-mail: [email protected]. † Present address: Department of Chemistry and Industrial Chemistry, University of Pisa, 56126 Pisa, Italy. (1) Gardner, J. W.; Bartlett, P. N. Sens. Actuators, B 1994, 18-19, 211-220. 10.1021/ac050139y CCC: $30.25 Published on Web 06/09/2005

© 2005 American Chemical Society

component analysis (PCA) or artificial neural networks, for discrimination or classification of the samples. Electronic noses are designed to permit a fast analysis of complex odors and to keep the time necessary to obtain and process the data short. Efforts are being made to have this equipment simple, fast, and low cost, so that it can be implemented as a routine, on-line, and real-time monitoring technique. Therefore, electronic noses have been receiving a tremendous interest in the analytical field as possible fast and low-cost alternatives for conventional analyses, but also for process monitoring and control. Doubtlessly the major field of application has been in the food industry, ranging from evaluation of raw materials,2 food processing monitoring,3 to quality control of end products,4 but also broadening to other areas, such as bioprocess monitoring,5 environmental monitoring,6 and medical diagnosis,7 among others. In practice, electronic noses have often proved to fail when assessing minor sample constituents in the presence of a major interference, such as ethanol or water (humidity). Examples are the ethanol interference observed in numerous electronic nose systems, based on different types of sensor array technologies, when analyzing alcoholic beverages.8-12 An attempt to overcome this strong limitation was tried by removing the respective interferences prior to sample analysis.9,12 However, removing bulk components involuntarily results in also removing part of the target solutes, and sample information is consequently lost for subsequent sensor evaluation. (2) Jonsson, A.; Winquist, F.; Schnu ¨ rer, J.; Sundgren, H.; Lu ¨ ndstrom, I. Int. J. Food Microbiol. 1997, 35, 187-193. (3) Zondervan, C.; Muresan, S.; de Jonge, H. G.; Thoden, van Velzen E. U.; Wilkinson, C.; Nijhuis, H. H.; Leguijt, T. J. Agric. Food Chem. 1999, 47, 4746-4749. (4) Aparicio, R.; Rocha, S. M.; Delgadillo, I.; Morales, M. T. J. Agric. Food Chem. 2000, 48, 853-860. (5) Bachinger, T.; Riese, U.; Eriksson, R. K.; Mandenius, C. F. Biosens. Bioelectron. 2002, 17, 395-403. (6) Fenner, R. A.; Stuetz, R. M. Water Environ. Resour. 1999, 71(3), 282-289. (7) Ping, W.; Yi, T.; Haibao, X.; Farong, S. Biosens. Bioelectron. 1997, 12, 10311036. (8) Pinheiro, C.; Rodrigues, C. M.; Scha¨fer, T.; Crespo, J. G. Biotechnol. Bioeng. 1000, 77 (6), 632-640. (9) Aishima, T. Anal. Chim. Acta 1991, 243, 293-300. (10) Privat, E.; Roussel, S.; Grenier, P.; Bellon-Maurel, V. T. Sci. Aliments 1998, 18, 459-470. (11) Guadarrama, A.; Ferna´ndez, J. A.; IÄ n ˜iguez, M.; Souto, J.; de Saja, J. A. Anal. Chim. Acta 2000, 411 (1-2), 193-200. (12) Muenchmeyer, W.; Walte, A.; Matz, G. Sens. Actuators, B 2000, 69, 397383.

Analytical Chemistry, Vol. 77, No. 15, August 1, 2005 4927

In our previous work, the approach of enriching the aroma compounds relative to ethanol was therefore chosen, employing a selective membrane separation processsorganophilic pervaporationsas a biocompatible and clean sample preparation method.8,13 In this process, a dense, nonporous, hydrophobic membrane separates the upstream side (feed) containing the feed solution from the downstream side (permeate) containing the compounds recovered. However, the hydrophobic membrane is not a mere barrier, but rather interacts with the solutes promoting or rendering more difficult their sorption and subsequent diffusion according to their mutual chemical nature; i.e., it acts as a selective transport barrier. More details regarding pervaporation as a sample preparation method can be found elsewhere.8,13 For the monitoring of a wine must fermentation, during which aroma compounds responsible for the complex organoleptic character are produced in the ppm range, while one of the main metabolic products, ethanol, is produced in much higher quantities (up to 10% w/w), this methodology proved to be feasible for discriminating samples based on the aroma compounds and hence to follow the fermentation based on information regarding the aroma profile evolution.8 In practice, however, this selective enrichment should not impair the main advantages of an electronic nose, namely, the simplicity of operation and the short analysis time, which are key criteria for this technique to be broadly implemented. According to this aim, the present work presents the development of a novel integrated pervaporation-electronic nose unit for simple, fast, and automated analysis. Although the direct integration is an elegant solution, it is not obvious that the on-line pervaporation is an adequate sampling procedure for the sensors, due to the vacuum applied, which results in a low molecular density of the sample obtained. Hence, the coupled system was thoroughly evaluated for the analysis of sample matrixes representing the wine must complexity. MATERIALS AND METHODS Reagents. Ethanol and the wine model aroma compounds used, ethyl acetate, isoamyl acetate, and isoamyl alcohol, were of analytical grade (Sigma-Aldrich). All standard solutions were prepared by weight in order to yield accurate concentrations. The wine model compounds were chosen either due to their relatively high concentration in the wine must, their organoleptic significance, or both.14 Electronic Nose. A commercially available electronic nose with an array of 25 organic conducting polymer-based sensors was employed. Conducting polymers have defined adsorptive surfaces that interact with volatile chemicals, which therefore display reversible changes on the electrical resistance of the polymers.15 The electronic nose was composed of an analyzer (A32S AromaScan, Osmetech Plc) that contained the array of sensors and a sample station (A8S, Osmetech Plc) used to (13) Pinheiro, C. Monitoring the Biological Production of Complex Aroma using an Electronic Aroma Sensing Technique. Ph.D. Thesis, Universidade Nova de Lisboa, Portugal, 2003. (14) Scha¨fer, T.; Bengtson, G.; Pingel, H.; Bo¨ddeker, K. W. Crespo, J. P. S. G. Biotechnol. Bioeng. 1999, 62, 412-421. (15) Persaud, K. C.; Pelosi, P. Sensor arrays using conducting polymers for an artificial nose. In Sensors and sensory systems for an electronic nose; Gardner, J. W., Bartlet, P. N., Eds.; NATO ASI Series E: Applied Sciences 212; Kluwer: Dordrecht, 1992; Chapter 15.

4928

Analytical Chemistry, Vol. 77, No. 15, August 1, 2005

humidify with precision an inert gas (N2, Air Liquide), then used as reference gas, and to equilibrate the samples at a defined temperature. Following the manufacturer’s advice, the humidity effect was minimized by matching the humidity of the reference gas to that of the sample. Characterization of the Sensors’ Response to Individual Compounds. The 25 sensors in the array were organized in groups of replicas of the same type of sensor, as follows: type I (sensors 1-3); type II (sensors 4-6); type III (sensors 7-10); type IV (sensors 11-14); type V (sensors 15-18); type VI (sensors 19-22); and type VII (sensors 23-25). Ten microliters of each pure compound was equilibrated during 1 h at 30 °C in a 22-mL closed vial. Each analysis consisted of a cycle of 2 min, 30 s, during which the sensors were exposed to different ambiences: 30 s to reference gas (N2 humidified to 50% relative humidity, established at 30 °C); 60 s to the sample headspace (transferred to the sensors by dry reference gas); and 60 s to reference gas for baseline recovery. During the period that the sensors were exposed to the sample, the flux was stopped over the sensors (“stopped flow technique”, as advised by the manufacturer) in order to obtain a steady-state response. three to five replicas of each sample were taken, and the period of 50-60 s of the steady-state sensors’ response was used for further processing. This procedure was used for characterization of the sensors’ response only. For the integrated system, the transient response was used, obtained with the system explained and operated as described in the following. Integrated System. The integrated system implemented in this work is depicted in Figure 1. The feed solution was placed in a temperature-controlled feed vessel (1). The feed solution was recirculated through the pervaporation module (2; GKSS Research Center) by means of a gear pump (3; Ismatec) and the feed flow measured by means of a flowmeter (F; Cole-Parmer). The upstream tubing was kept in PTFE (Omnifit). The membrane used was a modified silicone rubber-type composite membrane, poly(octylmethylsiloxane)-polyetherimide (POMS-PEI) (GKSS), with an active layer thickness of 10 µm (POMS) and an effective membrane area of 100 cm2. According to previous studies, this membrane had been found most suitable for recovering and selectively concentrating relevant aroma compounds that define the muscatel aroma profile.14 In the downstream compartment (indicated by the gray lines in Figure 1), the pressure was established by a rotary-vane vacuum pump (7; DUO 2.5, Pfeiffer Vacuum) and measured using a pressure gauge (P; Pirani, Pfeiffer Vaccum). To avoid sensor contamination due to back-diffusion of pump oil into the permeate duct, a glass cold trap (5) immersed in liquid nitrogen (Air Liquide) and a filter (6), composed of a mixture of molecular sieves (Merck) and activated carbon (Norit), were placed between the permeate recovery tubes and the pump. To recover a sample of permeate vapor and subsequently send it to the sensors, a set of “on-off” valves was used: two-way (B and D; Swagelok) and three-way (A; Pfeiffer Vacuum, and C; Swagelok); a set of tubes and a loop of defined volume (4) were also used. Tubes used to establish the vacuum in the permeate compartment, represented in Figure 1 as thick gray lines, were of aluminum, with the exception of the ones directly connected to the pervaporation module and pressure gauge, of stainless steel, as they were used to recover the permeate to be analyzed. All

Figure 1. Setup of the integrated system of on-line pervaporation with the electronic nose sensor array. T, temperature control; P, Pirani gauge; F, flowmeter. (1) feed vessel; (2) pervaporation module; (3) recirculation pump; (4) sampling loop; (5) condenser; (6) filter; (7) vacuum pump; (8) three-way valve of the analyzer; (9) electronic nose sensor array; (10) humidity and temperature sensors. (A-D) two- and three-way valves.

Figure 2. Sequence of the analysis for aqueous solutions using the integrated PV-EN system.

these tubes had an inner diameter of 17.2 mm (DN 16 ISO-KF; Pfeiffer Vacuum). Tubes and connections used to recover the permeate sample, including the loop, represented as thin gray lines, had an inner diameter of 2.1 mm (commonly designated as 1/ -in. tubes, Swagelok) and were of stainless steel. Valve A was 8 of stainless steel and valve C of brass. The tubes and connections used for recirculation of the reference gas, indicated by the thin black lines, were of brass, and the corresponding B and D valves of stainless steel. The total volume of the permeate recovery compartment was 200 mL, and the volume sent to the sensors of 3 mL, corresponding to the sampling loop’s volume. A three-way valve (8) inside the electronic nose analyzer allowed the entrance of a humidified reference gas, used as a baseline for the sensors, or the permeate sample, carried by dry pressurized carrier gas (N2, Air Liquide) at a flow rate of 10 mL/ min. The electronic nose sensor array (9) was located inside the analyzer, followed by a temperature and humidity sensor (10). System Operation. (a) Conditioning. Each solution to be analyzed was placed in the feed vessel (1) and conditioned during 5 min (recirculated at a flow rate of 1.6 L/min; pervaporation connected). The vacuum pump (7) was connected, and opening valve A to the left and valve C to the right allowed establishing the minimum pressure (maximum vacuum) in all the downstream compartment, including the loop (4), indicated by the gray lines. The minimum pressure observed with this configuration was 30 Pa, and experiments were run at a temperature of 22 ( 1 °C. (b) Sample Collection or Loop Fill Time. Valve C was closed, and valve A was turned to the right, allowing the accumulation of the vapor permeate on the downstream side of the membrane, including the loop, and consequently observing a pressure raise on the downstream compartment until a predefined

final pressure value (dependent on the sample). After this time, the permeate sample collection was terminated by turning valve A to the left. (c) Expansion and Stabilization. Valve B was opened during 4 s for expansion of the loop content until ambient pressure; after this period, the valve was closed again and the sample in the loop stabilized during 2 min to homogenize. (d) Sampling to the Sensors. Initially, the baselines of the sensors were established by passing humidified reference gas over them for 30 s (in this case, N2 humidified to 25% relative humidity, established at 30 °C). After this period, automatic actuation of valve 8 allowed the admission of the loop content, carried over the sensors by dry nitrogen gas at a flow rate of 10 mL/min, by opening valve B and valve C to the left. The transient response was recorded by continuously sending the sample over the sensors, until the sensors’ signal reached negative values. Finally, the initial sensors’ signal was recovered by passing humidified reference gas over the sensor again for 60 s, to ready the instrument for analysis of the next sample, already trapped in the sampling loop according to the procedure mentioned above. This procedure is illustrated in Figure 2, in particular for the analysis of an aqueous solution. The typical sensor array transient response obtained with this analysis cycle is depicted in Figure 3. Data Analysis. The software provided with the equipment (AromaScan) was utilized to acquire and process the data. For characterization of the sensors’ response to the individual compounds, sensor data were collected and averaged from the steady state of the sensor response, between 50 and 60 s. These data were then normalized relative to the total sensor response, to compensate for variations on vapor concentration due to Analytical Chemistry, Vol. 77, No. 15, August 1, 2005

4929

Table 1. Quality Factors for the Separation between an Aqueous Solution with 25 ppm Ethyl Acetate and Water, for the Different Intervals Represented in the Score Plot of Figure 5 sensor response interval (s)

quality factor

sensor response interval (s)

quality factor

40-45 45-50 50-55 55-60

30.64 64.75 47.05 33.27

60-65 65-70 70-75 75-80

20.22 15.11 4.50 2.01

Table 2. Quality Factors for the Separation between the Different Solutions Represented in the Score Plot of Figure 6

Figure 3. Transient response of the sensor array to an aqueous solution of 25 ppm ethyl acetate analyzed with the integrated PV-EN system.

experimental errors or deviations of sensors’ responses. This procedure allows eliminating the effect of concentration. The results were represented as radar plots (see Figure 4). For the results of the integrated system, sensor data were collected and averaged from different periods of the transient period of the sensor response (see Figure 3). The data for the different samples were then autoscaled (mean-centered and divided by their standard deviation, for rescaling with unit variance). Autoscaling is a common preprocessing method in PCA, employed to prevent high sensor responses from dominating the analysis and allowing evaluation of relevant information that may be contained in sensors exhibiting low responses. The data were visualized using PCA, a linear feature extraction method that reduces the dimensionality while preserving the maximum variance existent in the original data set. It consists of choosing a line along the axis of largest variation in the data: this first axis is called the first principal component, PC1, and accounts for the maximum variance. A second principal component, PC2, is drawn orthogonal to PC1 so that it accounts for the largest variation of the data points unexplained by PC1. PC3 is drawn orthogonal to PC1 and PC2 and accounts for the remaining variance; and so on, each new principal component is calculated by decreasing order of relevance, and accounts for less and less variance of the original data, and therefore less information. PCA has been widely used to analyze electronic nose responses, reducing the high dimensionality of a multivariate data set to usually two or three new variables (principal components) that contain the majority of the information, where the original data are projected: score plot. The principal components result from linear weighted combinations of the original variables. These weights are commonly named loadings, and they measure the correlation between the original variables (in the present case, the sensor responses) and the new variables (the principal components obtained with PCA), which can be easily visualized in a loading plot. More details regarding PCA can be obtained elsewhere.16 (16) Massart, D. L.; Vandeginste, B. G. M.; Buydens, L. M. C.; De Jong, S.; Lewi, P. J.; Smeyers-Verbeke, J. Data handling in science and technology 20A, Handbook of chemometrics and qualimetrics; Elsevier: Amsterdam, 1997.

4930 Analytical Chemistry, Vol. 77, No. 15, August 1, 2005

solutions discriminated

quality factor

water/10 ppm isoamyl acetate water/25 ppm ethyl acetate water/100 isoamyl alcohol 10 ppm isoamyl acetate/25 ppm ethyl acetate mixture 1a/10 ppm isoamyl acetate mixture 1a/25 ppm ethyl acetate mixture 1a/mixture 2b mixture 2b/100 ppm isoamyl alcohol

54.66 75.38 62.73 13.76 22.62 16.43 34.03 23.97

a Mxture 1/10 ppm isoamyl acetate + 25 ppm ethyl acetate. Mixture 2/10 ppm isoamyl acetate + 25 ppm ethyl acetate + 100 ppm isoamyl alcohol.

b

In the score plots defined by the selected principle components, a parameter designated as quality factor (QF) was obtained. This parameter is automatically given by the equipment, and it is based on the Mahalanobis distance, which is a measure of the class separation between different cluster centroids, and is presented in Tables 1 and 2 as an indication of the clusters/sample discrimination (see further discussion). The QF takes into account the degree of correlation (or covariance) between the variables, as well as the shape (circular or noncircular) and orientation. It originates an invariant distance between two clusters by pooling their respective within-group variance.17 More details regarding the Mahalanobis distance can be found elsewhere.18 A corollary generally accepted defines that if two clusters have group means that are within six standard deviations of one another, misclassification can occur due to overlap of the boundaries.17 RESULTS AND DISCUSSION Characterization of the Sensors’ Response to Individual Compounds. Parts a-e in Figure 4 depicts the typical patterns, in the form of radar plots, displayed by the array of sensors used to water, ethanol, ethyl acetate, isoamyl alcohol, and isoamyl acetate, respectively. Details regarding the analysis and data treatment in order to obtain these plots were presented in the Materials and Methods section. This kind of graphical representation is very useful as it allows one to easily visualize the typical response pattern produced by the array of sensors to a certain compound. It can be observed in general that for water and ethanol all sensors present an average response (defined by the manu(17) Mark, H. L. ; Tunnell, D. Anal. Chem. 1985, 57, 1449-1456. (18) De Maesschalck, R.; Jouan-Rimbaud, D.; Massart, D. L. Chemom. Intell. Lab. Syst. 2000, 50, 1-18.

Figure 4. Pattern of responses of the electronic nose sensor array to different individual compounds, represented as radar plots.

facturer as 2% < R < 5%). However, for the other compounds, there are sensors that stand out, presenting a strong response (defined as 5% < R < 9%), resulting in more characteristic patterns for these last compounds. For the aroma compounds studied, higher selectivity of sets of sensors were found as follows (in increasing order of the responses): ethyl acetate, sensor types II (4-6), VII (23-25), V (15-18), and IV (11-14); isoamyl alcohol, sensor types II (4-6), VII (23-25), V (15-18), and IV (11-); isoamyl acetate, sensor types II (4-6), VII (23-25), IV (11-14), and V (15-18). As can be easily observed from these plots, sensors of the same type respond in general in a very similar way, which would be expected as they correspond to replicas of the same sensor. In electronic noses, it is a common practice to use replicas of the same type of sensor in the overall sensor array. This procedure allows obtaining multiple responses of the same type of sensor in only one analysis, contrary to increasing the number of analyses/ replicas of the same sample. Furthermore, it permits one to evaluate the short-term and long-term reproducibility of the same type of sensor. Nevertheless, not always sensors of the same type present identical responses. As an example, in Figure 4, it can be noticed that sensor 13, belonging to type IV (11-14), presents in general a different response from the rest of the set. It was observed that this sensor presented an initial base resistance signal lower than the other sensors, which might explain in general its different behavior, and possibly these sensor data should not be processed by the pattern recognition techniques. Integration of “On-Line” Pervaporation as a Sampling Method for the Electronic Nose. In a previous work, the potential of integrating pervaporation as a sample preparation method for the electronic nose was demonstrated. In this case, the selective sample enrichment was performed in a standard laboratory-scale pervaporation unit and the enriched sample (permeate) quantitatively recovered by condensation.8 Nevertheless, this method was implemented “off-line”, imposing serious disadvantages to the analysis. A detailed examination of the procedure indicates that the permeate was quantitatively recovered in the liquid phase by condensation for 6 h, it was subsequently diluted to a defined ethanol concentration, and finally, the sample

was equilibrated for 1 h at 60 °C in order to obtain a vapor headspace that was then transferred to the sensors. Consequently, the sample pretreatment using off-line pervaporation turned the analysis by an electronic nose extremely laborious, slow, and difficult to automate. It was therefore necessary to establish a direct integration of this sampling method by pervaporation with an electronic nose, designated by on-line pervaporation, that allowed a fast and simple analysis. In pervaporation, vacuum is usually applied in the downstream side of the membrane in order to diminish the partial pressures of the compounds and, by this manner, establish and maximize the driving force for solute transport across the membrane. According to Dalton’s law, as the total downstream pressure is low, the partial pressures of the compounds are low and below their correspondent saturated vapor pressure. Consequently, in pervaporation, the solutes that cross the membrane to the downstream compartment are recovered already in the vapor phase, such that the enriched sample can then be directly conducted over the sensors. This allows avoiding the steps of condensation and generation of the liquid-vapor equilibrium, making this analysis much faster and simpler. During this work, different systems were designed and implemented that coupled pervaporation directly with the sensors of the electronic nose. The results presented in this paper are only focused on the last system developed, the outcome of a number of optimization procedures aimed at diminishing the length of the sample transfer lines, having the system simple and easy to operate, and foreseeing the possibility of full automation. In this system, a sample of vapor permeate was obtained, subsequently expanded, and sent to the sensors at atmospheric pressure. Furthermore, the generated downstream pressure was minimized (vacuum maximized), maximizing the driving force across the membrane in order to operate pervaporation under the best conditions. The configuration of the system developed is depicted in Figure 1. The operation of this system (presented in detail in the Material and Methods section) was optimized in order to minimize the time required for each analysis. This time length was the result of all time phases included in the sampling with the on-line Analytical Chemistry, Vol. 77, No. 15, August 1, 2005

4931

Figure 5. Score plot of PC1 and PC2 presenting the discrimination obtained for the different periods from 40 to 80 s of the transient response, for an aqueous solution of 25 ppm ethyl acetate (black symbols) and water (white symbols).

pervaporation system and sending this sample to the sensors, which are presented in detail in Figure 2. Nevertheless, after the first sample, the analysis time required for the following replicas could be shortened by using the time ranging from sample expansion up to the analysis by the sensors for the conditioning of the next sample. In the case of aqueous solutions, these simultaneous procedures allowed reducing the analysis time from 15 min for the first sample to 10 min to the following samples (see Figure 2). This time allows a high sample throughput over time. The total volume of the permeate recovery compartment was 200 mL (including the 3-mL loop volume). Most of the volume corresponded to the downstream compartment of the laboratorial pervaporation module. This volume can be easily reduced (requiring the design and manufacture of a new pervaporation module), which would have the advantage of further reducing the sample collection time and, consequently, the total analysis time. This system was first evaluated concerning its capability of detecting ethyl acetate in water. The response to water and to an aqueous solution with 25 ppm ethyl acetate was compared, using a time of permeate accumulation (sample collection or loop fill time) of 4 min. Figure 3 depicts the transient response for an aqueous solution with 25 ppm ethyl acetate. The aim was to evaluate whether the sensors’ response curves had information for discrimination due to the presence of ethyl acetate and to identify the region with more relevant and consistent information for discrimination. For this purpose, the overall sensor response to the sample (40-110 s) was subdivided in periods of 5 s and the averaged value of each of these intervals used for further processing. Figure 5 depicts the score plot of PC1 and PC2 for the first eight intervals of the transient response to water and to an aqueous solution with 25 ppm ethyl acetate, ranging from 40 to 80 s. Table 1 presents the corresponding quality factors, indicating the separation between clusters with and without ethyl acetate. From Figure 5 and Table 1 it can be concluded that the intervals that originate a higher separation between clusters and, consequently, a better and more consistent discrimination are between 45 and 4932

Analytical Chemistry, Vol. 77, No. 15, August 1, 2005

50, 50 and 55, and 55 and 60 s. Although containing information for discrimination, the beginning of the response, the interval of 40-45 s, presents a great dispersion within the clusters, which deteriorates the separation and consequently diminishes the corresponding quality factor value. This dispersion might have been due to a higher variability of the response caused by the actuation of the valve during sample introduction (see Figure 3). From 60 s on, the separation between samples with and without ethyl acetate diminishes, and, for example, the clusters of the last two intervals (70-75 and 75-80 s) can no longer be discriminated, as the quality factors are lower than 6. The Mahalanobis distance can be computed in the original variable space or in a lower dimension space defined by the principal components after PCA.18 In the case of the software used, the quality factor is calculated in the PC space, meaning they are not absolute values, but rather change when a sample or group of samples is introduced or excluded, redesigning the PCA and consequently the score plot. Nevertheless, it was verified that the variations in this parameter were not very high (remained in the same order of magnitude), and therefore, it can be considered a good indicator of separation between clusters. In Figure 5, it can be observed that the discrimination due to ethyl acetate occurs along PC2, while PC1 differentiates the response intervals used. The results from 80-110 s are not shown, but they present a high dispersion of the clusters formed, namely, for water, and hence should not be used for discrimination. This system was further evaluated, analyzing the following aqueous solutions: water and binary aqueous solutions with 10 ppm isoamyl acetate and 100 ppm isoamyl alcohol. Moreover, a ternary aqueous solution with 10 ppm isoamyl acetate plus 25 ppm ethyl acetate, and a quaternary aqueous solution with 10 ppm isoamyl acetate, 25 ppm ethyl acetate, and 100 ppm isoamyl alcohol were also analyzed. The concentrations of the different aroma compounds were chosen to simulate the typical concentrations toward the end of the muscatel wine must fermentation. The concentration of isoamyl acetate was higher than normally found but corresponded to the total concentration of esters with high organoleptic impact (commonly designated fruity esters).

Figure 6. Discrimination of one or more different aroma compounds dissolved in water obtained with the integrated PV-EN system.

Figure 7. Score plot of PC1 and PC2 presenting the separation of a solution of 100 ppm isoamyl alcohol and water.

Figure 6 depicts the score plot of PC1 and PC2 for all these solutions, using the period of response 45-50 s, because this interval proved once more to originate the best discrimination for all the solutions analyzed. Table 2 presents the quality factors for the discrimination of these solutions, where it can be observed that all the values obtained are high. Therefore, it can be concluded that the integrated system established allows evaluating the contribution of the different aroma compounds in water, independently and in more complex mixtures. As an example, Figure 7 presents the score plot of PC1 and PC2 for the discrimination between water and an aqueous solution of 100 ppm isoamyl alcohol. Figure 8 presents the corresponding loading plot, where the sensors more relevant for the discrimination are indicated. It can be concluded that using the integrated system and after an adequate selection of sensor information it is possible to obtain a very good discrimination between all solutions with the different aroma compounds, indicated by a quality factor of 75.78 for the discrimination between water and the 100 ppm isoamyl alcohol aqueous solution (as depicted in Figure 7): 65.50 for water-25 ppm ethyl acetate; and 55.06 for water-10 ppm isoamyl alcohol (these last two, not presented). As can be observed in Figure 7, sample discrimination occurs along PC1 (that represents 99.24% of the total information), while PC2 accounts for cluster dispersion (with only 0.55% of the total information). The corresponding loading plot (Figure 8) allows one to obtain information about the sensors, as the loadings

Figure 8. Loading plot for the separation of a solution of 100 ppm isoamyl alcohol and water.

measure the correlations between the original variables (in this case, the 25 sensor responses) and the new variables (the principal components). The nearer the loading value of 1 (positive or negative), the higher the association of the variable with the component; on the contrary, values near zero indicate that the variable contributes little to the formation of that component. In Figure 8, the variables/sensors that contribute most to the discrimination of isoamyl alcohol from water are the ones with the highest negative weights/loadings along PC1 (sensors 1518 and 11-14) and with the highest positive loadings (19-22 and 1-3). According to Figure 4c, the first set of sensors are the sensors presenting a stronger response to isoamyl alcohol, while the second set of sensors presents almost no response to this compound; this behavior corroborates that the discrimination observed is in fact due to isoamyl alcohol. As it is known that in this kind of graphical representation two variables are strongly correlated when there is a small angle between the lines connecting them with the origin,16 it is possible to identify in Figure 8 the variables/sensors that are correlated, namely, sensors 4-6, 11-14, 15-18, and 23-25. This result was expected as these sets consist of replicas of the same type of sensor. The results for the discrimination of water from an aqueous solution of 25 ppm ethyl acetate (data not shown) indicate sensors (in increasing order of importance) 15-18 and 11-14, and for the discrimination of water from an aqueous solution of 10 ppm isoamyl acetate (data not shown), sensors 11-18, as the ones Analytical Chemistry, Vol. 77, No. 15, August 1, 2005

4933

Figure 9. Score plot of PC1 and PC2 presenting the discrimination of a mixture of aroma compounds (10 ppm isoamyl acetate, 25 ppm ethyl acetate, and 100 ppm isoamyl alcohol) in the presence of 10 wt % ethanol.

more relevant for discrimination based of the respective aroma. In both cases, these sets of sensors are the ones presenting stronger responses to ethyl acetate and isoamyl acetate (see Figure 4d and e), which again corroborates that the separations are in fact due to ethyl acetate and isoamyl acetate, respectively. Moreover, in both cases. sensors 4-6 and 23-25 presented no significant contribution for the separation (weights on PC1 near zero). According to Figure 4d and e, these last sets of sensors, belonging to sensor types II and VII, were also indicated as responding to these aroma compounds, although exhibiting lower responses than sensors of types IV and V. The fact that these sensors do not contribute significantly to the discrimination might be due to the concentration reaching the sensors being inferior to their response threshold. The knowledge of the sensors’ response threshold and resolution capabilities is therefore of major importance and may be evaluated by determining the exact concentration of the vapor that effectively reaches the sensors. Finally, the integrated system developed was evaluated concerning its capability to analyze a mixture of aroma compounds in the presence of interfering compounds, namely, ethanol. Simulating the muscatel wine must, toward the end of the fermentation, a solution with 10 ppm isoamyl acetate, 25 ppm ethyl acetate, and 100 ppm isoamyl acetate in 10% w/w ethanol was analyzed against a solution of 10% w/w ethanol only. In this case, the analysis time by the sensors (sampling to the sensors, see Figure 2) increased by 2 min or more relative to the analysis of aqueous solutions, as the transient response required longer time to recover the baseline. This caused the total time necessary for each analysis to increase to 17 min for the first sample and 12 min for the following replicas. Figure 9 depicts the score plot of PC1 and PC2, showing the discrimination between these two solutions (with a quality factor of 14.67), using the data from interval 50-55 as this demonstrated to originate a better discrimination. Figure 10 presents the corresponding loading plot, where it can be observed that the sensors contributing to the discrimination due to the aroma compounds (with higher negative weights/ loadings along PC1) are sensor types V (15-18) and IV (11-14). Sensors 23-25 present values near zero, and four, five, and six positive loadings and, therefore, do not contribute for the aroma 4934 Analytical Chemistry, Vol. 77, No. 15, August 1, 2005

Figure 10. Loading plot for the separation of a mixture of aroma compounds in the presence of 10 wt % ethanol.

evaluation. Sensors of types IV and V are in fact the ones that in general respond and contribute to the perception of the aroma compounds considered, which allows concluding that this set of results has a real significance. It should be reminded that without coupling pervaporation as an aroma pre-enrichment sampling method, the sensors’ response to the aroma compounds would be completely overlapped by the ethanol interference.8 Indeed, the sampling is a very important issue, and it should be carefully evaluated and improved in order to improve the overall electronic nose performance.19 From the previous results, it can be concluded that the information for discrimination takes place in the transient phase of the response, namely, in the beginning of the response rise. As a matter of fact, although these data were not shown, it was observed during this work that if the sample was stopped over the sensors in order to create a steady-state response (commonly known by “stopped flow technique”, as presented in the Materials and Methods section), information for discrimination occurred in the initial transient part, while in the steady state there was no discrimination of the samples at all. Although it is a common practice in electronic noses, namely, in commercial systems, to use information in the steady-state region, it is becoming more and more usual to evaluate in detail the information also contained in the transient region and often use it in order to improve sample discrimination.20, 21 In fact, it is well known for gas sensors that their transient responses can provide valuable information of the sample chemical composition, indicating the importance of kinetic effects on the sensor’s surface.22 Another major advantage of using the transient response is that, by avoiding extending the analysis in order to obtain a steady-state response, the total cycle of the analysis is much faster. Nevertheless, the dynamic response of the sensors is strongly affected by how the sample is sent over the sensors, as it is very sensible to variations, such as, for example, flow fluctuations. (19) Mielle, P.; Marquis, F. Sens. Actuators, B 1999, 58, 526-535. (20) Wilson, D. M.; DeWeerth, S. P. Sens. Actuators, B 1995, 28, 123-128. (21) Eklo ¨v, T.; Mårtensson, P.; Lundstro ¨m, I. Anal. Chim. Acta 1997, 353, 291300. (22) Fort, A.; Gregorkiewitz, M.; Machetti, N.; Rocchi, S.; Serrano, B.; Tondi, L.; Ulivieri, N.; Vignoli, V.; Faglia, G.; Comini, E. Thin Solid Films 2002, 2-8.

CONCLUSIONS In this work, a system was developed that integrated directly a pervaporative sampling technique with the sensors of the electronic nose (PV-EN system). This system was evaluated using model solutions, simulating samples of the muscatel wine must fermentation with different degrees of complexity. It was demonstrated that this system allowed evaluation of the contribution of different relevant aroma compounds in more and less complex solutions, and also in the presence of ethanol, which is a strong interfering compound during the sensors’ response to aromas. Additionally, the analysis was simple and fast, allowing a high sample throughput over time. It was observed that the transient response contained significant and consistent information for discrimination, allowing simultaneously a short analysis cycle. Nevertheless, when using the dynamic information and time-dependent data extraction methods (as was the case in this system), it is extremely important to have very good control of the fluxes and the sample introduction time. Therefore, it is recommended that this system be totally automated and that there is a precise measure and control of the fluxes involved. It is of extreme interest to broaden this study with a quantitative analysis of the integrated system implemented, investigating in detail not only the performance of pervaporation as a sampling

method but also obtaining information regarding the sensors response threshold and resolution capability. The information obtained may then be used to optimize the integrated system for a more dedicated aim, if desired, such as taking advantage of the system versatility, improving the selectivity toward key aroma compounds. ACKNOWLEDGMENT The authors acknowledge GKSS, Germany, for the kind donation of the POMS-PEI membranes. We are grateful to the Department of Information Engineering and Department of Earth Sciences, University of Siena, Italy, for valuable ongoing collaboration and comments. We also thank Osmetech plc., United Kingdom, in particular Dr. Andrew Tummon for his assistance. Finally, we gratefully acknowledge Prof. Joaquim Vital, from Faculdade de Cieˆncias e Tecnologia, Universidade Nova de Lisboa, for his tireless assistance. C.P. acknowledges research grant PRAXIS XXI/BD/18289/9898, and T.S. grant SFRH/BPD/7172/ 2001, both from Fundac¸ ˜ao para a Cieˆncia e a Tecnologia, Portugal.

Received for review January 24, 2005. Accepted May 13, 2005. AC050139Y

Analytical Chemistry, Vol. 77, No. 15, August 1, 2005

4935