Anal. Chem. 2008, 80, 606-611
Frequency-Coded Chemical Sensors Francis Tsow, Erica S. Forzani, and N. J. Tao*
Department of Electrical Engineering Arizona State University, Tempe, Arizona 85287
We have developed a new method to intelligently sample analytes and introduce the analytes to sensors. The method automatically adjusts sampling duration according to the sensors’ response to the analytes and converts the amplitude of the sensor output to a frequency output, giving us another opportunity to reduce noise in the signal. It also addresses some of the common sensor issues such as response time, saturation, chemical dynamic range, and sensor protection, saving precious detection time, protecting sensors, and enabling sensitive sensors built for low-concentration detection to be used for high-concentration detection as well. We have put together a system using a tuning fork chemical sensor as a sample sensor to demonstrate the feasibility and benefits of the new sensing technique. There has been much progress in the area of chemical sensors, particularly in the sensitivity or detection limit. For example, detection limits down to single molecules have been demonstrated using different technologies.1,2 However, in order to use sensors in the real-world environment, many challenges remain. Among them, sensor saturation that limits the dynamic range, slow response time due to mass transport, and overexposure leading to incapacity of the sensor are a few common limitations. As the sensor is exposed to a large amount of the targeted analyte, it can potentially cause any or all of the issues listed above. In a simple model, we can see that the sensor can be saturated so that it will not respond to any more analyte molecules leading to a limited dynamic range.3 Depending on the particular sensor mechanism that is employed, this saturation/over exposure can even result in an irreversible damage to the sensor, rendering it useless. In addition, as sensors are exposed to high concentration of analytes, their response times will typically lengthen as it takes longer time to reach a steady output.4-10 * To whom correspondence should be addressed. E-mail:
[email protected]. (1) Ishijima, A.; Yanagida, T. Trends Biochem. Sci. 2001, 26, 438-444. (2) Lazzarino, M.; De Marchi, E.; Bressanutti, M.; Vaccari, L.; Cabrini, S.; Schmid, C.; Poetes, R.; Scoles, G. Microelectron. Eng. 2006, 83, 13091311. (3) Grassi, M.; Malcovati, P.; Baschirotto, A. IEEE J. Solid-State Circuits 2007, 42, 518-528. (4) Aguilar, A. D.; Forzani, E. S.; Li, X. L.; Tao, N. J.; Nagahara, L. A.; Amlani, I.; Tsui, R. Appl. Phys. Lett. 2005, 87, 193108/1-193108/3. (5) Althainz, P.; Dahlke, A.; Frietschklarhof, M.; Goschnick, J.; Ache, H. J. Sens. Actuators, B 1995, 25, 366-369. (6) Forzani, E. S.; Zhang, H. Q.; Nagahara, L. A.; Amlani, I.; Tsui, R.; Tao, N. J. Nano Lett. 2004, 4, 1785-1788. (7) Jang, J.; Bae, J. Sens. Actuators, B 2007, 122, 7-13. (8) Lu, C.; Chen, Z.; Saito, K. Sens. Actuators, B 2007, 122, 556-559. (9) Pancheri, L.; Oton, C. J.; Gaburro, Z.; Soncini, G.; Pavesi, L. Sens. Actuators, B 2003, 89, 237-239.
606 Analytical Chemistry, Vol. 80, No. 3, February 1, 2008
To avoid saturation or overexposure of the sensor, often times the dynamic range of the sensor must be restricted11 or multiple sensors/sensing properties are employed.12,13 Lewis et al. implemented a pulsed system in their silicon sensor illustrating both the ability to protect the sensor from high concentration and the ability to reduce the effects of noise sources.14 However, their pulsing frequency was chosen by avoiding noise sources and within the range of the mechanical and chemical limitations of their devices.14 Bartlett et al. has addressed the response time issue of microelectrochemical transistor sensors by measuring the time it takes their sensors to reach a preset current level.15 They have shown in their papers the time response measurement method correlated well with the amperometric measurements.15 Nature uses a neuronal shuttering mechanism in vision, tactile, and olfactory systems, enabling them to sense small differences in neurotransmitter concentrations even at high concentration.16 A particularly interesting strategy in neurons is the frequencyencoded signal, i.e., information is carried in the frequency instead of the amplitude of a signal. One of our goals is to mimic this methodology to solve some of the sensor problems described above. The sensing method introduced in this article differs from most other sampling methods by taking advantage of the transient kinetics rather than steady-state response of the sensor and combining it with a microcontroller-enabled automatic switching mechanism. In addition to addressing the saturation, sensor protection, chemical dynamic range, and response time issues, our system also made unique amplitude-to-frequency response possible by combining the transient kinetics together with automatic switching to convert the amplitude response of the sensor into a pseudo-oscillation, with the oscillation frequency showing dependence on concentration. This is akin to converting an amplitude modulation into a frequency modulation, potentially affording better noise isolation and enabling noise reduction filtering methods, depending on the specific noise source and type of noise.17-20 The ability of the new system to perform multiple samplings automatically can also help reduce the chance of (10) Zangooie, S.; Bjorklund, R.; Arwin, H. Sens. Actuators, B 1997, 43, 168174. (11) Fung, Y. S.; Wong, C. C. W. Anal. Chim. Acta 2002, 456, 227-239. (12) Endres, H. E.; Mickle, L. D.; Kosslinger, C.; Drost, S.; Hutter, F. Sens. Actuators, B 1992, 6, 285-288. (13) Yamada, K.; Nakano, T.; Yamamoto, S. Electr. Eng. Jpn. 1997, 120, 34-40. (14) Lewis, S. E.; DeBoer, J. R.; Gole, J. L. Sens. Actuators, B 2007, 122, 2029. (15) Bartlett, P. N.; Wang, J. H.; James, W. Analyst 1998, 123, 387-392. (16) Shahrezaei, V.; Delaney, K. R. J. Neurophysiol. 2005, 94, 1912-1919. (17) Chang, K. Encyclopedia of RF and Microwave Engineering; John Wiley & Sons: Hoboken, New Jersey, 2005. (18) Friedt, J. M.; Carry, E. Am. J. Phys. 2007, 75, 415-422. (19) Giessibl, F. J. Rev. Mod. Phys. 2003, 75, 949-983. 10.1021/ac7016162 CCC: $40.75
© 2008 American Chemical Society Published on Web 12/29/2007
Figure 1. Schematic illustrations showing (A) a block diagram of the frequency-encoded sensing method and (B) the tuning fork sensor amplitude detection method.
erroneous results. We demonstrated a flexible, multiplatform detection system using a polymer-modified tuning fork chemical sensor platform capable of avoiding saturation of the sensor and its related problems, at the same time measuring high-concentration analytes and performing frequency modulation of the output signal potentially affording any means of noise reduction. EXPERIMENTAL DETAILS Experimental Setup. The details of our tuning fork sensor were described elsewhere.21-23 Briefly, a polymer wire is stretched across the two prongs of a microfabricated quartz tuning fork (R38-32.768-12.5, Raltron) (Figure 1A).21 When the tuning fork is driven to oscillate, the polymer wire is stretched and compressed. Upon exposure to a target analyte, the analyte molecules bind to the polymer wire and change the spring constant of the polymer wire which can be measured by monitoring the resonant frequency of the tuning fork. Rather than detecting directly the resonant frequency shift, we drove the tuning fork sensor into oscillation at a fixed frequency and monitored the oscillation amplitude. As the resonant frequency shifted, the amplitude also changed according to the frequency profile (amplitude response vs frequency) of the tuning fork sensor (Figure 1B). Thus, by monitoring the output amplitude of the tuning fork sensor, given the frequency profile, we know the change in resonant frequency of the tuning fork sensor. This method not only requires simpler electronics but also affords us an almost real-time monitoring of minor changes in the resonant frequency of the tuning fork sensor without the need of strenuous data processing after the data is (20) Grober, R. D.; Acimovic, J.; Schuck, J.; Hessman, D.; Kindlemann, P. J.; Hespanha, J.; Morse, A. S.; Karrai, K.; Tiemann, I.; Manus, S. Rev. Sci. Instrum. 2000, 71, 2776-2780. (21) Boussaad, S.; Tao, N. J. Nano Lett. 2003, 3, 1173-1176. (22) Ren, M. H.; Forzani, E. S.; Tao, N. J. Anal. Chem. 2005, 77, 2700-2707. (23) Tsow, F.; Tao, N. J. Appl. Phys. Lett. 2007, 90, 174102/1-174102/3.
collected. We used ethylcellulose (EC T10, Aqualon T10 Pharm EC, Hercules) polymer and poly(vinylphosphonic acid) polymer for ethanol and humidity detections, respectively. Detection Method Description. The exposure time of the tuning fork sensor to analyte is regulated by switching using threeway valves (990-000501-001 V2-12V-VCP 6psi E 7 C, Pneutronics) controlled by a microcontroller (MSP430F2013, Texas Instrument). The analyte vapor stored in a 40 L Tedlar bag is pumped into the housing of the tuning fork sensor which triggers a change in the sensor signal. When the signal reaches a preset level (Vhi), the microcontroller activates the valve which shuts off the analyte and at the same time allows purge gas (N2 or air) from another 40 L bag to flow into the housing. As a result, the signal falls until it reaches another preset level, Vlo, at which the microcontroller re-exposes the sensor to analyte by activating the valve (see Supporting Information Figure S-1 for a cartoon). This sequence of events, except the first ramp up, will then repeat itself until the experiment is stopped, resulting in a pseudo-oscillation from which switching frequency, output rising slope, time to first switching, etc. information can be extracted. From the extracted information, information on the presence and concentration of the analyte can be obtained. RESULTS AND DISCUSSION To demonstrate the principle, we selected the tuning fork sensor platform as the sensing element. We made this decision based on our familiarity, ease of use, high sensitivity, and relatively fast response of the sensor. A sensing system as detailed in the previous section was constructed with a T-10 polymer wire modified tuning fork sensor. We tested our sensing method using the mentioned sensing system and compared it with a typical conventional sensing system. We used water and ethanol as our sample analytes, and both have demonstrated the expected pseudo-switching in the output of the sensor system. We chose Analytical Chemistry, Vol. 80, No. 3, February 1, 2008
607
Figure 2. Response of the frequency-encoded sensing using ethanol vapors at 47.3 mg/L (A) and 3.9 mg/L (B) as model analytes. The purging gas was nitrogen, and the data was digitally filtered during data processing with a low-pass filter with a cutoff frequency at 10 Hz.
to focus on experiments using ethanol as the analyte, however. The response of the tuning fork sensor is typically quicker with ethanol than water. Notice that all concentrations used in this paper are defined as mass of analyte/L of diluting gas (N2 in the following data) under ambient temperature and pressure. Also, in our concentration calculations, we assume no analyte molecules were absorbed onto surfaces of the sample preparation bags or the detection system. Figure 2A shows the output of the system when 47.3 mg/L ethanol was used as the analyte while 3.9 mg/L data is shown in Figure 2B. We first ran various concentrations of ethanol to the new system to illustrate how the new method can be used in determining analyte concentration and at the same time addressing the issues mentioned in the introduction section. Notice the pseudo-oscillation frequency in the output changed as the concentration of the analyte changed (Figure 2). In fact, this pseudo-oscillation of our sensing method can be used not only to extract analyte concentration information, but also the amplitudeto-frequency conversion nature of the output can be used to potentially reduce noise by filtering out noises at uninterested frequencies.19 Figure 3 shows how the data obtained using the new sensing system can be used to determine concentration information. Concentration information can be extracted from how fast the sensor output changes (the slope in the data) (Figure 3A), the time it took to reach the first microcontroller switch point (Figure 3B), or the pseudo-oscillation frequency of the new system (Figure 3C). As can be seen from the plots (Figure 3), the sensor showed reasonable sensitivity even at or near saturated ethanol vapor of 142 mg/L. Multiple measurements (two to four measurements) were made. In general, the measurements were more reproducible at lower concentrations than at higher concentrations as can be seen in the plots with the error bars. The lower repeatability was likely due to insufficient purging time between different runs at the same concentration (approximately 5 min). In comparison with conventional methods, if we assume the controlling electronics to be properly designed, in other words, the electronics do not introduce significant additional noises, at lower concentrations, the repeatability should be the same. However, at higher concentrations, the conventional method would require even longer time to purge as the sensing element would be more saturated with analytes. If insufficient time was provided for purging, the conventional sensor output will be less reproducible in comparison with the frequency608
Analytical Chemistry, Vol. 80, No. 3, February 1, 2008
Figure 3. (A) Rising slope vs ethanol concentration for the second pseudo-oscillation cycle. (B) Time for the microcontroller to switch from analyte to purging gas for the first time vs ethanol concentration. (C) Normalized frequency vs ethanol concentration. (Normalized frequency is obtained by performing fast Fourier transform on the sensor output data for the first few cycles and subsequently dividing by the amplitude of the overshoot.)
coded sensor. In the extreme case, if the concentration was so high that irreversible damage was done to the conventional sensor, obviously, the output will be nonrepeatable. Figure 3A is a plot of the rising slope of the sensor output versus analyte concentration. We deliberately selected the second cycle instead of the first cycle for the following two reasons: (1) to try to avoid including the initial ramp up time from the start voltage to the lower of the two preset switching voltages, Vlo; instead, we would like to include the time the detection system took to go from Vlo to the peak voltage only, as this is repeated cycle after cycle rather than just the initial ramp up cycle (Figure 2, parts A and B); (2) to minimize the potential effects of different stabilization times before each run. The plot suggests that higher analyte concentration gives steeper slopes in the sensor voltage output. It is also evident from Figure 3A that the rate of change of the slope of the sensor output with respect to analyte concentration is larger for higher analyte concentration within the sub 15.8 mg/L regime (please refer to the following paragraph for details). It appears that Figure 3A suggests at least two rather distinct slopes for the sub 15.8 mg/L concentration and the above 15.8 mg/L concentration levels. A similar two-regime phenomenon also appears in Figure 3, parts B and C, time to first switch and normalized frequency versus concentration plots, respectively. In fact, this two-regime phenomenon was also observed using a separate tuning fork sensor. At this point, we do not know the exact reason behind this two-regime phenomenon; we believe that a combination of surface binding and diffusion models can be a possibility. As analyte molecules begin to fill up most of the binding sites, the rate of change of the output will now be dominated by diffusion inside the polymer wire, which is a slower process. As a result, after a certain concentration is reached, the rate of change of the output slope will change as the underlying dominant responsible mechanism is changed. The change in slope appears to be quite abrupt in the figures. We expect the change to be more gradual if more data points were taken around where the current switchover occurred. Another interesting observation that can also be related to diffusion inside the polymer wire is overshooting of the sensor output passing beyond the preset voltages. If we assume a simple model that the rate of change in sensor output (such as the rising slope of the output) is proportional to analyte concentration, one would expect the sensor output to reach the preset switch voltage (Vhi) faster for higher concentration. As a result, one would also expect the time to first switch to shorten and, similarly, the pseudooscillation frequency to increase. This simple model is confirmed at lower concentrations in both the time to first switch plot (Figure 3B) and the normalized frequency plot (Figure 3C). However, at higher analyte concentrations, the time to first switch (Figure 3B) seems to plateau and the normalized frequency (Figure 3C) decreases instead of increases. We do not have definitive evidence as to why the results deviate from the simple model as mentioned before. However, we attribute these deviations to either a direct or indirect result of the overshooting of the sensor output passing beyond the preset voltages. Notice that the sensor output continued to go up shortly after the microcontroller had switched the input valve to purging mode (overshoot) as can be seen in Figure 2, parts A and B. The plateau phenomenon of the time to
first switch plot (Figure 3B) is a result of additional contribution from the time lapse between Vhi and the peak output voltage (overshoot period). Notice there was also an undershoot (the sensor output continued to go down shortly after the microcontroller had switched the input valve back to sampling the analyte) (Figure 2A), but the amount of undershoot is relatively mild in comparison with the overshoot for the same concentration, particularly at higher concentrations. As the concentration of the analyte increased, the amount of overshoot also increased, offsetting the reduction in time to reach Vhi (as would have otherwise been expected using the simple model mentioned before) due to a steeper sensor output rising slope. As a result, the overall time to first switch appeared to remain approximately the same. As for the decrease in pseudo-oscillation frequency at higher analyte concentrations as analyte concentration increases, once again, overshoot in the sensor output is largely responsible. The overshoot reduces the pseudo-oscillation frequency in two ways: (1) It increases or offsets the reduction in rise time as would have been expected using the mentioned simple model due to higher analyte concentration. (2) It increases the fall time of the output voltage as the initial value for the recovery is now higher due to the overshoot. In fact, the increase in fall time had a greater impact on the overall oscillation frequency. As evident from the pseudo-oscillation plot for higher concentration analytes (Figure 2A), the time constant of the fall time is considerably larger than that for the rise time. We essentially believe that this is an inherent property of the tuning fork sensor and attribute internal diffusion as one of the possible reasons for this overshoot phenomenon as will be further discussed in the following paragraph. This overshoot can be mitigated in the future by implementing a control algorithm (such as proportional-integral, PI, or proportional-integral-derivative, PID, control systems) via adjusting the flow rate of the sample. We believe this overshooting effect is more than simply the dead volume of the test cell. In order to further investigate the overshoot mechanism, we have constructed a second sensing system employing a second three-way valve and another higher flow rate pump (2 L/min) dedicated to help purging of the analyte (see Supporting Information Figure S-2 for a schematic). This system is fundamentally very similar to the previous system except for the additional pump and valve, with the latter controlled by the complement (inverse) of the control voltage of the first valve. As a result, the second three-way valve turns on when the first valve is in purging mode, helping to sweep out the analytes in the test cell with a high flow rate pump. When the first valve is in sampling mode, the second valve will be switched such that it will become essentially transparent to the rest of the system. This resulted in basically the same sampling operation as the first system and achieving a purging rate approximately 8 times faster. We did not observe any obvious difference between data taken from the original system and the new two-valve system. The experiment indicates that the overshooting effect is independent of external diffusion-convective conditions of the analyte in the sensing system. Given the dead volume (estimated to be
