Gold Nanoparticle Chemiresistor Sensor Array ... - ACS Publications

simple system of linear alcohols (methanol, ethanol, propan-1-ol, and butan-1-ol) dissolved in .... Marc D. Woodka and Vincent P. Schnee , Michael...
1 downloads 0 Views 1MB Size
Anal. Chem. 2010, 82, 3788–3795

Gold Nanoparticle Chemiresistor Sensor Array that Differentiates between Hydrocarbon Fuels Dissolved in Artificial Seawater James Scott Cooper,*,†,‡ Burkhard Raguse,†,‡ Edith Chow,‡ Lee Hubble,‡ Karl-Heinz Mu¨ller,‡ and Lech Wieczorek†,‡ Wealth from Oceans Flagship and Future Manufacturing Flagship, CSIRO Materials Science and Engineering, Lindfield NSW 2070, Australia Gold nanoparticle films (AuNPF) functionalized with a range of hydrophobic and hydrophilic thiols were assembled in chemiresistor sensor arrays that were used to differentiate between complex mixtures of analytes in the aqueous phase. A chemiresistor array sampled a simple system of linear alcohols (methanol, ethanol, propan-1-ol, and butan-1-ol) dissolved in water over a range of concentrations. Discriminant analysis confirmed that the response patterns of the array could be used to successfully distinguish between the different alcohol solutions at concentrations above 20 mM for all of the alcohols except methanol, which was distinguished at concentrations above 200 mM. Alcohol solutions more dilute than these concentrations had response patterns that were not consistently recognizable and failed cross validation testing. This defined the approximate limit of discrimination for the system, which was close to the limits of detection for the majority of the individual sensors. Another AuNPF chemiresistor array was exposed to, and successfully identified crude oil, diesel, and three varieties of gasoline dissolved in artificial seawater at a fixed concentration. This work is a demonstration that the pattern of responses from an array of differently functionalized AuNPF sensors can be used to distinguish analytes in the aqueous phase. Thiol-functionalized gold nanoparticle films (AuNPF) have been used in vapor phase sensing devices that can successfully identify organic vapors1,2 or even diagnose lung cancer from exhaled breath.3 We have shown that by maintaining the electrode dimensions below a critical size, it is possible to also operate AuNPF sensors in aqueous solutions, directly detecting dissolved * Corresponding author. E-mail: [email protected]. † Wealth from Oceans Flagship. ‡ Future Manufacturing Flagship. (1) Han, L.; Shi, X. J.; Wu, W.; Kirk, F. L.; Luo, J.; Wang, L. Y.; Mott, D.; Cousineau, L.; Lim, S. I. I.; Lu, S.; Zhong, C. J. Sens. Actuators, B: Chem. 2005, 106, 431–441. (2) Zhong, Q.; Steinecker, W. H.; Zellers, E. T. Analyst 2009, 134, 283–293. (3) Peng, G.; Tisch, U.; Adams, O.; Hakim, M.; Shehada, N.; Broza, Y. Y.; Billan, S.; Abdah-Bortnyak, R.; Kuten, A.; Haick, H. Nat. Nanotechnol. 2009, 4, 669–673.

3788

Analytical Chemistry, Vol. 82, No. 9, May 1, 2010

hydrocarbon analytes.4-7 The AuNPF property that changes when it detects an analyte, in either the vapor or aqueous phase, is its electrical resistance.8 This is because when analyte interacts with a AuNPF it alters the film’s principal method of electron transport, electron tunneling.9 Under equilibrium conditions, the electron tunneling resistance RT of the AuNPF is given by10-12 RT ∝ exp(βL) exp(EA /kT)

(1)

where L is the average separation gap between adjacent nanoparticles, β is the electron tunneling decay constant which depends on the matrix of thiols between the nanoparticles, EA is the activation barrier energy, k is the Boltzmann constant, and T is the temperature.13 If the nanoparticle film is exposed to an analyte, then the analyte molecules will diffuse into the matrix of thiols between the nanoparticles and cause swelling such that the average separation gap between nanoparticles will increases from L to L + ∆L. The nanoparticles in this study are relatively large (>5 nm) so Coulomb blockade effects, local core thermal motion, and changes in the dielectric permittivity can be neglected at room temperature.13-16 The relative change (4) Raguse, B.; Chow, E.; Barton, C. S.; Wieczorek, L. Anal. Chem. 2007, 79, 7333–7339. (5) Chow, E.; Herrmann, J.; Barton, C. S.; Raguse, B.; Wieczorek, L. Anal. Chim. Acta 2009, 632, 135–142. (6) Raguse, B.; Barton, C. S.; Mu ¨ ller, K.-H.; Chow, E.; Wieczorek, L. J. Phys. Chem. C 2009, 113, 15390–15397. (7) Chow, E.; Gengenbach, T. R.; Wieczorek, L.; Raguse, B. Sens. Actuators, B: Chem. 2010, 143, 704–711. (8) Krasteva, N.; Fogel, Y.; Bauer, R. E.; Mullen, K.; Joseph, Y.; Matsuzawa, N.; Yasuda, A.; Vossmeyer, T. Adv. Funct. Mater. 2007, 17, 881–888. (9) Zabet-Khosousi, A.; Dhirani, A. A. Chem. Rev. 2008, 108, 4072–4124. (10) Terrill, R. H.; Postlethwaite, T. A.; Chen, C. H.; Poon, C. D.; Terzis, A.; Chen, A. D.; Hutchison, J. E.; Clark, M. R.; Wignall, G.; Londono, J. D.; Superfine, R.; Falvo, M.; Johnson, C. S.; Samulski, E. T.; Murray, R. W. J. Am. Chem. Soc. 1995, 117, 12537–12548. (11) Mu ¨ ller, K.-H.; Herrmann, J.; Raguse, B.; Baxter, G.; Reda, T. Phys. Rev. B 2002, 66, 75417 1–8. (12) Herrmann, J.; Mu ¨ ller, K.-H.; Reda, T.; Baxter, G. R.; Raguse, B.; de Groot, G. J. J. B.; Chai, R.; Roberts, M.; Wieczorek, L. Appl. Phys. Lett. 2007, 91, 183105 1–3. (13) Mu ¨ ller, K.-H.; Wei, G.; Raguse, B.; Myers, J. Phys. Rev. B 2003, 68, 155407. (14) Herrmann, J.; Bray, D. J.; Mu ¨ ller, K.-H.; Wei, G.; Lindoy, L. F. Phys. Rev. B 2007, 76, 212201 1–4. (15) Choi, J. P.; Coble, M. M.; Branham, M. R.; DeSimone, J. M.; Murray, R. W. J. Phys. Chem. C 2007, 111, 3778–3785. (16) Mu ¨ ller, K.-H.; Herrmann, J.; Wei, G.; Raguse, B.; Wieczorek, L. J. Phys. Chem. C 2009, 113, 18027–18031. 10.1021/ac1001788 Published 2010 by the American Chemical Society Published on Web 04/12/2010

of the nanoparticle film resistance ∆R/R0 upon exposure to the analyte is dominated by the change to the interparticle separation and is given by8 ∆R R(L + ∆L) - R(L) ) exp(β∆L) - 1 ) R0 R(L)

(2)

In the case of small swelling, one finds ∆R/R0 ≈ β∆L and the change in separation distance is given by6 Fa,water ∆L ) ηP Fa

(3)

where P is the AuNPF/H2O partition coefficient of the analyte, the ratio Fa,water/Fa is the concentration of the analyte in parts per million (ppm), and the quantity η is a geometrical factor that depends on the degree of interdigitation of adjacent selfassembled monolayers (SAMs) of thiols.6 From eqs 2 and 3 it follows that the nanoparticle film’s relative resistance is linearly proportional to both the partition coefficient of the film itself and the concentration of the analyte molecules it is exposed to. The partition coefficient toward a particular analyte has been shown to depend on the functional groups that make up the SAM.5,6 Thus, by varying the physicochemical properties of the thiolate ligands around the nanoparticles, we expect to be able to change the degree of partitioning of the different analytes, effectively tuning the film’s sensitivity to different chemicals of interest. When an array of AuNPF chemiresistors, each functionalized with different thiolate ligands, is exposed to an analyte, each sensor in the array has a unique response. The overall pattern of responses forms a signature that is defined by the chemistry and concentration of that particular analyte.17 Analysis of these signatures with statistical tools can confirm whether the chemical species can be discriminated by the array.18 Statistical analyses are commonly used to quantify the performance of sensor arrays.19-24 In the current work we describe the formation of arrays of AuNPF chemiresistors, functionalized with a variety of thiols and their use in detecting and differentiating between different alcohols and between different, complex hydrocarbon fuels such as crude oil, diesel, and three types of gasoline dissolved in artificial seawater. Although complex mixtures, like different fuel blends, have been examined before, the research was either directed at measuring undiluted fuel samples using (17) Legin, A.; Rudnitskaya, A.; Vlasov, Y. In Comprehensive Analytical Chemistry; Alegret, S., Ed.; Elsevier: Amsterdam, The Netherlands, 2003; Vol. XXXIX, pp 437-486. (18) Jurs, P. C.; Bakken, G. A.; McClelland, H. E. Chem. Rev. 2000, 100, 2649– 2678. (19) Lu, Y. J.; Partridge, C.; Meyyappan, M.; Li, J. J. Electroanal. Chem. 2006, 593, 105–110. (20) Ying, Z. H.; Jiang, Y. D.; Du, X. S.; Xie, G. Z.; Yu, J. S.; Tai, H. L. Eur. Polym. J. 2008, 44, 1157–1164. (21) Yoshikawa, G.; Lang, H. P.; Akiyama, T.; Aeschimann, L.; Staufer, U.; Vettiger, P.; Aono, M.; Sakurai, T.; Gerber, C. Nanotechnology 2009, 20, 5. (22) De, M.; Rana, S.; Akpinar, H.; Miranda, O. R.; Arvizo, R. R.; Bunz, U. H. F.; Rotello, V. M. Nat. Chem. 2009, 1, 461–465. (23) Xu, Z.; Shi, X. J.; Wang, L. Y.; Luo, J.; Zhong, C. J.; Lu, S. S. Sens. Actuators, B: Chem. 2009, 141, 458–464. (24) Brudzewski, K.; Ulaczyk, J. Sens. Actuators, B: Chem. 2009, 140, 43–50.

ultraviolet light spectroscopy25 or measuring the vapors originating from undiluted samples using metal oxide sensor arrays.26-28 The ability to discriminate between different hydrocarbon fuels represents an interesting challenge for a sensor array as the fuels consist of similar hydrocarbons albeit of different molecular weights.29,30 Additionally, the development of simple and potentially portable sensors to detect and discriminate between hydrocarbons in seawater could have interesting applications in oil prospecting or environmental monitoring. EXPERIMENTAL SECTION Materials. Gold(III) chloride trihydrate (HAuCl4 · 3H2O), tetraoctylammonium bromide (TOAB), 4-(dimethylamino)pyridine (DMAP), N-methyl-2-pyrrolidone (NMP), acetonitrile, sodium borohydride, potassium chloride, potassium iodide, iodine, propan-1-ol, and butan-1-ol were from Sigma-Aldrich, Australia. (3-Mercaptopropyl)triethoxysilane (MPTES), 1-hexanethiol, 2-naphthalenethiol, 2-phenylethanethiol, 1,10-decanedithiol, 4-fluorothiophenol, 4-bromothiophenol, 1-thioglycerol, and trans4,5-dihydroxy-1,2-dithiane were from Fluka, Australia. Sulfuric acid, hydrochloric acid, and nitric acid were from AJAX, Australia. Toluene, methanol, ethanol, and acetone were from LabScan, Thailand. Dichloromethane (DCM) was from BDH, Australia, acetone was from Chem-Supply, Australia, ethanol was from CSR, Australia, propan-2-ol was from Merck, Germany, and Deconex OP-120 glass cleaning solution was from Borer Chemie, Switzerland. All reagents are of analytical grade and were used as received. The E10 ethanol blended unleaded, regular unleaded, U8000 premium unleaded, and diesel gasolines were sourced from local fuel stations, and the crude oil was from the Northwest Shelf oil field, Western Australia, Australia. Solutions were prepared using Milli-Q deionized water (>18.0 MΩ Millipore, MA) unless otherwise stated. The salts used in the artificial seawater were sodium chloride from Aldrich; sodium sulfate from Merck, Germany; potassium chloride and boric acid from Standard Laboratories, Australia; sodium bicarbonate from BDH laboratory chemicals division, England; potassium bromide from By Products and Chemicals Pty. Ltd., Australia; sodium fluoride from Baker Chemical Co.; magnesium chloride from Fluka Analytical; calcium chloride from Ajax Chemicals, Australia; strontium chloride from Chemical Supply, Australia. Gold Band Microelectrodes. Gold band microelectrodes (nominal length of 3 mm, width of 5 µm, and an electrode spacing of 5 µm) were prepared on a glass microscope slide (Borofloat 33, Schott, Australia) using standard photolithographic techniques as described previously.4 Nanoparticle Film Deposition and Functionalization. Gold nanoparticles (6 nm nominal diameter) were synthesized by the (25) Steers, D.; Gerrard, C.; Hirst, B.; Sibbett, W.; Padgett, M. J. J. Opt. A: Pure Appl. Opt. 1999, 1, 680–684. (26) Lauf, R. J.; Hoffheins, B. S. Fuel 1991, 70, 935–940. (27) McCarrick, C. W.; Ohmer, D. T.; Gilliland, L. A.; Edwards, P. A.; Mayfield, H. T. Anal. Chem. 1996, 68, 4264–4269. (28) Sepcic, K.; Josowicz, M.; Janata, J.; Selby, T. Analyst 2004, 129, 1070– 1075. (29) Potter, L. P.; Simmons, K. E. In Total Petroleum Hydrocarbon Criteria Working Group Series, Vol. 2; Amherst Scientific Publishers: Amherst, MA, 1998; p 14. (30) Collins, C. D. In Methods in Biotechnology: Phytoremediation: Methods and Reviews; Willey, N., Ed.; Humana Press: Totowa, NJ, 2007; Vol. 23, p 100.

Analytical Chemistry, Vol. 82, No. 9, May 1, 2010

3789

Brust Method31 and stabilized in water as described by Gittins and Caruso32 with minor modifications.4 A printing ink was made up of 1% DMAP stabilized gold nanoparticles and 4% N-methyl2-pyrrolidone (NMP). An Autodrop printing system (Microdrop Technologies, Germany) dispensed 90 nL of the ink onto the center of the electrodes. At room temperature the water evaporated within 5 min and left a flat circular film of gold nanoparticles.5 A Teflon plate with wells that had Buna-N O-rings incorporated at the bottom, which aligned with the microelectrodes, was affixed to the glass slide. Various thiol solutions (150 µL, 10 mM in acetonitrile) were dispensed into each well and the wells were sealed with a glass lid. The thiols were given 1 h to displace the DMAP from the surface of the gold nanoparticle.7 The wells were flushed with acetonitrile followed by deionized water and then dried under a stream of nitrogen. Two different arrays were used for measuring the two different analyte groups, for the alcohol testing the sensors in the first array were functionalized with 1-hexanethiol, 2-phenylethanethiol, 2-naphthalenethiol, 1,10-decanedithiol, 4-fluorothiophenol, trans-4,5-dihydroxy-1,2-dithiane, and 4-bromobenzenethiol. The second array, used for hydrocarbon testing, was similar to the first array except the 1,10-decanedithiol functionalized sensor was replaced with a 1-thioglycerol functionalized sensor. The flexibility of the functionalization method means that it is trivial to tailor arrays with any combination of thiols and optimize the arrays specificity toward particular analytes. The conductivity of the eight different SAM coated gold nanoparticle films, in air, was determined from the film resistance and the film geometry. The height and diameter of dry films were measured with a Sloan Dektak 3030 stylus profilometer. For a nanoparticle film prepared from 90 nL of a 1% w/v solution of DMAP stabilized gold nanoparticle solution, functionalized with thiol in acetonitrile, the active area between the electrodes was approximately 180 nm high with a width of approximately 500 µm. Preparation of Aqueous Analyte Solutions. Artificial seawater was prepared by the Kester method33 except the final solution was not aerated. The saturated solutions of hydrocarbon fuels were prepared by carefully pipetting 10 mL of the fuel into a vial that contained 30 mL of artificial seawater. The vial was then sealed, and the mixture was magnetically stirred overnight without vortexing. The following day an aliquot of the aqueous layer (25 mL) was taken without disturbing the hydrocarbon fuel layer. Immediately prior to testing, these saturated aqueous solutions were diluted to 1:20, 1:10, and 1:1 with more artificial seawater. In contrast, the aqueous alcohol solutions were prepared by simply diluting with water. For butan-1-ol, propan-1-ol, and ethanol, a range of concentrations, 10, 50, 100, and 200 mM, of the alcohols was examined. For methanol, the concentrations examined were 100, 200, 500, and 1000 mM. These concentrations were selected because they are close to the limits of detection of the sensors in the array. Analyte Determination. The fluidic system consisted of four sample vials and a single analyte bottle of water that were (31) Brust, M.; Walker, M.; Bethell, D.; Schiffrin, D. J.; Whyman, R. J. Chem. Soc., Chem. Commun. 1994, 801–802. (32) Gittins, D. I.; Caruso, F. Angew. Chem., Int. Ed. 2001, 40, 3001–3004. (33) Kester, D. R.; Duedall, I. W.; Connors, D. N.; Pytkowic, Rm. Limnol. Oceanogr. 1967, 12, 176–179.

3790

Analytical Chemistry, Vol. 82, No. 9, May 1, 2010

Scheme 1. Equivalent Circuit Model That Shows How the Double Layer Capacitance (Cdl) Decouples the Chemiresistor Sensor Resistance (R) from the Analyte Solution Resistance (RS)

connected with Teflon tubing (1/16 in.) to a nine port MVP selection valve (Hamilton, NV). Stainless steel tubing (1/16 in.) connected the valve to the sensor array unit. The sensor array unit consisted of three parts; the glass slide with electrodes, a silicone gasket, and a stainless steel top plate. The silicone gasket in the sensor array unit formed the walls of a channel (2.9 mm wide and 1.6 mm high) that contained the chemiresistor sensors and separated the analyte from the electrical connections. A stainless steel top plate with 1/16 in. nanoport (Upchurch Scientific, WA) inlet and outlet ports completed the unit and sealed the roof of the channel. The outlet port from the sensor array unit was connected to a peristaltic pump (Ismatec, GlattbruggCH). The pump speed was typically 0.75 mL min-1. The pump and selection valve were controlled remotely by programs that were developed in-house using visual basic (Microsoft Corporation, WA). Previously, we employed low frequency (i.e., 1 Hz or less) impedance spectroscopy to monitor the sensor response.4,5 The sensor response could be adequately modeled using a simple RC network as shown in Scheme 1, where RS is the analyte solution resistance, R is the nanoparticle film resistance, and Cdl is the ionic double layer capacitance of the band electrode surface. It was shown that for testing at frequencies lower than 1 Hz, the small double layer capacitance imposes a high impedance that prevents the electrical current from traveling through the surrounding electrolyte solution. Given the apparent validity of the simple network, it follows that the sensor should also function on application of a dc bias to measure the chemiresistor response. Hence, results shown in the present paper were carried out with an applied dc bias voltage of 100 mV while measuring the current. Previously, we also found that the I-V characteristics of the nanoparticle films were linear between 0 and 0.5 V, and Ohm’s law can be used to obtain the change in nanoparticle film resistance in the presence of analyte.34 It should be noted that during dc testing, degradation of the film due to the possible electromigration of nanoclusters35 was not observed with the baseline resistance and response to analytes remaining consistent for many days, as long as the sensor array was continually immersed in water. The measurement system was based on a two probe resistance measurement.36 The electronics that biased the electrodes, converted the resultant currents to potential signals, and then amplified the signals were built in-house. The amplified potential (34) Raguse, B.; Herrmann, J.; Stevens, G.; Myers, J.; Baxter, G.; Mu ¨ ller, K.-H.; Reda, T.; Molodyk, A.; Braach-Maksvytis, V. J. Nanopart. Res. 2002, 4, 137–143. (35) Wohltjen, H.; Snow, A. W. Anal. Chem. 1998, 70, 2856–2859. (36) Maissel, L. I. In Handbook of Thin Film Technology; Maissel, L. I., Glang, R., Eds.; McGraw-Hill: New York, 1970; Vol. 1, pp 13-16.

Table 1. Thiols Used to Functionalize the Gold Nanoparticle Chemiresistors and the Measured Conductivity of the Respective Nanoparticle Films

signal was measured with a Powerlab 16-channel system (ADInstruments, NSW, Australia), and the data were recorded with Chart 5.5.6 software (ADInstruments, NSW, Australia). Within the Chart software a 10 Hz low-pass filter was applied to each signal, and with this filter the resistance measurements had a 0.02% variation from a stable baseline. Data were processed with Excel 2003 (Microsoft Corporation, WA) Xlstat v2008.6.4 (Addinsoft, NY), and all figures were plotted with OriginPro v8.0891 software (OriginLab Corporation, MA). RESULTS AND DISCUSSION Previously we have shown that by changing the nature of the SAM, e.g., using 6-hydroxyhexanethiol and 1-hexanethiol or by using mixtures of 1-hexanethiol and 4-mercaptophenol, it was possible to partially change the selectivity of the chemiresistor sensor toward nonpolar analytes (e.g., hexane)6 and polar analytes (e.g., ethanol).7 The development of such semiselective chemiresistors was encouraging for the further development of a sensor array of differently functionalized gold nanoparticle films. Statistical methods, such as discriminant analysis, can potentially be used to extract useful analytical information from arrays of semiselective sensors. This could lead to the development of so-called “electronic noses” and “electronic tongues” that are used to discriminate between complex mixtures of analytes.37 (37) Gouma, P.; Sberveglieri, G.; Dutta, R.; Gardner, J. W.; Hines, E. L. MRS Bull. 2004, 29, 697–700.

In the present work the chemiresistor array is initially produced by inkjet printing a DMAP stabilized gold nanoparticle solution. As shown previously, the DMAP is displaced by thiols allowing us to readily produce arrays of differently functionalized gold nanoparticle films. This is preferable to the de novo synthesis of the different thiol-functionalized gold nanoparticle materials and furthermore avoids having to optimize the printing conditions for each different thiol-gold nanoparticle combination. In the present work, a range of hydrophobic, alkane and aromatic thiols, as well as two hydrophilic (hydroxy-functionalized) thiols that were used in the chemiresistor array. The electrical conductivities of the various functionalized nanoparticle films obtained from this study are shown in Table 1. Comparing the conductivity of these AuNPF with literature values is difficult because there are several variables that can influence the conductivity (e.g., thiol to nanoparticle ratio,38 nanoparticle size,39 interparticle linking,15 etc.). However, it was found that the conductivities of the nanoparticle films were in broad agreement with previously published literature values.39-46 In order to test the ability of the sensor array to discriminate between different solutions, each containing a single analyte, the chemiresistor sensor array was initially exposed to aqueous (38) Snow, A. W.; Wohltjen, H. Chem. Mater. 1998, 10, 947–949. (39) Evans, S. D.; Johnson, S. R.; Cheng, Y. L.; Shen, T. J. Mater. Chem. 2000, 10, 183–188. (40) Wuelfing, W. P.; Green, S. J.; Pietron, J. J.; Cliffel, D. E.; Murray, R. W. J. Am. Chem. Soc. 2000, 122, 11465–11472.

Analytical Chemistry, Vol. 82, No. 9, May 1, 2010

3791

Figure 1. Resistance change for 4-fluorothiophenol-AuNPF and trans4,5-dihydroxy-1,2-dithiane-AuNPF sensors exposed to 0.2 M propan1-ol in water. Note that the scales of the two curves are the same, but the origins have been shifted for simpler comparison. The markers indicate the time that the sensor array was exposed to the aqueous propan-1-ol solution.

solutions of simple linear aliphatic alcohols (i.e., methanol, ethanol, propan-1-ol, and butan-1-ol). Methanol is completely miscible in water and is the most polar alcohol with a high relative dielectric constant of 33.0.47 In comparison, butan-1-ol is the least polar with a dielectric constant of 17.8 and limited solubility in water (79 g per kg of H2O).47,48 Qualitatively, one would expect that a polar alcohol would only weakly partition into the hydrophobic 4-fluorothiophenol-AuNPF but would strongly partition into the trans4,5-dihydroxy-1,2-dithiol-AuNPF which has a more polar organic matrix. Figure 1 shows a typical response of a 4-fluorothiophenol-AuNPF and a trans-4,5-dihydroxy-1,2-dithiane-AuNPF chemiresistor toward a 0.2 M aqueous solution of propan-1-ol. The response time of the sensor array was reasonably quick with all of the sensors stabilizing within 60 s after being exposed to the analyte. The response of a sensor to an analyte is defined as the relative change in resistance ∆R/R0, 60 s after exposure. As can be seen in Figure 1, the trans-4,5-dihydroxy-1,2-dithianeAuNPF chemiresistor had an approximately 3 times larger response toward propan-1-ol than the more hydrophobic 4-fluorothiophenol-AuNPF chemiresistor (i.e., ∆R/R0 ) 4.2% versus 1.5%, respectively). Examination of ∆R/R0 for all of the data generated by the array showed that there was generally a positive skew to their distributions. To account for this and improve the validity of the following discriminant analysis, eq (41) Wuelfing, W. P.; Murray, R. W. J. Phys. Chem. B 2002, 106, 3139–3145. (42) Joseph, Y.; Besnard, I.; Rosenberger, M.; Guse, B.; Nothofer, H. G.; Wessels, J. M.; Wild, U.; Knop-Gericke, A.; Su, D. S.; Schlogl, R.; Yasuda, A.; Vossmeyer, T. J. Phys. Chem. B 2003, 107, 7406–7413. (43) Leopold, M. C.; Donkers, R. L.; Georganopoulou, D.; Fisher, M.; Zamborini, F. P.; Murray, R. W. Faraday Discuss. 2004, 125, 63–76. (44) Yang, C.-Y.; Li, C.-L.; Lu, C.-J. Anal. Chim. Acta 2006, 565, 17–26. (45) Joseph, Y.; Peic, A.; Chen, X. D.; Michl, J.; Vossmeyer, T.; Yasuda, A. J. Phys. Chem. C 2007, 111, 12855–12859. (46) Wang, G. R.; Wang, L.; Rendeng, Q.; Wang, J.; Luo, J.; Zhong, C.-J. J. Mater. Chem. 2007, 17, 457–462. (47) Wohlfarth, C. In CRC Handbook of Chemistry and Physics; Lide, D. L., Ed.; CRS Press, Taylor and Francis: Boca Raton, FL, 2009; p 6-150 and 6-154. (48) Goral, M.; Wisniewska-Goclowska, B.; Maczynski, A. J. Phys. Chem. Ref. Data 2006, 35, 1391–1414.

3792

Analytical Chemistry, Vol. 82, No. 9, May 1, 2010

Figure 2. Average chemiresistor responses to 100 mM solutions of the alcohols The thiols used to functionalize the sensors, labeled 1, 2, 3, 4, 5, 6, and 8, are defined in Table 1. The error bars are defined by the standard deviation of four repeated measurements of each sample.

4 was used to normalize the responses of the sensors. This smoothed the positive skew in the distribution of the data but did not amplify the noise from sensors that did not respond.

(∆R/R0)Norm )

{

4 + log(∆R/R0) for ∆R/R0 > 10-4 0

for ∆R/R0 < 10-4 (4)

An array with seven active electrode sensors was used to examine solutions of alcohols at a variety of concentrations. Figure 2 shows the average normalized response from the seven chemiresistors to 100 mM solutions of methanol, ethanol, propan1-ol, and butan-1-ol. The responses of the chemiresistors to each analyte and at each tested concentration are available in Figure S-1 in the Supporting Information. As expected lower concentrations elicited weaker responses from the sensors and the longer alkyl chain length alcohols elicited larger responses from each sensor. This indicated that more hydrophobic alcohols partition more readily into the hydrophobic moiety of the SAM coating. The average response of the chemiresistors approximately followed the trend: trans-4,5-dihydroxy-1,2-dithiane-AuNPF > 1-hexanethiol-AuNPF > 2-phenylethanethiol-AuNPF > 2-naphthalenethiolAuNPF > 1,10-decanedithiol-AuNPF > 4-bromobenzenethiol-AuNPF > 4-fluorothiophenol. If the concentration is fixed, then an array is superfluous and a single sensor can discriminate the different analytes, i.e., the strongest response must be from butan-1-ol and the weakest response must be from methanol. However, in practical circumstances, concentration would not necessarily be the same between samples or even be known. If concentration is varied then the response of one sensor to a high concentration of a weakly partitioning analyte (e.g., 1000 mM methanol) and a low concentration of a strongly partitioning analyte (e.g., 10 mM butan-1-ol) would be the same. In these circumstances, the magnitude of an individual sensor’s response cannot be used to distinguish different analytes. In this scenario of variable concentration, it is only possible to distinguish analytes by using an array of sensors and analyzing all of their responses simultaneously. The relative

responses of the sensors in an array form a pattern that is defined by the partition coefficient of the analyte and is consistent across a range of concentrations. To ensure that the discrimination achieved in this study was valid, a range of concentrations were tested. The concentrations were selected so that the magnitude of the sensors’ responses to the weakest partitioning analyte (methanol) overlapped with the responses to the strongest partitioning analyte (butan-1-ol). The sensitivity of the sensors can be estimated by fitting a straight line to the response data and measuring the gradient across the different concentrations.49 Extrapolating the line to the point where ∆R/R0 ) 3(rmsnoise) is a way to estimate the sensors’ limit of detection (LOD).49,50 The accuracy of this estimate relies heavily upon how the noise is measured. We used 500 data points (∼9 min of sampling time) while the sensor array was exposed to water. The complete details for the calculation of rmsnoise and the LOD of the sensors are given in the Supporting Information. Reasonable estimates of the intrinsic noise gave consistent, conservative limits of detection. For methanol, the average LOD of the sensors was ∼100 mM with the sensors being more sensitive as the alcohols’ alkyl chain got longer, such that the sensors had an average LOD of ∼5 mM for butan-1-ol. It should be noted that the point of interest of the present study is the limit of discrimination of the sensor array and not the limit of detection of the individual sensors. It is still possible to discriminate two different analytes even when the concentration is lowered to the point where some of the sensors fail to respond. As long as a few of the sensors in the array respond and form a distinct pattern then it is still possible to recognize the analyte. Obviously the pattern from weak responses with high contributions of noise and error is more likely to be misclassified than a clear pattern of strong responses. To assess whether the responses generated by the sensor array can be used to discriminate different analytes, we use a supervised statistical technique. The data was examined by a multivariate discriminant analysis (MDA) using a stepwise forward model and equality of covariance assumed. The stepwise forward method starts with no variables in the model and then one by one includes each statistically significant variable. First an analysis was performed on only the high concentration responses (200 mM butan-1-ol, propan-1-ol, ethanol, and 1000 mM methanol). For this concentration, the four different analytes were clearly discriminated and the confusion matrix from the cross validation test showed 100% perfect recognition. As lower concentrations were introduced into the analysis (200, 100, and 50 mM for butan-1-ol, propan-1-ol, and ethanol and 1000, 500, and 200 mM for methanol), the results showed that the different analytes could still be clearly discriminated as can be seen in Figure 3. For this analysis, the within class variance was reduced by omitting the first and last points from each data set, which accounted for the nonequilibrium conditions of the first measurement and any unrelated noise spikes that occurred. There was high collinearity, which the stepwise forward model dealt with by only including the responses from five sensors: 2-phenylethanethiol-AuNPF, 1,10-decanedithiol-AuNPF, trans-4,5-dihydroxy-1,2-dithiane-AuNPF, 2-naphthalenethiol-AuNPF, (49) Currie, L. A. Pure Appl. Chem. 1995, 67, 1699–1723. (50) Li, J.; Lu, Y. J.; Ye, Q.; Cinke, M.; Han, J.; Meyyappan, M. Nano Lett. 2003, 3, 929–933.

Figure 3. Discriminant analysis of the sensor arrays response to the four alcohol solutions each tested at three different concentrations; stars indicate the centroids of the groups and the circles are the 95% confidence ellipses.

and 4-fluorothiophenol-AuNPF. Figure 3 is a good representation of the entire data set with the two principal components representing 92.3% of data’s total variance. Cross validation of the samples was still perfect with 100% correct classification, as would be expected when the principal component plot shows no overlap occurring between the groups. When the complete data set which encompasses the entire concentration range for each analyte is analyzed, the quality of discrimination is reduced. A cross validation report of the entire data set showed that 6 out of 64 samples were misclassified: 2 each from the methanol, ethanol, and butan-1-ol groups but none from the propan-1-ol group. Unsurprisingly it was the low concentration observations that were misclassified showing that at a concentration of 10 mM for butan-1-ol and ethanol and 100 mM for methanol the patterns of responses from the array are not clear enough to guarantee discrimination. This coincides closely with the limits of detection for these analytes, showing that as the sensor responses gets weaker with stronger contributions of noise, the power to discriminate the different analytes is also reduced. In comparison to the simple alcohols, hydrocarbon fuels consist of more complex mixtures of chemical species but with similar functional groups. These analytes therefore represent a greater challenge for the sensor array. Testing aqueous solutions of these fuels will help to ascertain whether the sensor array is capable of functioning when more than one analyte interacts with AuNPF chemiresistors, and by using artificial seawater we can demonstrate that the sensor array will function in an unusual environment. One of the challenges of field testing a sensor array in real seawater, temperature fluctuation, was not replicated in these laboratory tests because it can be compensated for by altering the system (e.g., incorporating a heater) or by optimizing the dimensions of the nanoparticles.16 The analytes that were dissolved in the artificial seawater were gasoline, diesel, and crude oil and they consist of a range of aliphatic and aromatic hydrocarbons. The molecules that make up gasoline typically have between 4 and 12 carbon atoms, whereas diesel has a distribution of chain lengths between 8 and 21 carbon atoms and the crude oil is expected to contain Analytical Chemistry, Vol. 82, No. 9, May 1, 2010

3793

Figure 4. Sensor responses to 1:1 diluted saturated hydrocarbon solutions in artificial seawater. The thiols used to functionalize the sensors are defined in Table 1. The error bars are defined by the standard deviation of six repeated measurements of each fuel-seawater mix.

a complex mixture of all the components.29,30 Of the three gasolines, the premium gasoline has a higher octane rating than the regular gasoline, indicating that it has a higher concentration of antiknock additives and the ethanol blend is advertised to contain up to 10% ethanol. Solutions of artificial seawater spiked with hydrocarbon fuels were tested with another sensor array that had seven chemiresistors: 1-hexanethiol-AuNPF, 2-phenylethanethiol-AuNPF, 2-naphthalenethiol-AuNPF, 4-bromobenzenethiol-AuNPF, 4-fluorothiophenol-AuNPF, 1-thioglycerol-AuNPF, and trans-4,5-dihydroxy-1,2dithiane-AuNPF (labeled 1, 2, 3, 5, 6, 7, and 8, respectively, in Figure 4 and Table 1). During testing the sensor was first flushed with artificial seawater and then exposed to the diluted saturated solution of fuel and artificial seawater for 60 s and then flushed with artificial seawater again. This generated response curves similar to those in Figure 1 where again the response of the sensor was defined as the relative change in the resistance 60 s after exposure (∆R/R0) and is plotted in Figure 4 for each sensor. The patterns of the sensors responses to the three gasolines were quite similar with the largest response from the 1-hexanethiolAuNPF sensor and the weakest from the trans-4,5-dihydroxy1,2-dithiane-AuNPF sensor. In contrast, the patterns for the crude oil and diesel are dissimilar, with relatively weaker responses from the 4-bromobenzenethiol-AuNPF and 1-thioglycerol-AuNPF sensors and a stronger response from the trans-4,5-dihydroxy1,2-dithiane-AuNPF sensor. The response data from the 1:1 dilution sample set was normalized with eq 4 and processed by discriminant analysis to generate the principal component plot depicted in Figure 5. The data was again processed by a stepwise forward model with only two sensors not making a statistically significant contribution to the model, 4-fluorothiophenol-AuNPF and trans-4,5-dihydroxy-1,2dithiane-AuNPF. A strong correlation existed between sensors 1-hexanethiol-AuNPF, 2-phenylethanethiol-AuNPF, 2-naphthalenethiol-AuNPF, and 4-bromobenzenethiol-AuNPF indicated by correlation coefficients greater than 0.9. The majority of variation (92.4%) from these sensors is accounted for in the first principal component. In contrast, the 1-thioglycerol-AuNPF sensor had correlation coefficients that were less than zero with 3794

Analytical Chemistry, Vol. 82, No. 9, May 1, 2010

Figure 5. Principal component plot generated from the discriminant analysis of the chemiresistor array’s responses to artificial seawater saturated with hydrocarbon fuels and diluted 1:1. Stars indicate centroids and the circles indicate the 95% confidence ellipses for that group. The inset is the zoomed-in view of the gasoline data with the third principal component also plotted.

each of those four sensors. The variance accounted for by the second principal component (6.8%) mainly comes from the response of the 1-thioglycerol-AuNPF sensor. The aqueous solutions of diesel and crude oil, depicted in Figure 5, are clearly separated and discriminated from each other and the blends of gasoline. When considering the gasolines, the regular gasoline appears to overlap the ethanol blend which in turn appears to overlap the premium gasoline. Yet cross validation of the gasoline samples based on squared distances of four principal components shows there is no confusion between the samples. This is because the third principal component which only contains 0.5% of the total variance contributes to the discrimination of the gasoline samples, as can be seen in the inset frame of Figure 5. Thus the patterns of responses generated by the sensor array distinguish 1:1 diluted saturated solutions of different gasolines in artificial seawater. The sensor array was further tested to see if it could discriminate the hydrocarbon-seawater solutions at different concentrations. As well as the 1:1 dilution, two more concentrations were made by diluting saturated solutions by 1:10 and 1:20. Six repeat exposures of each hydrocarbon-seawater solution, each at the three different dilutions, were performed. This generated a total of 90 sensor response patterns which were again examined by multivariate discriminant analysis. The principal component plot output by the discriminant analysis is shown in Figure 6. With the inclusion of a range of concentrations in the analysis, the between class variance is decreased; this is visually evidenced by the groups in Figure 6 being closer together than the groups in Figure 5. The solutions of crude oil and diesel still remain distinct and cross validation testing of these groups continues to achieve 100% correct recognition for all of their samples. In contrast, the gasoline solutions are now thoroughly entwined in Figure 6. A cross validation report misclassifies many of these samples: 2 out of the 18 in the ethanol blend group, 8 out of 18 in the premium gasoline group, and 11 out of 18 for the regular unleaded group. Thus the sensor array can discriminate

cessfully discriminate between complex mixtures of hydrocarbons such as crude oil, diesel fuel, and gasoline dissolved in artificial seawater. It achieved these results even when testing over a range of concentrations. A major advantage of the AuNPF chemiresistor arrays is that their sensing properties can be readily changed by functionalizing the AuNPF with different thiols. The sensors are produced by inkjet printing of DMAP stabilized gold nanoparticles onto microelectrodes, followed by displacement of the DMAP stabilizer with different thiols. This method appears to be particularly suitable for efficiently producing arrays of gold nanoparticle films and opens up the possibility of tuning the response of the sensor array toward different analytes depending on the particular selection of thiols used to functionalize the gold nanoparticle films. Figure 6. Principal component plot generated from the discriminant analysis of the chemiresistor array’s responses to 1:1, 1:10, and 1:20 dilutions of artificial seawater saturated with different hydrocarbon fuels. Stars indicate centroids, and the circles indicate the 95% confidence ellipses for that group.

between artificial seawater samples that contain traces of crude oil, diesel, or gasoline, but it cannot, at this stage, accurately define the type of gasoline if the concentration is varied. CONCLUSIONS This work is the first demonstration that arrays of AuNPF chemiresistors, functionalized with different thiols, can discriminate between complex mixtures of analytes directly in aqueous solutions. With the use of multivariate discriminant analysis, the chemiresistor arrays were able to successfully discriminate between four simple alcohols dissolved in water. More significantly, the sensor arrays were also able to suc-

ACKNOWLEDGMENT The authors gratefully acknowledge Mark Roberts for building the electronic testing equipment, Roger Chai for his photolithography work, and Jan Myers for help synthesizing gold nanoparticles. We would also like to thank Andy Ross providing samples of Crude Oil from the North West Shelf. SUPPORTING INFORMATION AVAILABLE Further information on the sensor responses to aqueous solutions of the alcohols, the limit of detection of the individual sensors, and the method used to calculate the limit of detection. This material is available free of charge via the Internet at http://pubs.acs.org.

Received for review January 20, 2010. Accepted March 29, 2010. AC1001788

Analytical Chemistry, Vol. 82, No. 9, May 1, 2010

3795