Anal. Chem. 1999, 71, 1338-1341
Detection of Volatile Compounds with Mass-Sensitive Sensor Arrays in the Presence of Variable Ambient Humidity F. L. Dickert,* O. Hayden, and M. E. Zenkel
Institute for Analytical Chemistry, University of Vienna, Wa¨ hringerstrasse 38, A-1090 Vienna, Austria
Mass-sensitive sensor arrays were established for the detection of isomeric or highly analogue analyte mixtures, which show similar physical and morphological properties. Supramolecular host-guest chemistry and arrays of four mass-sensitive quartz crystal microbalances have been successfully combined with multivariate calibration techniques in the presence of variable air moisture. This system enabled even the separation of xylene isomerssa task that might be crucial even by gas chromatography. The data of the sensor arrays were analyzed with partial least squares and artificial neural networks. The xylene isomers could be detected with an accuracy of ∼1% in the range of 0-200 ppm, nearly eliminating the residual water cross-sensitivity of the sensor coatings, which allows effective work place or environmental monitoring of toxic compounds with fast response levels. Due to the growing interest in environmental and work place monitoring, the detection of analytes such as aromatic and halogenated hydrocarbons is of increasing economical importance. Supplementary to instrumental analytical methods, smart sensor systems are desirable to enable quick and easy on-the-spot investigations. The feasibility of continuous surveillance would promote a higher security level in companies handling toxic chemicals. Sensor arrays have already been used for the quantitative recognition of analyte mixtures.1-4 The methods of multivariate data analysis5 are widely used, for example, to recognize several liquors6 or for drift correction purposes.7 Further applications are the detection of hazardous compounds8-10, odorants,11-14 or alcohols.15 (1) Weimar, U.; Schierbaum, K. D.; Go¨pel, W.; Kowalski, R. Sens. Actuators 1990, B1, 93-96. (2) Ema, K.; Yokoyama, M.; Nakamoto, T.; Moriizumi, T. Sens. Actuators 1989, B18, 291-296. (3) Pearce, T. C.; Gardner, J. W.; Friel, S.; Bartlett, P. N.; Blair, N. Analyst 1993, 118, 371-377. (4) Janata, J. ; et al. Anal. Chem. 1988, 70, 473-481 (5) Lavine, B. K. Anal. Chem. 1998, 70, 209R-228R. (6) Aishima, T. Anal. Chim. Acta 1991, 243, 293-300. (7) Smits, J. M.; Melssen, W. J.; Derksen, M. W. J.; Kateman, G. Anal. Chim. Acta 1993, 248, 91-105. (8) Mu ¨ ller, R.; Lange, E. Sens. Actuators 1986, 9, 39-48. (9) Carey, W. P.; Kowalski, B. R. Anal. Chem. 1986, 58, 3077-3084. (10) Stetter, J. R.; Jurs, P. C.; Rose, S. L. Anal. Chem. 1986, 58, 860-866. (11) Chang, S.-M.; Iwasaki, Y.; Suzuki, M.; Tamiya, E.; Karube, I.; Muramatsu, H. Anal. Chim. Acta 1991, 249, 323-329. (12) Kalman, E. L.; Winquist, F.; Lundstroem, I. Atmos. Environ. 1997, 31, 17151719.
1338 Analytical Chemistry, Vol. 71, No. 7, April 1, 1999
The chemical industry extensively uses aromatic solvents such as xylene isomers, toluene, and halogenated solvents such as tetrachloroethylene. The detection of these VOCs is not trivial because of the low functionality of these analytes. Their chemical and physical differences are very small; for instance, the boiling points of m- and p-xylene differ by less than 1 °C. Even with methods with a high separation efficiency, such as gas chromatography, the resolution of xylene isomers can be difficult. Highly sensitive and selective coating materials for mass-sensitive devices that can resolve closely related isomers would be desirable. Masssensitive quartz crystal microbalances (QCMs) operating at 10 MHz and coated with supramolecular host-guest materials, which are highly adaptable to the desired analytes,16,17 offer the selectivity required. For that purpose, the host-guest materials have been chosen from among β-cyclodextrins and calix[4]resorcinearenes. Different pairs of the analytes mentioned have been examined in the presence of varied relative humidity. Humidity has been included in the study due to the fact that, despite the low moisture absorption of the supramolecular compounds, the incorporation of humidity into the sensor materials can induce a change of polarity and lead to nonlinearities in the sensor characteristics. Therefore, one sensor has been coated with a material containing a high degree of hydroxyl groups serving as a humidity sensor. The data records of these sensor arrays have been analyzed by partial least squares (PLS) and artificial neural networks (ANNs.)18-22 EXPERIMENTAL SECTION Measurements. The sensor array consists of four 10-MHz quartz microbalances, whose frequencies are measured with Hewlett-Packard 53131A frequency counters and transferred to a (13) Hanaki, S.; Nakamoto, T.; Moriizumi, T. Sens. Actuators 1996, A57, 6571. (14) Lavigne, J. J. J. Am. Chem. Soc. 1998, 120, 6429-6430. (15) Shurmer, H. V.; Gardner, J. W.; Corcoran, P. Sens. Actuators 1990, B1, 256-260. (16) Dickert, F. L.; Haunschild, A. Adv. Mater. 1993 5, 887-893. (17) Hauptmann, P.; Ro¨sler, S.; Lucklum, R.; Borngra¨ber, R.; Hartmann, J. Sens. Actuators 1998, B48, 415-424. (18) Geladi, P.; Kowalski, B. R. Anal. Chim. Acta 1986, 185, 1-17. (19) Dickert, F. L.; Haunschild, A.; Hofmann, P. Fresenius J. Anal. Chem. 1994, 350, 577-581. (20) Sommer, V.; Tobias, P.; Kohl, D.; Sundgren, H.; Lundstroem, I. Sens. Actuators 1995, B28, 217-236. (21) Hierlemann, A.; et al. In Sensors Update; Baltes, H. P., Goepel, W., Hesse, J., Eds.; VCH: Weinheim, Germany 1996; Vol. 2, pp 119-180. (22) Carey, W. P.; Yee, S. S. Sens. Actuators 1992, B9, 113-122. 10.1021/ac981014e CCC: $18.00
© 1999 American Chemical Society Published on Web 02/24/1999
Figure 1. C8-CDMe: schematic structure of β-cyclodextrin crosslinked with diiodooctane. epi-CDnMe instead has been cross-linked with epichlorohydrin.
PC via an IEC bus. The gas mixing apparatus has also been controlled and fully automated through the PC. The evaluation of the array data is very sensible to frequency instabilities of the QCM devices; thus the oscillating circuit as well as the test gas stream has been temperature controlled to minimize drift phenomena. A frequency stability of ∼0.1 Hz/day could be achieved. The concentrations of the organic analytes have been varied in steps of 40 ppm between 0 and 200 ppm, and the relative humidity of the gas stream has been adjusted in steps of 15% between 0% and 60% relative humidity. In this way, the concentration space was spanned via 180 different compositions of the gas phase. Multivariate Calibration Procedures. The calibration was performed with Matlab for Windows using the Chemometrics Toolbox and the Neural Network Toolbox. In the first step, the 180 data points measured as described above were randomly split into 150 points for calibration and 30 points for cross-validation. For the evaluation with PLS, the measured data Si are used to calculate the analyte concentrations cj with linear polynomials of varying degree: 4
cj ) β 0 +
∑ i)1
4
β′iSi +
∑ i)1
4
β′′iS2i + ... +
∑β
n′ n i Si
i)1
The optimal results with minimal prediction errors have been obtained for polynomials of the third degree. The evaluation with ANNs resulted from standard backpropagation networks with sigmoidal neurons in the hidden layer and linear neurons in the output layer. Predicting all concentrations with only one network requires at least three neurons in the hidden layer, and many networks have to be trained due to the initialization of the latter with random numbers. Another improvement could be obtained, when the training of the network was stopped at the point of optimal validation (cf. Figure 8). Coating Materials. All measurements were performed with the same coated QCM devices. The layers of the four QCMs used in the array consisted of the supramolecular host-guest materials CAVS,23 VAS4S,24 C8-CDnMe,25 and epi-CDMe, with the chemical structures shown in Figures 1 and 2. The coatings showed a frequency shift of 4 kHz, which corresponds to a layer thickness (23) Dickert, F. L.; Ba¨umler, U. P. A.; Stathopulos, H. Anal. Chem. 1997, 69, 1000-1005. (24) Dickert, F. L.; Ba¨umler, U. P. A.; Zwissler, G. K. Synth. Met. 1993, 61, 47-52. (25) Dickert, F. L.; Geiger, U.; Keppler, M.; Reif, M.; Bulst, W.-E.; Knauer, U.; Fischerauer, G. Sens. Actuators 1995, B26-27, 199-202.
Figure 2. Calix[4]resorcinearenes CAVS and VAS4S.
on average of 160 nm. CAVS was prepared by refluxing resorcinol with decanal in ethanol and concentrated HCl. The substance was recrystallized twice in acetone.26 VAS4S was synthesized in dried DMSO with CAVS and 2,3-dichloroquinoxaline and recrystallized in ethyl acetate-chloroform (9:1).27 epi-CDnMe was prepared via cross-linking β-cyclodextrin with epichlorohydrin in a sodium hydroxide solution, followed by chromatographical purification and a subsequent methylation step28 in absolute DMF. C8-CDMe is a cross-linked β-cyclodextrin with 1,8-diiodooctane in DMF and was again methylated in analogy to epi-CDnMe. Thus, the polymerized β-cyclodextrin derivatives are polymerized host molecules unlike the calix[4]resorcinarene products, which are monomeric with linked aliphatic spacers. Every product has been characterized with NMR. RESULTS AND DISCUSSION (1) Resolution of m- and p-Xylene. The limitations of the procedure’s capacity have been tried with m- and p-xylene, which show no significant difference in their chemical and physical properties. The boiling point of m-xylene is 139.1 °C with a calculated dipole moment of 0.26 D. The nonpolar p-xylene boils slightly lower at 138.3 °C. Both analytes are therefore primarily detectable due to a selective inclusion process in the host molecule, whereby the minor inherent cross-sensitivities of the sensor coatings to the humidity may adversely affect the sensitivity pattern of the sensor array. The sensitivities of the array to both isomers including relative humidity are displayed in Figure 3. Sensor CAVS shows identical (26) Tunstad, L. M.; Tucker, J. A.; Dalcanale, E.; Weiser, J.; Bryant, J. A.; Sherman, J. C.; Helgeson, R. C.; Knobler, C. B.; Cram, D. J. J. Org. Chem. 1989, 54, 1305-1312. (27) Moran, J. R.; Ericson, J. L.; Dalcanale, E.; Bryant, J. A.; Knobler, C. B.; Cram, D. J. J. Am. Chem. Soc. 1991, 113, 5712-5714. (28) Boger, J.; Cornoran, R. J.; Lehn, J. R. Helv. Chim. Acta 1978, 61, 2190.
Analytical Chemistry, Vol. 71, No. 7, April 1, 1999
1339
Figure 3. Sensitivity pattern of the array with 10% RH and 1000 ppm p- and m-xylene. Table 1. Root-Mean-Squared Errors of Prediction of Relative Humidity and VOCs with PLS and ANNa measurement
analyte
PLS (%)
ANN (%)
1
RH m-xylene p-xylene RH m-xylene p-xylene RH toluene tetrachloroethylene
0.9 2.7 2.5 0.6 2.5 1.4 0.5 3.1 1.0
0.2 1.0 1.0 0.5 1.8 1.2 0.5 2.5 1.0
2 3
Figure 4. Comparison of predictions obtained with PLS and ANN, respectively. The analyte concentrations are collected to a vector and normalized to unity.
a For the analysis with PLS, polynomials of the third degree were used. The architecture of the ANN is explained in the text.
sensitivities to the xylene isomers with the predicted greatest sensitivity to humidity of the whole array. Sensor C8-CDnMe shows a much greater response to p-xylene than to m-xylene, whereas this selectivity is inverted for sensor epi-CDMe. Sensor VAS4S supplies surplus information for the pattern recognition, so that this array should enable the detection of m- and p-xylene in the ppm range. The resulting root-mean-squared errors of prediction (rmsep) of the analysis with the PLS procedure are summarized in Table 1. With ANNs, a more precise prediction by a factor of 2 is possible than with PLS (Table 1). Relative humidity (RH) can be predicted with an accuracy of 0.1% RH and the xylene isomers with an accuracy of 2 ppm. The results of PLS and ANN are compared in Figure 4. (2) Resolution of o- and m-Xylene. The excellent result for the detection of m- and p-xylene is made possible by the structural difference of the analytes. The resolution of highly similar o- and m-xylene is feasible by the more distinct difference in the boiling points of 144.4 °C for o-xylene and 139.1 °C for m-xylene in relation to the isomer mixture (section 1); thus a superior sensitivity for o-xylene can be expected. The results are shown in Figure 5. In contrast to the first array measurement, all sensors showed maximum sensitivity for oxylene and no inversion of the sensitivity seen with m- and p-xylene occurred. Therefore, the prediction is expected to be worse than in the first case. The procedure of evaluation with PLS and ANN including the formation of a calibration and validation set is the same as with the first array measurement with the errors of 1340 Analytical Chemistry, Vol. 71, No. 7, April 1, 1999
Figure 5. Sensitivity pattern of the array with 10% RH and 1000 ppm o- or m-xylene.
Figure 6. Sensitivity pattern of the array with 10% RH and 1000 ppm tetrachloroethylene or toluene.
prediction shown in Table 1. The PLS model allows on average a comparably accurate prediction to ANN. (3) Resolution of Toluene and Tetrachloroethylene. In addition to the examination of aromatic analytes, a mixture of an aromatic and a halogenated hydrocarbon has been investigated. For this purpose, toluene and tetrachloroethylene were selected as model compounds because they both have a flat lengthened shape and boiling points differing by only 10 °C. Furthermore, the sensitivities of mass-sensitive sensors are always greater for tetrachloroethylene than for toluene, due to the greater molecular weight of tetrachloroethylene.
the sensor responses still showed a strong relationship to the analyte level with variable humidity. Anyhow, the highest accuracy of prediction is obtained using ANN. In Figure 8, the training process of the ANN for the detection of humidity in the presence of toluene and tetrachloroethylene using a random validation control is shown. In contrast to the calibration curve, the prediction of the validation data shows a global minimum (enlarged figure) after 200 training cycles. This quite common shape of the validation curve is largely the consequence of an overfitting of noise artifacts.29
Figure 7. Comparison of the linear (×) and quadratic (O) PLS predictions for water concentrations using four sensors. Regression line of the quadratic PLS model (solid); the quadratic PLS predictions have a linear behavior with a rmsep of 0.5%; instead the linear PLS model shows nonlinearity for variable humidity with a rmsep of 1,6%. The best fit is obtained with cubic PLS or ANN.
Figure 8. Optimization of the network training for the analyte mixture (section 3) with a cross-validation procedure. A network of maximum predictive power is obtained at the minimum in standard error of prediction of validation (inset). Subsequent training results in overfitting.
The minor sensitivities for toluene in relation to tetrachloroethylene can be seen in Figure 6; the slightly worse recognition of toluene compared to tetrachloroethylene occurred as expected. The results found after evaluation are in the manner discussed above in section 2sagain all sensors showed similar selectivities for the organic vapors. The results for toluene differ by a significant amount, and the accuracy of the toluene prediction is worse compared to the other analytes (Table 1). As an example, Figure 7 visualizes the potential of eliminating the water cross-sensitivity with linear PLS. Linear models with higher polynomial orders yielded fairly good predictions, since (29) E. V. Thomas, Anal. Chem. 1994, 66, 795A-804A.
CONCLUSION The discussed mass-sensitive sensor array enables the highly accurate separation of mixtures of aromatic and halogenated hydrocarbons within the concentration range of 0-200 ppm even in variable humidity environments. The separation of m- and p-xylene, which was likely the most crucial task, was successfully obtained. Moreover, the m- and p-xylene mixture had the minimum rmsep of all separations presented. This result points out the effectiveness of supramolecular host molecules as coating materials for QCM resonators. Various sensitivity patterns are accessible using differently adapted host molecules. The analysis of the multidimensional array data with PLS works best using polynomials of the third degree to describe the functional relationship between sensor responses and analyte concentrations. The results gathered with linear PLS and ANN are of comparable accuracy, which emphasizes the minor humidity influence to the sensor response within the observed concentration range. The minimum rmsep were yielded with back-propagation networks with three sigmoidal neurons in the hidden layer, resulting in an accuracy better than 3 ppm concerning most organic analytes. Relative humidity can be predicted with an accuracy of 0.2-0.5% rmsep. Summarizing, mass-sensitive devices coated with supramolecular layers and multivariate data analysis provide the separation even of isomeric analytes. Furthermore, the low-priced sensor array with mass-sensitive devices is very appropriate for the gas phase, in particular for applications under harsh conditions, since chemical sensors based on resonators such as QCMs or SAWs have pronounced chemical and physical stability; this makes the continuous monitoring of the environment or of work places possible. Our results support future applications of mass-sensitive devices combined with computational chemistry due to the increasing commercial interest for sensor elements for multicomponent analysis. ACKNOWLEDGMENT Financial support by the “Jubila¨umsfonds-O ¨ stereichische Nationalbank” Project 5943 is gratefully acknowledged.
Received for review September 14, 1998. Accepted January 7, 1999. AC981014E
Analytical Chemistry, Vol. 71, No. 7, April 1, 1999
1341
