A Single Input Multiple Output (SIMO) Variation-Tolerant Nanosensor

Subscriber access provided by University of Sunderland

Article

A Single Input Multiple Output (SIMO) Variation-Tolerant Nanosensor Dong-Il Moon, Beomseok Kim, Ricardo Andre Peterson, Kazimieras Badokas, Myeong-Lok Seol, Debbie G. Senesky, Jin-Woo Han, and M. Meyyappan ACS Sens., Just Accepted Manuscript • DOI: 10.1021/acssensors.8b00510 • Publication Date (Web): 27 Aug 2018 Downloaded from http://pubs.acs.org on September 1, 2018

Just Accepted “Just Accepted” manuscripts have been peer-reviewed and accepted for publication. They are posted online prior to technical editing, formatting for publication and author proofing. The American Chemical Society provides “Just Accepted” as a service to the research community to expedite the dissemination of scientific material as soon as possible after acceptance. “Just Accepted” manuscripts appear in full in PDF format accompanied by an HTML abstract. “Just Accepted” manuscripts have been fully peer reviewed, but should not be considered the official version of record. They are citable by the Digital Object Identifier (DOI®). “Just Accepted” is an optional service offered to authors. Therefore, the “Just Accepted” Web site may not include all articles that will be published in the journal. After a manuscript is technically edited and formatted, it will be removed from the “Just Accepted” Web site and published as an ASAP article. Note that technical editing may introduce minor changes to the manuscript text and/or graphics which could affect content, and all legal disclaimers and ethical guidelines that apply to the journal pertain. ACS cannot be held responsible for errors or consequences arising from the use of information contained in these “Just Accepted” manuscripts.

is published by the American Chemical Society. 1155 Sixteenth Street N.W., Washington, DC 20036 Published by American Chemical Society. Copyright © American Chemical Society. However, no copyright claim is made to original U.S. Government works, or works produced by employees of any Commonwealth realm Crown government in the course of their duties.

Page 1 of 27 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Sensors

A Single Input Multiple Output Variation-Tolerant Nanosensor

(SIMO)

Dong-Il Moon1,2, Beomseok Kim1,2, Ricardo Peterson3, Kazimieras Badokas1,4, Myeong-Lok Seol1,2, Debbie G. Senesky3, Jin-Woo Han*,1,2, and M. Meyyappan1 1

Center for Nanotechnology, NASA Ames Research Center, Moffett Field, CA 94035, USA

2

Universities Space Research Association, NASA Ames Research Center, Moffett Field, CA

94035, USA 3

Department of Aeronautics and Astronautics, Stanford University, Stanford, CA 94305 USA

4

Institute of Photonics and Nanotechnology, Vilnius University, Vilnius, LT 10257, Lithuania

*Address correspondence to [email protected]

ABSTRACT:

Successful transition to commercialization and practical implementation of

nanotechnology innovations may very well need device designs that are tolerant to the inherent variations and imperfections in all nanomaterials including carbon nanotubes, graphene and others. As an example, a single walled carbon nanotube network based gas sensor is promising for a wide range of applications such as environment, industry, biomedical and wearable devices due to its high sensitivity, fast response and low power consumption. However, a longstanding issue has been the production of extremely high purity semiconducting nanotubes, thereby contributing to the delay in the market adoption of those sensors. Inclusion of even less than 0.1% of metallic nanotubes, which is inevitable, is found to result in a significant deterioration of sensor-to-sensor uniformity. Acknowledging the coexistence of metallic and semiconducting nanotubes as well as all other possible imperfections, we herein present a novel variation-tolerant 1 ACS Paragon Plus Environment

ACS Sensors 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 2 of 27

sensor design where the sensor response is defined by a statistical Gaussian measure in contrast to a traditional deterministic approach. The single input and multiple output data is attained using multiport electrodes fabricated over a relatively large area single nanotube ensemble. The data processing protocol discards outlier data points and the origin of the outliers is investigated. Both the experimental demonstration and complementary analytical modeling support the hypothesis that the statistical analysis of the device can strengthen the credibility of the sensor constructed using nanomaterials with any imperfections. The proposed strategy can also be applied to physical, radiation and biosensors as well as other electronic devices.

KEYWORDS: carbon nanotube, gas sensor, statistical analysis, sensor variation, printed electronics, variation tolerance

2 ACS Paragon Plus Environment

Page 3 of 27 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Sensors

Nanomaterials such as carbon nanotubes (CNTs), graphene and others have been successfully considered in a wide range of applications over the last two decades, but implementation of these advances in practical systems and commercialization has been rather slow.1 One major reason is the lack of consistency in device performance and device-to-device variations, causing reliability and reproducibility issues and impeding commercialization.2 While material quality and processing issues – which can be solved in the long run - can lead to this problem, the inherent nature of some of the nanomaterials may very well make it unrealistic to solve the problems using conventional approaches. An example can be made with the current situation of most applications using single walled carbon nanotube (SWCNTs), which lacks reliable and cost effective way to control the conductivity type (metallic versus semiconducting) and chirality, either during growth or post-processing.3 Likewise, graphene lacks a reliable and cost effective way to manipulate the number of layers precisely and uniformly and also to introduce a pre-specified bandgap in a controlled manner.4 Application development for these materials in sensors, electronics, photonics and others to date has acknowledged the above deficiencies, but largely relied on using device platforms that have long been used for thin films of silicon and other established materials. Successful transition to commercialization may very well need device designs that are tolerant to the inherent variations in nanomaterials and other imperfections. Here we propose and demonstrate such a concept with SWCNT gas sensor as a case study. The proposed methodology can be infused to bio and radiation sensors, physical sensors and other devices, as well as utilize other nanomaterials besides carbon nanotubes. Carbon nanotube based gas sensors can be constructed with either individual semiconducting SWCNT or an ensemble of SWCNTs.5-8 Sensors with individual semiconducting SWCNT have been shown to yield greater response than the ensemble nanotube


ACS Sensors 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 4 of 27

devices.9 The electrical response of the ensemble sensor is deteriorated by the inherent nature of semiconducting and metallic nanotube mixture in as-produced material as well as purified samples which are not sorted for nanotube type.10 Therefore, a certain degree of metallic nanotubes present in the network alters the response characteristics since the metallic portion is insensitive to charge transfer and other molecular interactions involved in gas sensing.11 In spite of the superiority of the individual nanotube based devices, most of the sensor studies reported in the literature as well as early commercialization attempts have focused on using CNT networks due to ease of fabrication.5,6 Also, the ensemble-type sensors are considered to be closer to volume manufacturing in the absence of any breakthrough in individual nanotube process, purity/type control and alignment issues.12,13 In fact, ensemble CNT devices have been demonstrated for various applications including integrated circuit, energy storage, displays and sensors.14-18 Even in all these demonstrations, however, the device to device variability inevitably remains as a fundamental challenge for commercialization because of the statistical randomness of metallic vs. semiconducting fraction, network formation and nanotube density.13 In order to tackle the foregoing variability issues, a variation-aware and variation-tolerant sensor design and data analysis strategies are presented here. The fabricated sensor consists of a single sensing material surrounded by multiple electrodes, resulting in a combination of data set that can be post-processed to improve the sensor reliability. The critical information from outlier data points, if any, which represent failure in conventional two terminal sensor devices, is deliberately excluded from the data set. The origins of such outliers are also investigated. The prototype sensors are fabricated fully by inkjet printing technology though the design would apply equally well to silicon and other substrates conventionally used in past nano chemsensor development.5-9 Recently, research and development in the field of printed electronics has


Page 5 of 27 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Sensors

attracted great deal of attention along with internet-of-things (IoT).19 The printing technology offers electronic devices built on flexible substrates such as paper, textile and plastic while providing reasonable performance and creating flexible, foldable, biodegradable, disposable and wearable applications.20-23 In this regard, almost all fundamental building blocks of electronic systems including lab-on-paper, thin film transistor, memory, display, photovoltaics, battery, supercapacitor and gas sensors have been demonstrated.15,24-28 The process uniformity in printing technology is significantly poorer than in solid state microfabrication technology. Therefore, the printed gas sensors can suffer from process-induced variability on top of the material-inducedvariability mentioned before.29

RESULTS AND DISCUSSION The presented multi-terminal single sensor is contrasted with the traditional two-terminal sensor array in Figure 1a and 1b. If N electrodes are placed, then N/2 individual devices and resultantly the same number of data set are produced in the conventional two terminal sensors (Figure 1a) whereas a total number of independent measurement set of N(N-1)/2 is possible in the single multiport sensor (Figure 1b). For example, a sixteen electrode system results in eight data points from the array of eight two-terminal devices while 120 measurement points are imported in one multiport sensor. As the number of electrodes N increases, the data size of the traditional two-terminal sensor array increases proportional to N while that of the multi-terminal single sensor increases proportional to N2. Therefore, the multiport sensor produces a great deal of data points for a given footprint, which is ideal for using machine learning techniques for the selective identification of an analyte from a complex background in potential electronic nose applications. In the multiport sensor, a resistance is measured from any arbitrary pair of


ACS Sensors 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 6 of 27

electrodes, and repeating the measurements for all possible combinations of electrode pairs creates the data set. Such measurement scheme is similar to impedance tomography30,31 or electrical resistance tomography32 of a two-dimensional film, which has been used for damage diagnosis in structural components. Inks of metallic silver nanoparticle and semiconducting SWCNT were respectively used for the electrode and active sensing layers here. Note that the commercial semiconducting nanotube ink intrinsically contains a small fraction of metallic nanotubes. Figure 1c shows the inkjet-processed sensor where a two-dimensional film of SWCNTs is connected to sixteen silver electrodes along the circumference. The fabrication details and the electrical characterization setup can be found in the Supporting Information (Figures S1 – S3). Figure 1d shows the distribution of the 120 measured baseline resistances from the printed device. As expected, a random distribution of baseline resistance is seen. As the electrode-to-electrode distance (EED) is different from one pair to another, the raw resistance in each case is converted to a so-called unit resistance by normalizing the measured resistance for each electrode pair to the corresponding EED. The distribution of unit resistance follows a Gaussian distribution function because the contact resistance is independent from the electrode distance (Figure 1e). The gas response tests were conducted with ammonia (NH3) in order to demonstrate the multiport sensor and validate the proposed methodology, though any other air pollutant could have been an equally good candidate as the objective here is not to focus on achieving the best sensitivity for any single analyte but to minimize variabilities discussed in the Introduction. In the real-time measurements here, purging and sensing cycles were repeatedly applied to evaluate the sensor performance with various NH3 concentrations in the range of 1-50 ppm. The details of the gas sensing setup can be found in the Supporting Information (Figure S4).


Page 7 of 27 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Sensors

Figure 1f presents representative measured ammonia response data from a single conventional two terminal sensor architecture, qualifying the intrinsic sensor material properties and confirming conventional sensor behavior. The resistance begins to increase upon exposure to ammonia. NH3 molecules interact through charge transfer when in contact with the sidewall of the nanotubes; in other words, the molecules bind to the nanotubes via physisorption.33 Because the nanotubes exhibit a p-type semiconducting behavior under ambient conditions, NH3 molecules adsorbing on the sidewall transfer electrons to the nanotubes due to an electrical potential difference between the two, and the electron transfer depletes the concentration of holes in the nanotubes, resulting in an increase in resistance.5,12,34 The same measurement was iterated 15 times and provided consistent results as shown in the error bar of Figure 1g. The sensor showed a linear response from 1 to 50 ppm and the detection limit extrapolated by the 3σ method is 115 ppb. The detection limit can be extrapolated from the linear calibration curve (see Section 8 of SI). Herein, the sensor response is defined as the ratio of the resistance shift over the initial resistance (Rt - Ro) / Ro, where Rt and Ro are resistance upon gas exposure and initial resistance, respectively.6 The sensor response time is the time needed to reach a stable output signal when an external stimulus is introduced. Typically, 95% of the final value is used to estimate the response time, assuming the stimulus is a step change.7 Similarly, the sensor recovery time is the time needed to reach its baseline when the stimulus is taken away. The response and recovery times depend on the time of exposure to the gas, gas concentration and purge time. Here, the response and recovery times were 243 seconds and 576 seconds, respectively for 20 ppm of ammonia. All these results are comparable to previous conventional chemiresistive sensor arrays fabricated on silicon and printed circuit board34 and no attempt was made to optimize performance. However,


ACS Sensors 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 8 of 27

the latter sensors show reduced sensitivity after only a few cycles whereas the nanotubes synthesized here and the SIMO sensor did not exhibit significant performance degradation over time during a 60-day test period (see Figure S5). The baseline resistance showed stability over 60 days, the sensor response of the SIMO sensor showed repeatability over time and the response of various sub-sensors in the array was reproducible as illustrated in Figure S5. In addition, the gas response is often influenced by the moisture in the environment.5,6 Operating the sensor at elevated temperatures is one way to nullify the impact of moisture. Another approach is to measure the gas response along with environmental relative humidity (RH) and compensate the result based upon the RH. The important factor in the latter case is whether the sensor itself responds to various humidity levels. The details on the impact of humidity on the present SIMO sensor can be found in the Supporting Information (Figure S6). One may argue that a large number of conventional two-terminal sensors can provide the same performance as the single input and multiple output (SIMO) sensor and possibly the consequence may not be dramatically different in terms of statistical results. First, the value of the SIMO sensor is in a readout system design. For example, generation of 120 data points from a multiple array of conventional two-terminal sensors needs 240 electrodes (or 121 electrodes if one electrode is common for all), whereas the SIMO scheme needs just 16 electrodes. Such a large number of electrodes demand a large sensor footprint. In addition, 120 devices in the conventional approach require 120 input-output (I/O) interfaces in the readout circuitry. The I/O lines are connected to a switching circuit, the so called multiplexer, that selects one of several input lines and forward it into one output line. The size of the multiplexing circuit increases approximately n x n, which means 120 I/O counts of the conventional approach consume not only large circuit area but also large operation power. A statistical analysis is carried out for both


Page 9 of 27 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Sensors

the multiple array of conventional two-terminal sensors and the SIMO sensor. Figure 2a and 2b show the initial resistance distribution and gas response respectively for both cases. The average and standard deviation of the initial unit resistance of the conventional sensors are 23.8 Ω/µm and 23.2 Ω/µm, respectively. The corresponding values for the SIMO sensor are 14.3 Ω/µm and 12.7 Ω/µm, respectively. When comparing the same number of data points, the distribution of the SIMO sensor is more uniform than that of the conventional two-terminal sensors. This result would be attributed to the fact that the SIMO includes only 16 contact resistances. This base resistance variability is directly reflected on the NH3 response too. The SIMO sensor has a 0.6 % higher average response and a 0.43 % lower standard deviation than the conventional sensors. Further comparison between the conventional and SIMO sensors is summarized in Table S1. The NH3 response values measured from each electrode pair of the SIMO sensor are distributed over 2.5 % to 5 % as seen in Figure 2b and Figure S7b showing the distribution spectrum. The spread among sub-sensors could be potentially worse, depending on several material/process quality factors. This spread is discouraging for the nanotube-based sensors, requiring further calibration. For that matter, any nanomaterial-based sensor is expected to behave similarly and require calibration schemes and it is possible to argue that every sensor needs calibration to ensure sensor accuracy and system integrity because of natural fluctuations. The calibration parameters are then stored in a microprocessor to accommodate and deal with the variability. However, considering the explosive demand of gas sensors in the IoT era, this need for full number calibration (as opposed to calibrating a representative number of devices) demands significant system overhead, reduces the manufacturing throughput and increases the cost.35 Furthermore, there may be some applications that the full number calibration is not possible at all. For example, the nature of printed electronics implies the on-demand


ACS Sensors 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 10 of 27

manufacturing of a certain component at certain specified time and location, but the gas test facility for calibration may not be available in that location. One such scenario includes the International Space Station with in space manufacturing capabilities where the printing equipment is already deployed but with no gas calibration equipment; such scenarios are easy to envision in terrestrial applications as well.

In this regard, the calibration-free application

scenario – which is highly desirable - demands a reproducible but credible gas sensor. Therefore, instead of the deterministic single reading in the traditional gas sensor, a statistical reading is used in the present variation-tolerant approach. The distribution of unit resistances measured from the SIMO sensors is shown in Figure 2c before (black) and after (red) the 10ppm ammonia exposure. The data shows normal distribution, and the simple average of all 120 responses shows 3.88 %, which can represent the SIMO response. In principle, the average of the samples of observations converges into nearly normal distribution only when the number of observations is sufficiently large. In practice, however, the data size of a traditional multiple sensor array is inevitably limited, which can lead to statistical errors. In contrast, the SIMO design in our example has a 120 data set and these are fitted into a continuous probability distribution function (Figure 2c). When the raw data is fitted into a Gaussian distribution using a least square parameter method, the sensor response based on the mean value is also 3.88 %. Considering the identical result from the simple average and the peak value of Gaussian fitting, the proposed SIMO scheme can provide reliable sensing results from the ensemble of CNT sensors. The aforementioned methods are based on random errors. There is, in principle, always a possibility of systematic error generating outlier points. Such outliers, normally considered as bad data, may be purposely excluded as long as the detection of erroneous data is credible. Two possible categories of outliers can be considered. The first one is structural outlier wherein the


Page 11 of 27 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Sensors

device itself is structurally defective so that the initial resistance deviates from the intrinsic resistance distribution. Another is a functional outlier wherein the device shows normal distribution but the gas response deviates a lot from the response distribution. As the sensor uses the baseline data as its reference, the impact of the initial resistance is important. An example of a structural outlier is shown in Figure 3a. The initial resistance distribution includes two groups that show two orders of magnitude greater resistance from some of the printed SIMO sensors. Using these outliers without examination might result in failure. However, the multiple readout from one SIMO sensor makes the self-detection of outliers possible. Examination of the fabricated SIMO sensor confirmed that these high resistances commonly came from a specific electrode. The inspection of that physical location corresponding to that electrical address confirmed the root of the systematic error to be the defect in the metal electrode (inset of Figure 3a). In spite of the failure in those two electrodes, the SIMO sensor still functions if such systematic erroneous data are detected followed by selective removal. A simple outlier test method is the Z-score method, Z = (Yi – Ym) / s, where Yi, Ym, and s are the sample value, mean and standard deviation, respectively.36 Other outlier test methods can be considered as well, for example, extreme studentized deviate (ESD) can be a good alternative, but Grubbs test or Tietjen-Moore test would not be suitable because the suspected number of outliers must be specified, which is impractical.37,38 In the present case, the resistance may be the sample value. Depending on the maturity of the material and process and the application tolerance, the Z-score threshold is determined. After excluding the outliers from the initial device, functional outlier checking can be carried out on the fly. In the functional outlier checking, the response can be the Y value in the Z-score test. As explained earlier, the sensor response is related to the initial resistance. The response points are then re-plotted as a function of each


ACS Sensors 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 12 of 27

initial resistance (Figure 3b). The general trend is that the higher initial resistance shows higher response. This result is consistent with our intuition that the higher resistance may contain less metallic nanotubes, resulting in greater response. However, a few outliers (red dots in Figure 3b) appear to deviate markedly from other observations in the sample. It should be emphasized that if the response values were simply distributed as shown in Figure S7b, the outlier points might be hidden in the tail of the normal distribution, thereby passing the outlier test. As seen in the scattered graph of Figure 3b, the functional outlier is bivariate outlier, showing unusual score on two variables. The red outlier points are not part of the centroid and clearly seen in visual inspection. Therefore, a simple Z-score test used to detect structural outliers from onedimensional histogram graph may be incomplete. In order to systematically detect the multivariate outlier, Mahalanobis Distance (MD), a distance of a data point from the calculated centroid, can be useful.39 The MD score accounts for the covariances between variables and takes into consideration that the variances in each variable are different. The squared MD score is MD2 = (X - µ)T Σ (X - µ), where X and µ are the two dimensional vector of observation and mean, respectively and Σ is the covariance matrix. Despite these outliers showing greater response, including them in the post measurement data processing may cause overestimation of the sensor performance. Accordingly, the erroneous data should be discarded. The red points in Figure 3b denote low initial resistance and high response. This fact is rather counterintuitive to common understanding since they represent possibly a sample with high metallic nanotube fraction. In order to further investigate the root of deviation, the physical addresses of the two red outlier points were inspected to gain insight. There was no physical signature of abnormality from optical and microscopical inspection and both points were from the shortest channel length in the combination of electrode pairs.


Page 13 of 27 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Sensors

Assuming that the resistivity of the nanotube film is uniform, the resistance is proportional to the length of the film according to Ohm’s law. So, it is natural to expect that those points are from the small resistance regions. However, when Figure 3b is replotted by replacing the resistance into the unit resistance in x-axis, the red outlier group appears at greater than the mean unit resistance region (Figure 3c). This fact leads to the possibility that the red outlier devices may suffer from relatively less formation of percolation path of metallic nanotubes. Further analysis is needed to analyze the origin of the outliers.10 Despite this metallic-CNT free devices presenting large sensor response,9 the deviation from average characteristics should be considered as functional outlier and thus excluded from the signal. The mean response including the outlier data was 3.88 % with a standard deviation of 0.55 % while the mean response excluding the outliers was 3.82 % with the standard deviation of 0.50 %. This similar response is due to the large sample size compared to few outliers. However, the lowered standard deviation signifies that the distribution becomes tighter and the response becomes reproducible. Based on the lesson from Figure 3, the systematic defects can add significant error but a few random point outliers (Figures 3a and 3b) can add a small error. This fact does not necessarily apply to all possible sensor cases. Sensors with different materials processed differently may have their own unique pattern of outliers. Thus, the noise patterns need to be classified after examining their characteristics. In addition, the threshold of outlier removal varies depending on the use case by trading off the need for sensor accuracy and the complexity of data analytics. In order to verify the consistency across one SIMO to another SIMO sensor, three different SIMO devices were fabricated and characterized. Whereas the individual two-terminal sub-sensor responses spread from 2.0 % to 5.5 %, all three SIMO sensors exhibited fairly


ACS Sensors 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 14 of 27

consistent results of 3.88 %, 3.88, and 3.91%, respectively. The detailed statistical data can be found in Table S1. Finally, analytical modeling was carried out to support the experimental results. The model is based on the total conductance of a bundle of nanotubes where the diameter, metal-tosemiconductor ratio and the total tube counts are considered as random variables but captured by Gaussian distribution. The nanotube density and the metal-to-semiconductor ratio to form a single ensemble are considered as variables in the model. The mean and the standard deviation of the density and the ratio are determined from the model to fit the experimental results. The average and standard deviation can be defined depending on the raw material of nanotubes, purification process, film formation method and device fabrication technology. The response to ammonia exposure is implemented to increase the resistance of the semiconducting nanotubes. The basis of the resistance modeling follows previous literature40 and the details and the model parameters are given in the Supporting Information. Figure 4 shows the probability distribution of the unit resistance response based on the model and the experimental results shown in Figure 2c. The agreement between them supports the hypothesis that the statistical analysis of the gas sensor can strengthen the reliability of the randomly distributed nanotube network based sensor. This modeling study can be further utilized to suggest design guidelines: how much variability of raw material is responsible to cause the device variability, how tight the printing process should be controlled (film density), and how many data samples are required to provide consistent SIMO response. As discussed here, the motivation for the present work is to increase the number of sampling in one sensor to lead to the convergence of population mean of the response. The results demonstrate the success of the approach in terms of reducing variability. Still, sensor


Page 15 of 27 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Sensors

material development will remain fundamental to improve the detection limit, sensitivity, response time and selectivity.6 Improvements in the material quality with negligible batch-tobatch variations in properties and reproducibility in sensor processing (for example, inkjetting of the material onto the sensor array) will help to reduce the number of samplings and relax the data mining complexity. In this work, the data analytics including outlier detection and statistic processing was conducted using computer software. From an engineering point of view, such data analytics need to be implemented in the hard ware itself for stand-alone sensor device.

CONCLUSIONS A variation-tolerant single input multiple output (SIMO) approach is experimentally demonstrated to improve the reliability of carbon nanotube ensemble based gas sensors. A single mat of carbon nanotube network is arranged with multiple electrodes on the circumference. The proposed device structure generates quadratic number of data set to manifest one sensor. As opposed to a discrete data point of a conventional two terminal chemiresistor, the SIMO sensor provides a probabilistic approach that can overcome the fundamental drawback of sensor-tosensor variability of the traditional chemiresistor based gas sensors. The methods to detect and exclude erroneous data are also demonstrated which can further improve the credibility of the sensor performance. Overall, the SIMO network of single walled carbon nanotube based gas sensor can offer high sensitivity, fast response and low power consumption for a wide range of applications in environmental monitoring, industrial leak detection, biomedical and healthcare. It is also possible to construct an electronic nose6 with this design by printing multiple sensor materials – each giving different response for the same gas input - in designated regions over the active area. Furthermore, the presented generic method can be applied to bio, radiation and a


ACS Sensors 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 16 of 27

variety of physical sensors and other nano devices as well as extending to other nanomaterials with inherent imperfections of any kind.


Page 17 of 27 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Sensors

ASSOCIATED CONTENT Supporting Information The Supporting Information is available free of charge on the ACS Publication website at DOI: Details on the fabrication process, electrical characterization, gas sensing experiments, sensor reliability/repeatability/stability, effect of humidity, comparison of conventional sensors with the proposed SIMO design and analytical model are discussed. AUTHOR INFORMATION Corresponding Author *Jin-Woo Han, E-mail: [email protected]; Phone: +1 650 604 3985 Notes The authors declare no competing financial interest. ACKNOWLEDGMENT This work was in part supported by the In-Space Manufacturing Program of the Advanced Exploration Systems (AES) Office of the NASA HEOMD. The authors are grateful to Niki Werkheiser and Jessica Koehne for their support. KB was an international summer intern during the summer of 2017.


ACS Sensors 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 18 of 27

REFERENCES (1)

Fadel, T.R.; Farrell, D.R.; Friedersdorf, L.E.; Griep, M.H; Hoover, M.D.; Meador, M.A.; Meyyappan, M. Toward the Responsible Development and Commercialization of Sensor Nanotechnologies. ACS Sensors 2016, 1, 207-216.

(2)

Franklin, A. D.; Tulevski, G. S.; Han , S.-J.; Shahrjerdi, D.; Cao, Q.; Chen, H.-Y.; Wong, H.-S. P.; Haensch, W. Variability in Carbon Nanotube Transistors: Improving Device-toDevice Consistency. ACS Nano 2012, 6, 1109–1115.

(3)

Zhang, Q.; Huang, J.-Q.; Qian, W.-Z.; Zhang, Y.-Y.; Wei F. The Road for Nanomaterials Industry: A Review of Carbon Nanotube Production, Post-Treatment, and Bulk Applications for Composites and Energy Storage. Small 2013, 9, 1237–1265.

(4)

Allen, M. J.; Tung, V. C.; Kaner, R. B.; Honeycomb Carbon: A Review of Graphene. Chem. Rev. 2010, 110, 132–145.

(5)

Kauffman, D. R.; Star, A. Carbon Nanotube Gas and Vapor Sensors. Angew. Chemie Int. Ed. 2008, 47, 6550–6570.

(6)

Meyyappan, M. Carbon Nanotube-Based Chemical Sensors. Small 2016, 12, 2118–2129.

(7)

Li, J.; Lu, Y.; Ye, Q.; Cinke, M.; Han, J.; Meyyappan, M. Carbon Nanotube Sensors for Gas and Organic Vapor Detection. Nano Lett. 2003, 3, 929–933.

(8)

Mittal, M.; Kumar, A. Carbon Nanotube (CNT) Gas Sensors for Emissions from Fossil Fuel Burning. Sens. Actuat. B 2014, 203, 349-362.

(9)

Kong, J.; Chapline, M. G.; Dai, H. Functionalized Carbon Nanotubes for Molecular Hydrogen Sensors. Adv. Mater. 2001, 13, 1384–1386.


Page 19 of 27 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Sensors

(10)

Hu, L.; Hecht, D. S.; Gruner, G. Percolation in Transparent and Conducting Carbon Nanotube Networks. Nano Lett. 2004, 4, 2513–2517.

(11)

Kong, J.; Dai, H. Full and Modulated Chemical Gating of Individual Carbon Nanotubes by Organic Amine Compounds. J. Phys. Chem. B 2001, 105, 2890–2893.

(12)

Qi, P.; Vermesh, O.; Grecu, M.; Javey, A.; Wang, Q.; Dai, H.; Peng, S.; Cho, K. J. Toward Large Arrays of Multiplex Functionalized Carbon Nanotube Sensors for Highly Sensitive and Selective Molecular Detection. Nano Lett. 2003, 3, 347–351.

(13)

Cao, Q.; Rogers, J. A. Ultrathin Films of Single-Walled Carbon Nanotubes for Electronics and Sensors: A Review of Fundamental and Applied Aspects. Adv. Mater. 2009, 21, 29– 53.

(14)

Unalan, H. E.; Fanchini, G.; Kanwal, A.; Du Pasquier, A.; Chhowalla, M. Design Criteria for Transparent Single-Wall Carbon Nanotube Thin-Film Transistors. Nano Lett. 2006, 6, 677–682.

(15)

Kaempgen, M.; Chan, C. K.; Ma, J.; Cui, Y.; Gruner, G. Printable Thin Film Supercapacitors Using Single-Walled Carbon Nanotubes. Nano Lett. 2009, 9, 1872–1876.

(16) Zhang, D.; Ryu, K.; Liu, X.; Polikarpov, E.; Ly, J.; Tompson, M. E.; Zhou, C. Transparent, Conductive, and Flexible Carbon Nanotube Films and Their Application in Organic LightEmitting Diodes. Nano Lett. 2006, 6, 1880–1886. (17)

Han, J.-W.; Kim, B.; Li, J.; Meyyappan, M. A Carbon Nanotube Based Ammonia Sensor on Cotton Textile. Appl. Phys. Lett. 2013, 102, 193104.

(18)

Han, J.-W.; Kim, B.; Li, J.; Meyyappan, M. Carbon Nanotube Based Humidity Sensor on Cellulose Paper. J. Phys. Chem. C 2012, 116, 22094–22097.


ACS Sensors 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

(19)

Page 20 of 27

Perelaer, J.; Smith, P. J.; Mager, D.; Soltman, D.; Volkman, S. K.; Subramanian, V.; Korvink, J. G.; Schubert, U. S. Printed Electronics: The Challenges Involved in Printing Devices, Interconnects, and Contacts Based on Inorganic Materials. J. Mater. Chem. 2010, 20, 8446.

(20)

Tobjörk, D.; Österbacka, R. Paper Electronics. Adv. Mater. 2011, 23, 1935–1961.

(21)

Ko, S. H.; Pan, H.; Grigoropoulos, C. P.; Luscombe, C. K.; Fréchet, J. M. J.; Poulikakos, D. All-Inkjet-Printed Flexible Electronics Fabrication on a Polymer Substrate by LowTemperature High-Resolution Selective Laser Sintering of Metal Nanoparticles. Nanotechnology 2007, 18, 345202.

(22)

Yu, W. W.; White, I. M. Inkjet Printed Surface Enhanced Raman Spectroscopy Array on Cellulose Paper. Anal. Chem. 2010, 82, 9626–9630.

(23)

Qi, Y.; Jafferis, N. T.; Lyons, K.; Lee, C. M.; Ahmad, H.; McAlpine, M. C. Piezoelectric Ribbons Printed onto Rubber for Flexible Energy Conversion. Nano Lett. 2010, 10, 524– 525.

(24)

Wang, S.; Ge, L.; Song, X.; Yu, J.; Ge, S.; Huang, J.; Zeng, F. Paper-Based Chemiluminescence ELISA: Lab-on-Paper Based on Chitosan Modified Paper Device and Wax-Screen-Printing. Biosens. Bioelectron. 2012, 31, 212–218.

(25)

Lau, P. H.; Takei, K.; Wang, C.; Ju, Y.; Kim, J.; Yu, Z.; Takahashi, T.; Cho, G.; Javey, A. Fully Printed, High Performance Carbon Nanotube Thin-Film Transistors on Flexible Substrates. Nano Lett. 2013, 13, 3864–3869.

(26)

Lien, D. H.; Kao, Z. K.; Huang, T. H.; Liao, Y. C.; Lee, S. C.; He, J. H. All-Printed Paper Memory. ACS Nano 2014, 8, 7613–7619.


Page 21 of 27 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Sensors

(27)

Hoth, C. N.; Choulis, S. A.; Schilinsky, P.; Brabec, C. J. High Photovoltaic Performance of Inkjet Printed polymer:Fullerene Blends. Adv. Mater. 2007, 19, 3973–3978.

(28)

Yang, L.; Zhang, R.; Staiculescu, D.; Wong, C. P.; Tentzeris, M. M. A Novel Conformal RFID-Enabled Module Utilizing Inkjet-Printed Antennas and Carbon Nanotubes for GasDetection Applications. IEEE Antennas Wirel. Propag. Lett. 2009, 8, 653–656.

(29) Chen, K.; Gao, W.; Emaminejad, S.; Kiriya, D.; Ota, H.; Nyein, H. Y. Y.; Takei, K.; Javey, A. Printed Carbon Nanotube Electronics and Sensor Systems. Adv. Mater. 2016, 28, 4397–4414. (30)

Loh, K. J.; Hou, T.-C.; Lynch, J. P.; Kotov, N. A. Carbon Nanotube Sensing Skins for Spatial Strain and Impact Damage Identification. J. Nondestruct. Eval. 2009, 28, 9–25.

(31)

Han, J.-W.; Kim, B.; Park, Y. C.; Meyyappan, M. Note: Two-Dimensional Resistivity Mapping Method for Characterization of Thin Films and Nanomaterials. Rev. Sci. Instrum. 2011, 82, 86117.

(32)

Kim, B.; Lu, Y.; Kim, T.; Han, J.-W.; Meyyappan, M.; Li, J. Carbon Nanotube Coated Paper Sensor for Damage Diagnosis. ACS Nano 2014, 8, 12092–12097.

(33)

Feng, X.; Irle, S.; Witek, H.; Morokuma, K.; Vidic, R.; Borguet, E. Sensitivity of Ammonia Interaction with Single-Walled Carbon Nanotube Bundles to the Presence of Defect Sites and Functionalities. J. Am. Chem. Soc. 2005, 127, 10533–10538.

(34)

Hannon, A.; Lu, Y.; Hong, H.; Li, J.; Meyyappan, M. Functionalized-Carbon Nanotube Sensor for Room Temperature Ammonia Detection. Sens. Lett. 2014, 12, 1469–1476.


ACS Sensors 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

(35)

Page 22 of 27

Gubbi, J.; Buyya, R.; Marusic, S.; Palaniswami, M. Internet of Things (IoT): A Vision, Architectural Elements, and Future Directions. Futur. Gener. Comput. Syst. 2013, 29, 1645–1660.

(36)

Grubbs, F. E. Procedures for Detecting Outlying Observations in Samples. Technometrics 1969, 11, 1–21.

(37)

Kreyszig, E. Advanced Engineering Mathematics; 4th ed.; Wiley, 1979; p 880.

(38)

Tietjen, G. L.; Moore, R. H. Some Grubbs-Type Statistics for the Detection of Several Outliers. Technometrics 1972, 14, 583–597.

(39)

Mahalanobis, P. C. On the Generalised Distance in Statistics. Proc. Natl. Inst. Sci. 1936, 49–55.

(40) Haruehanroengra, S.; Wei Wang. Analyzing Conductance of Mixed Carbon-Nanotube Bundles for Interconnect Applications. IEEE Electron Device Lett. 2007, 28, 756–759.


Page 23 of 27 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Sensors

Figure 1. (a) Schematic of a traditional sensor array. One sensor provides one output from two electrodes. (b) Schematic of the proposed multiport sensor. One sensor provides multiple outputs from combinations of multiple electrodes. (c) All-printed gas sensor consisting of a SWCNT network, sixteen silver electrodes and a polyimide film. (d) Resistance distribution extracted from the CNT ensemble using the multiport scheme. (e) Unit resistance distribution. Unit resistance is defined as the resistance of a given electrode pair used to measure that resistance divided by the distance between that pair of electrodes. (f) Real-time detection of NH3 using the printed conventional two terminal CNT gas sensor. Ammonia admittance timing diagram (top) and the resultant resistance change of the SWCNT network (bottom). (g) Sensor response according to NH3 concentration from (f) above. The error bar is the standard error of the mean from fifteen data points.


ACS Sensors

b

100

6 5

Response (%)

Unit resistance (Ω Ω /µ µ m)

a 120 80 60 40 20 0

4 3 2 1 0

Conv

SIMO

Conv

SIMO

c 12 10 Count (#)

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 24 of 27

8 6 4 2 0 0

5

10

15

20

25

30

Unit resistance (Ω Ω /µm)

Figure 2. (a) Distribution of the initial unit resistance. (b) Distribution of the sensor response. 120 two-terminal sensors and one SIMO sensor constitute each box plot. The SIMO sensor outputs 120 data points. The box plot represents 25 % and 75 % of the distribution with whiskers from minimum to maximum. The middle line of the box plot is the median of the data. (c) Unit resistance distribution (bar) and its Gaussian fitting (line) before (black color) and after (red color) gas sensing experiments of the SIMO sensor. All data in Figure 2 are from the 10 ppm NH3 experiment.


Page 25 of 27 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Sensors

Figure 3. (a) The resistance distribution in the printed gas sensor. Abnormally high resistances (blue color) from electrode #1 and #2 originated from structural defects on silver electrodes, which are shown in the inset. (b) Response versus resistance. (c) Response versus unit resistance. In Figure 3b and 3c, the experiment was performed with 10 ppm NH3 and there are two data points (red color) showing significant departure from others. From the SIMO scheme, it is possible to exclude the outliers and thereby obtain reliable data. The regression fitting blue lines in Figures 3b and 3c are drawn to guide the eye.


ACS Sensors

12 10 Count (#)

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 26 of 27

8 6 4 2 0 0

5

10

15

20

25

30

Unit resistance (Ω Ω /µm)

Figure 4. Unit resistance distribution (bar) and its Gaussian fitting from the analytical model (solid line) and experimental data (dashed line). Both modeling and experimental results are for before (black) and after (red) 10 ppm NH3 exposure.


Page 27 of 27 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

ACS Sensors

Graphical Table of Contents

For TOC only


A Single Input Multiple Output (SIMO) Variation-Tolerant Nanosensor

Recommend Documents