Graphene Nanoplatelet-Polymer Chemiresistive Sensor Arrays for the

response of all the select polymer sensors in response to an analyte provides a unique signature that ... reactive plasma.18 Although these methods ca...
1 downloads 12 Views 5MB Size
Download PDF
Article Cite This: ACS Sens. XXXX, XXX, XXX-XXX

pubs.acs.org/acssensors

Graphene Nanoplatelet-Polymer Chemiresistive Sensor Arrays for the Detection and Discrimination of Chemical Warfare Agent Simulants Michael S. Wiederoder,† Eric C. Nallon,‡,∥ Matt Weiss, Shannon K. McGraw,† Vincent P. Schnee,‡ Collin J. Bright,‡ Michael P. Polcha,‡ Randy Paffenroth,§ and Joshua R. Uzarski*,† †

Natick Soldier Research, Development and Engineering Center, United States Army, Natick, Massachusetts 01760, United States Communications-Electronics Research, Development and Engineering Center, United States Army, Fort Belvoir, Virginia 22060, United States § Department of Mathematical Sciences, Worcester Polytechnic University, Worcester, Massachusetts 01609, United States ∥ Black Cow Analytics LLC, Charlottesville, Virginia 22936, United States ‡

S Supporting Information *

ABSTRACT: A cross-reactive array of semiselective chemiresistive sensors made of polymer-graphene nanoplatelet (GNP) composite coated electrodes was examined for detection and discrimination of chemical warfare agents (CWA). The arrays employ a set of chemically diverse polymers to generate a unique response signature for multiple CWA simulants and background interferents. The developed sensors’ signal remains consistent after repeated exposures to multiple analytes for up to 5 days with a similar signal magnitude across different replicate sensors with the same polymer-GNP coating. An array of 12 sensors each coated with a different polymer−GNP mixture was exposed 100 times to a cycle of single analyte vapors consisting of 5 chemically similar CWA simulants and 8 common background interferents. The collected data was vector normalized to reduce concentration dependency, z-scored to account for baseline drift and signal-to-noise ratio, and Kalman filtered to reduce noise. The processed data was dimensionally reduced with principal component analysis and analyzed with four different machine learning algorithms to evaluate discrimination capabilities. For 5 similarly structured CWA simulants alone 100% classification accuracy was achieved. For all analytes tested 99% classification accuracy was achieved demonstrating the CWA discrimination capabilities of the developed system. The novel sensor fabrication methods and data processing techniques are attractive for development of sensor platforms for discrimination of CWA and other classes of chemical vapors. KEYWORDS: graphene, polymer sensors, chemical warfare agents, machine learning, chemical discrimination

D

diminishing their capabilities in low-resource and remote environments. In addition, existing portable sensors are often only responsive to a select number of analytes and perform poorly in the presence of background interferents.1 Thus, lowcost, low-power, miniaturized sensors capable of detection and discrimination of a broad range of CWA’s are needed for rapid deployment to mitigate risk of exposure.

etecting chemical warfare agents (CWA’s), such as tabun (GA), sarin (GB), soman (GD), and VX, to prevent exposure events is of great interest given previous and possible future use by individuals, terrorist organizations, and state sponsored militias.1 Current detection technologies include infrared and Raman spectroscopy for remote or standoff monitoring, ion mobility spectrometry, surface acoustic wave sensors, flame photometry, photoionization and electrochemical detection, and carbon nanotube gas ionization sensors for on-site testing.1 Many of these technologies require complex instrumentation, high costs, and expert personnel to operate, © XXXX American Chemical Society

Received: August 7, 2017 Accepted: October 11, 2017 Published: October 11, 2017 A

DOI: 10.1021/acssensors.7b00550 ACS Sens. XXXX, XXX, XXX−XXX

Article

ACS Sensors Table 1. Polymers Used to Create GNP-Polymer Composites for Each Sensor Material polymer

abbreviation

solvent

molecular weight (Da)

Polycaprolactone Polyepichlorohydrin Nafion 117 Polyvinyl alcohol Polyisobutylene Poly(4-vinylphenol-co-methyl methacrylate) Poly(1-vinylpyrrolidone)-graf t-(1-triacontene) Poly(vinylphosphonic acid) Polyacenaphthylene Polytetrafluoroethylene Poly(ethylene-co-vinyl acetate) Poly(4-vinylphenol)

PCL PECH Nafion PVA PIB PVPH-MMA PVPDY_gt PVPA PACN PTFE PEVA PVPH

Chloroform Chloroform Water/alcohols Water Chloroform Tetrahydrofuran Chloroform Tetrahydrofuran Chloroform Chloroform Chloroform Tetrahydrofuran

14,000 70,000 Unspecified 31,000 500,000 3,000−5,000 643 Unspecified 5,000−10,000 Unspecified 55,000 25,000

Table 2. CWA Simulant Compounds and Interferent Compounds Used to Challenge the GNP−Polymer Sensor Array with Their Respective Abbreviation, Vapor Pressure, and Concentration Used during Testing

polymers because of their low cost and response diversity.4 Each polymer possesses unique physical and chemical properties which affect the adsorption and desorption of vapor molecules.5 The collective response of all the select polymer sensors in response to an analyte provides a unique signature that aids in detection and discrimination. Technologies that have used this principle include mass sensitive polymer coated acoustic wave oscillators,6 colorimetric detection with fluorescent organic polymers7 and quantum dot composites,8 and electronic based detection with conductive polymers9 and

One method to address this problem is to use an array of semiselective chemical sensors that respond to many analytes simultaneously, creating a unique analytical signature for detection and classification of multiple analytes using one platform.2 To optimize classification the developed sensor arrays should be designed in a cross-reactive manner with high chemical diversity for differential response to a large set of analytes. The collected data can then be analyzed using machine learning algorithms for subsequent detection and discrimination.3 Many array based sensors utilize commercial B

DOI: 10.1021/acssensors.7b00550 ACS Sens. XXXX, XXX, XXX−XXX

Article

ACS Sensors

Figure 1. A custom vapor generation system was used to infuse test samples through a Teflon gas flow cell (A). Each device contains 24 interdigitated electrode arrays wire-bonded to a leadless chip carrier (B) inserted into the flow cell. The electrode dimensions (C) and an example GNP−polymer composite coated sensor (D) are shown for reference.

carbon black composites.10 For CWA’s specifically, crossreactive arrays using surface acoustic wave sensors,11 parallelplate capacitance based sensors,12 conductive polymers,13 and chemiresistive sensors made of polymer/carbon black composites14 and polymer/carbon nanotube composites15 have been demonstrated. Recent research demonstrates that single element graphene chemical vapor sensors are capable of discriminating multiple vapor compounds with similar structures.16 Graphene is of particular interest for chemical sensing applications due to its 2D structure, where every carbon atom is a surface atom, unlike other conductive carbon materials such as carbon black and multiwall carbon nanotubes.17 This provides the greatest possible surface area per unit volume so electron transport is highly sensitive to adsorbed molecular species.17 Unfortunately, single graphene sensors are limited in their classification capabilities for some similar compounds as their response signatures are not distinguishable using machine learning algorithms.16 To increase response diversity across sensor elements, graphene can be modified using covalent functionalization methods such as the introduction of radicals, nitrenes, carbenes, arynes, and reactive plasma.18 Although these methods can provide the desired diversity, the process to apply multiple functionalities on a monolithic substrate is challenging and difficult to repeat consistently causing significant variations in measured signal across sensing elements. In contrast, it is possible to overcoat a set of unique polymers on planar graphene sensors to increase response diversity for discrimination of single analytes and complex vapors.19 Despite these advantages, adoption of single layer graphene sensors is constrained by limited production volumes and relatively high costs.20 In this study we focus on polymer−graphene nanoplatelet (GNP) composite coatings to create an array of semiselective chemiresistive sensors to detect and discriminate CWA simulants and common background interferent compounds. GNP’s can be produced in large quantities, overcoming the costs of single layer graphene.20 By using polymer−GNP composites a facile approach to sensor fabrication can be utilized while taking advantage of the response diversity of polymers and the sensing potential of graphene. First multiple individual electrodes coated with the same polymer−GNP composite were evaluated to characterize repeatability of individual sensor response over a period of 5 days and across similarly coated sensors. Next, a device containing 12 unique polymer−GNP coated sensors was fabricated to generate a large data set with high response diversity. The device was

exposed a total of 1300 times to 13 analytes to demonstrate discrimination of 5 CWA simulants with similar chemical structures from one another and from 8 dissimilar background interferents. The data is processed using vector normalization to reduce concentration dependence, z-score to normalize baseline drift and represent signal-to-noise ratio, and a Kalman filter to reduce noise. The processed data was dimensionally reduced using principal component analysis (PCA) before analysis with the machine learning algorithms k-nearest neighbors, support vector classifier, random forest, or linear discrimination analysis to determine classification accuracy. For the four machine learning algorithms used 95−99% classification success for all analytes was achieved demonstrating good discrimination capacities. These examined techniques for sensor fabrication and data analysis are adaptable for development of remote sensing platforms for CWA detection.

■

EXPERIMENTAL SECTION

Materials. Graphene nanoplatelets (GNP) (Cheap Tubes Inc., Grafton, VT) were acquired with a platelet shape of 1−3 μm in diameter and an overall thickness of approximately 3−10 nm. All polymers were acquired from Sigma-Aldrich, St. Louis, MO and their names and properties are listed in Table 1. Diisopropyl methylphosphonate (DIMP) was purchased from Fisher Scientific (Hanover Park, IL). All other solvents and analytes are listed in Table 2 and were purchased from Sigma-Aldrich, St. Louis, MO. RoundUp (41% aqueous glyphosate, Scotts Company LLC, Marysville, OH), antifreeze, and diesel (Gulf Oil, Bowling Green, VA) were purchased from commercial sources. Sensor Fabrication. The fabrication of the sensor array consisted of simple drop-casting of GNP polymer−composite solutions on prefabricated interdigitated electrodes (IDE). First, Si/SiO2 substrates were cleaved into approximately 1.3 cm × 1.3 cm pieces followed by solvent cleaning in acetone, isopropanol, and DI water. Conventional lift-off photolithography was then used to pattern the Si/SiO2 substrate with the IDE design which consists of 24 identical devices, each comprising 12 electrode fingers with length, width, and separation of 750 μm, 50 μm, and 50 μm, respectively, producing a total sensor area of 0.44 mm2 (Figure 1C). The patterned device was then transferred to a Temescal BJD-1800 electron beam evaporator (Ferrotec, Livermore, CA), where a layer of Ti/Au (25 nm/300 nm) was deposited once an appropriate base pressure was achieved. Following metal deposition, the device was left in a 100% acetone bath for approximately 2 h to achieve metal lift-off. The metal patterned sensors were then rinsed with isopropanol and deionized water and dried with nitrogen. Once the metal deposition process was complete, a layer of SU-8-2005 photoresist (MicroChem, Westborough, MA) was spun onto the device at a thickness of approximately 5 μm. An additional photolithography mask design was used to provide an opening well to each sensor with length, width, and depth of 1300 μm, C

DOI: 10.1021/acssensors.7b00550 ACS Sens. XXXX, XXX, XXX−XXX

Article

ACS Sensors

Figure 2. Measured resistance of a single polymer−GNP sensor in response to toluene vapor is shown as raw signal and after baseline subtraction. Each graph contains 10 representative trials for a single sensor with a PCL or PECH coating during baseline (0−60 s), response (60−90 s), and recovery (90−270 s) phases. The maximum resistance change ΔR is also shown. 1200 μm, and 5 μm, respectively. The device was then hard baked at 150 °C for 5 min to create a permanent structure. Finally, the device was mounted in a 68-pin leadless ceramic chip carrier and each individual sensor wire bonded. The final device can be seen in Figure 1B. Commercial polymers used to create the sensor array are shown in Table 1, along with their abbreviation, respective solvent (if applicable), and molecular weight. The chemical structure of each polymer can be found in SI Figure S-1. The polymers were chosen based on previous research of polymer-coated single-layer graphene sensors.19 Each solution contained 12 mg/mL of polymer and 8 mg/ mL of GNP in the solvent specified in Table 1 for a total solids content of 20 mg/mL. All sensor solutions were sonicated after mixing for 30 min to disperse GNP and again immediately before deposition. Sensor fabrication was completed by drop-casting of 0.2 μL of each GNP polymer−composite solution into individual IDE wells using a microliter syringe (Hamilton, Reno, NV). Following coating deposition the entire device was heated to 70 °C for 10 min to remove residual solvents. An example sensor is shown in Figure 1D. Instrumentation and Analysis. The vapor generation system and gas flow cell used in these experiments is shown in Figure 1A was described previously.21 Briefly, a 68-pin breakout board was inserted into a custom machined Teflon block to provide a 0.25 mL flow channel to the sensor surface. For each trial, time vs resistance measurements were collected and divided into three sections, baseline, sample, and recovery periods set at 60, 30, and 180 s, respectively. The flow rate through the sensor was a constant 40 mL/min controlled by a digital mass flow controller (MKS Instruments, Andover, MA). Flow through the analyte vials was controlled by computer actuated solenoid valves. Vial bubblers containing analyte consisted of an inlet tube with a sparger at the end for bubble creation and an exit tube leading to the flow cell. During baseline and recovery periods, 100% nitrogen is passed across the sensor while during the response period a ratio of nitrogen and saturated vapor of the analyte (Table 2) is combined before infusion into the flow cell.

The device was exposed for 80−100 cycles with each cycle consisting of one trial for each analyte. The final data set contained 80−100 trials for each tested analyte for classification and cross validation. It is assumed a concentration equal to the saturated vapor of each analyte is infused for each trial because the fluid volume within the sealed vial remains constant and the temperature is steady during measurements. The sampling system and sensor were contained in a custom plastic enclosure outfitted with a fume extraction system and experiments were conducted at temperatures ranging between 21.8 and 22.4 °C and 63.1 and 67.3% relative humidity. The resistance of each coated electrode was measured using a data acquisition and control cube configured with two 12-channel, fully isolated resistance input boards (United Electronics Industry, Walpole, MA), creating 24 total channels with simultaneous data collection. Each channel was configured to operate in 3-wire mode, simultaneously collecting resistance measurements at a rate of approximately 15 Hz with a current of 50 μA. A custom Python script was used to read a user generated input CSV file specifying the measurement parameters for a single measurement which are baseline time, sample time, recovery time, active channels, minimum/maximum resistance, and sample flow rate. All post-processing, data analysis, and plotting was performed in Python 3.6 using Numpy, Scikit Learn, and Matplotlib libraries. Table 2 lists the CWA simulant and interferent compounds and chemical structures used to challenge the discrimination capability of the sensor array. The CWA simulants (DIMP, DEMP, DMMP, TEP, and TMP) are commonly used as chemical surrogates for G-series CWA.22 The simulants chosen have similar chemical structures, differing in functional group substitution around the central phosphoryl group. The interferents include common solvents (acetone, ethanol, hexane, and toluene) known to generate a good response for polymer based sensors,5b water, and complex mixtures (antifreeze, diesel, RoundUp) that are examples of background interferents expected in a real-world scenario. The antifreeze is a mixture of ethylene glycol (∼90−95%), water, and stabilizers; the RoundUp is a mixture of 41% glyphosphate, water, and stabilizers; and D

DOI: 10.1021/acssensors.7b00550 ACS Sens. XXXX, XXX, XXX−XXX

Article

ACS Sensors

Figure 3. Maximum relative resistance response to select analytes over time is shown for individual sensors coated in PCL PECH (A) and PECH PCL (B). Average relative resistance response of all trials to select analytes for sensors with same polymer is shown with error bars ±1 SD (C). The maximum and minimum relative maximum resistance response (ΔR/R) for PCL and PECH to each analyte (except water) is shown (D). the diesel contains diesel and trace amounts of naphthalene. These background interferents provide a realistic stressor to evaluate discrimination capabilities of the sensor arrays.

clumping after sonication), and show a stable signal response over time.19a These parameters are important to enhance discrimination capabilities and sensor repeatability. Sensor Array Response. The sensor response was investigated by measuring the resistance of the coated electrode arrays upon exposure to select analytes from Table 2. Figure 2 shows a subset of 10 trials to demonstrate the signal of sensors composed of PCL and PECH in response to toluene vapor exposure. The measured time series includes three distinct regions of baseline, response, and recovery for each trial with 80 trials total for each analyte. During the baseline period only nitrogen is flowing and the resistance remains flat. In the response section the resistance increases upon initial exposure and the rate of increase diminishes as the analyte absorption rate decreases as a function of time. Finally, during the recovery period the resistance decreases as analyte desorbs at an exponential rate while nitrogen flows over the sensor. The resistance increase results from analyte adsorption into the polymer causing swelling that increases the distance between immobilized GNP’s. In contrast, resistance decreases as analyte desorbs and the coating shrinks bringing the immobilized GNP’s closer together. This phenomena is also observed in other polymer-conductive carbon vapor sensors.14,15 As shown in Figure 2, raw resistance measurements for PCL and PECH upon repeated exposures to toluene exhibit baseline resistance drift over time. In Figure 2 the baseline drift is normalized by subtracting the average baseline resistance from all resistance measurements for each trial, demonstrating the magnitude of signal remains similar across trials. The magnitude and shape of the response for each polymer coating to each analyte is unique and varies as a function of the

■

RESULTS AND DISCUSSION Coating Formulation. The formulation of the coatings is important to optimize the discrimination capabilities of sensor array. The GNP loading of 40% w/w was chosen based on previous work with carbon black−polymer based vapor sensors that show a linear response vs analyte exposure which is important for subsequent data processing methods.10b This loading is greater than the 1−5% required for conductivity, according to the manufacturer, resulting in a highly conductive film with a consistent linear sensor response to analyte. The IV curve of the coatings (Figure S5) shows a linear slope over the entire voltage range measured, including at the 50 μA current used for this study, which indicates device resistance. Many studies with sorption based chemical sensors utilize linear solvation energy relationship (LSER) parameters as a theoretical model to optimize polymer selection for sensor arrays.23 In contrast for this study the 12 polymers employed were down selected from a library of 23 polymers reported in a previous study with polymer overcoated graphene sensors.19a The diversity of the molecular weight and chemical structure (Figure S-1) of the selected polymers generates a wide range of analyte sorption rates for each sensor to create a resistance response signature (magnitude and shape) for subsequent classification. The down selected polymers that were chosen display high signal response magnitude, represent response diversity as demonstrated by a hierarchical cluster analysis, facilitate even dispersion of GNP’s based on visual analysis (no E

DOI: 10.1021/acssensors.7b00550 ACS Sens. XXXX, XXX, XXX−XXX

Article

ACS Sensors

Figure 4. Averaged sensor response after vector normalization, z-score, and Kalman filter for each GNP−Polymer composite for 100 measurements of DMMP, DIMP, DEMP, TEP, TMP, and acetone. Note: y-axis scale is not uniform and data is dimensionless after transformations.

properties of both. The shape of the curve during both the absorption and desorption has unique properties that yield additional response diversity for discrimination.24 Sensor Repeatability. Figure 3 shows the relative maximum resistance response (ΔR/R), where the maximum resistance change (ΔR) is divided by the average baseline resistance (R), of single PECH (Figure 3A) and PCL (Figure 3B) sensors to select analytes. Even when only evaluating one component of the total sensor response, there are clear differences in measured signal between tested analytes. In addition there are differences in signal magnitude between the two polymer coatings, demonstrating a unique response signature for each analyte. For both polymers tested the relative maximum response of the sensor remains consistent after 1040 unique exposures or 80 cycles of 13 different analytes over a period of 5 days. Some differences in response magnitude across trials can be explained by minor variations in analyte concentration due to temperature changes, flow rate fluctuations, analyte exposure history, and baseline drift (Figure 2). To validate sensor reproducibility a device containing multiple electrodes with the same polymer−GNP coating was tested. Replicate coatings of the polymers PCL and PECH were each deposited on four different electrode arrays for testing (Figure 3C). For the evaluated PCL sensors the initial resistance range was 35.0−152 Ω and for PECH sensors it was 196−660 Ω due to variations in hand spotted drop cast coatings. The range of baseline drift over a 5 day test period for the PCL sensors was 4.7−11.9% and for the PECH sensors it was 1.2−7.2%. For all sensors the ΔR/R for each analyte was

statistically significant from a blank sample (nitrogen) using a Student’s t test (p < 0.05). In Figure 3C ΔR/R for each PCL sensor is statistically significant (p < 0.05) from each PECH sensor for exposure to acetone, hexane, and toluene (i.e., error bars do not overlap). The range of measurements for all PCL sensors does not overlap with any PECH sensor for acetone, antifreeze, diesel, hexane, and toluene demonstrating clear differences in response even with the wide range of initial resistances across sensors validating repeatability of sensor fabrication. For responses that overlap the shape of the response curve provides a unique signature for classification as shown in subsequent sections. Further optimization of coating fabrication repeatability and polymer−GNP ratio could enhance sensor performance for future systems using these methods. Simulant and Interferent Classification. Having demonstrated sensor stability over time and repeatability of sensor fabrication the focus was shifted to demonstrating discrimination capabilities of a larger array of polymer-GNP coated sensors. A device consisting of 12 individual sensors with unique single polymer-GNP coatings (Table 1) was used to discriminate 5 CWA simulants and 8 interferents (Table 2) tested as single analyte streams. Response diversity was achieved by using 12 different polymers to discriminate individual analytes within class (CWA simulant and interferents) and among all analytes. The resulting data set contained 100 trials for 13 total analytes for all 12 sensors with each trial containing 4000 individual resistance measurements (15 Hz for 267.67 s). F

DOI: 10.1021/acssensors.7b00550 ACS Sens. XXXX, XXX, XXX−XXX

Article

ACS Sensors

Figure 5. Visual projection of the extracted components after PCA from each trial is shown to demonstrate separation of tested CWA simulants and interferents using the same principal components. The classification accuracy of all analytes, CWA simulants, and interferents are shown for four different machine learning algorithms in the table above.

during the baseline collection phase to minimize the effects of previous analyte exposures. The calculated z-score accounts for baseline drift across all 100 trials by subtracting the average baseline resistance (μ) from each individual resistance. It also includes system noise (σ) in the denominator, resulting in a nondimensional signal-to-noise ratio for each sensor which maximizes the robustness of classification algorithms to the intrinsic scale of the signals. Noise reduction was then performed to enhance classification using a Kalman filter. The Kalman filter, sometimes known as linear quadratic estimation and commonly used in radar signal processing, estimates the signal at each time step based on the weighted history of the time series to more accurately represent the true signal.26 While a common signal processing method in many fields, its use in chemical sensing is minimal with only a few recent studies demonstrating the benefit of Kalman filtering to enhance data analysis.27 Kalman filtering contrasts with commonly employed moving average techniques by more accurately representing the data during response events (sudden decreases and increases) that are useful for classification.24 The resulting 3-D data tensor is the same as before transformation, namely, 12 (sensors) × 4000 (resistance measurements) × 1300 (100 trials for 13 analytes). The average signal by polymer after vector normalization, zscore, and Kalman filtering across 100 trials for select analytes is shown in Figure 4 with all other analytes displayed in the SI Figure S-3. By employing multiple unique polymer coatings a diverse signal response is generated creating a large data array for analyte discrimination. The magnitude, direction, and shape of the processed signal is dependent on the interaction between the polymer and the analyte. While the measured resistance for all sensors increases during response phase and decreases during recovery phase, vector normalization causes some of the curves to increase and some to decrease in magnitude during the response phase. This scaling is a result of the relative response from each of the 12 sensors. During the baseline

Preprocessing of the raw resistance measurements from each trial consisted of three steps: vector normalization, z-scoring, and Kalman filtering. This was done to reduce the effects of concentration, ensure all measurements were dimensionless, and reduce signal noise for classification respectively (Figure S2). Vector normalization of the raw data was performed on each of the 4000 samples in each trial independently. As there are 12 sensors, each sample is a point in 12-dimensional space. Vector normalization projects each of 4000 12-dimensional samples onto a unit hypersphere in 12-dimensional space by dividing each of the 12 values in the row by the Euclidean norm of that row.25 The formula for vector normalization is shown in eq 1 xi Vi = i=1 ∑12 xi2 (1) where Vi is the vector normalized value, xi is the untransformed i=1

resistance measurement, and ∑12 xi2 is the Euclidian norm. This technique reduces the concentration dependence of the data and is specifically useful for discriminative analysis for systems where all sensors have similar signal response shape to analytes.25b The reduction of concentration dependence demonstrates a more principled method to compare across analytes at different concentrations and within analyte type when concentration varies. Next, each column of the feature vector was normalized by computing the z-score (Zi) of each entry across a specified time window. This technique known as autoscaling or the z-score25b is shown in eq 2 Zi =

Vi − μ σ

(2)

where μ is the mean baseline resistance and σ is the standard deviation of the baseline resistance and Vi is the vector normalized sensor resistance at time i. The mean and standard deviation were calculated over the time interval of 45 to 55 s G

DOI: 10.1021/acssensors.7b00550 ACS Sens. XXXX, XXX, XXX−XXX

Article

ACS Sensors

100% which shows that the selected polymers provide great response diversity even for similarly structured CWA simulants. Among the 8 interferents the classification accuracy is lower than for all analytes at a range of 89.5−98.3% for the polymers selected. For the RF classifying method for all analytes a confusion matrix was generated to determine where the misclassification occurred within the data set (Figure S-4). The only misclassifications were RoundUp was misclassified as diesel and water was misclassified as antifreeze for some select data points. However, there were no CWA simulants or solvents that were misclassified with the techniques used. The addition of sensors with selectivity to those misclassified interferents would enhance discrimination capabilities if desired. The results demonstrate the discrimination capabilities (100% classification) of an array of polymer−GNP sensors for the 5 CWA simulants against themselves and against background interferents.

phase, a large response from one sensor will dominate forcing the remaining sensors to a small value as the L2 norm of each 12-sensor sample is normalized. However, once an analyte is introduced a different sensor may have a larger reaction or react more quickly than the previous dominant sensor. As a result, the more reactive sensor’s contribution to the normalized unit vector will be increased. Correspondingly, the contribution of the sensor which dominated during baseline will be decreased, reducing its normalized value below its previous value causing an inflection downward during the response phase. After baseline shift during z-scoring and the vector translates so the yvalue is zero during the baseline period and the curve is negative during the response phase. The absolute value of the maximum signal after z-scoring is indicative of the signal-tonoise ratio of each polymer−GNP sensor and is higher for solvents (acetone, ethanol, hexane, and toluene) and the CWA simulants compared to the other interferents (antifreeze, diesel, RoundUp, water). For some analytes such as water, RoundUp, and antifreeze the baseline phase appears sloped instead of flat due to incomplete desorption of the previous analyte and the yaxis scale used due to a relatively low signal response compared to other analytes. After data preprocessing the 3-D feature tensor is converted into a 2-D vector that is 1300 (trials) × 48,000 (sensors × resistance measurements). These 2-D vectors are analyzed with principal component analysis (PCA) to identify the principal components that best explain the variance in response between analytes. The PCA identified components only from the response and recovery phase and not the baseline phase to minimize the impact of previous analyte exposures on discrimination capabilities. The first component explains 61.78% of the variance, the second 27.17%, and the third 4.13% so on until more than 99% of the variance is explained by the first 10 components. Because the combined explained variance of the first 3 components is 93% only these 3 components were used for classification as three components was sufficient for accurate classification and minimized computation time. A 3-D visual representation of these principal components for every trial for each analyte is shown in Figure 5. Each CWA simulant is visually distinct from each other showing good discrimination potential using PCA. The interferent compounds are also well separated with a cluster for antifreeze, diesel, RoundUp, and water at the axes scale shown in Figure 5. Classification accuracy using the three PCA derived components for each trial was assessed with four different machine learning algorithms: k-nearest neighbors (KNN), support vector classifier (SVC), random forest (RF), and linear discriminant analysis (LDA). The four common machine learning algorithms chosen are general methods that cover many data structures and range from assuming almost nothing about the data (KNN) to assuming a lot about the data (LDA) to demonstrate robust discrimination independent of the chosen classification method. To minimize variance, the accuracy of each algorithm was determined using 10-fold cross validation with each fold having a 90/10 training/testing set split. For all CWA simulants, interferents, and all analytes the classification accuracy was 89−100% depending on the combination of analytes and algorithm considered (Figure 5). For all 13 tested analytes the classification accuracy was 95.3− 98.9% demonstrating good discrimination. Among the 5 CWA simulants themselves the classification ranged from 99.4% to

■

CONCLUSIONS An array of chemiresistive vapor sensors consisting of polymerGNP coated electrodes was shown to successfully detect and discriminate 5 CWA simulants and 8 background interferent compounds. By employing a diverse set of polymers, a high response variation with respect to signal magnitude and shape was observed. The novel sequence of data preprocessing methods including vector normalization, z-scale, and Kalman filter provided a principled method to account for concentration dependence, baseline drift, signal-to-noise ratio, and sensor noise that is applicable to many classes of chemical sensors. The processed data was dimensionally reduced using PCA and the first three principal components were used for classification with four different machine learning algorithms. The results demonstrate 100% classification for 5 similar CWA simulants against themselves and 99% classification for all analytes including CWA simulants and the tested background interferents. The fabricated sensors display a stable signal after 1300 distinct exposures over a period of 5 days to a diverse set of analytes and the response is similar for different sensors with the same coating type. The developed chemiresistive sensor array is versatile and adaptable for detection of other analyte classes not examined in this study through optimization of different polymer−GNP combinations and machine learning algorithms. Further investigation of coating formulation including GNP−polymer weight ratio, deposition method, and GNP particle shape is planned to improve detection sensitivity and discrimination capability. Overall, the demonstrated fabrication methods and data processing and analysis techniques are compatible with future development of CWA detection platforms.

■

ASSOCIATED CONTENT

S Supporting Information *

The Supporting Information is available free of charge on the ACS Publications website at DOI: 10.1021/acssensors.7b00550. Chemical structures of all polymers used, a visual example of the data preprocessing steps, the average signal response for all analytes not shown here, a confusion matrix that identifies analytes that caused classification errors, and an IV curve for the GNPpolymer coatings formulated (PDF) H

DOI: 10.1021/acssensors.7b00550 ACS Sens. XXXX, XXX, XXX−XXX

Article

ACS Sensors

■

Using Carbon Black−Chiral Polymer Composite, Chemically Sensitive Resistors. Anal. Chem. 1998, 70, 1440−1443. (c) Fonollosa, J.; Rodriguez-Lujan, I.; Shevade, A. V.; Homer, M. L.; Ryan, M. A.; Huerta, R. Human Activity Monitoring Using Gas Sensor Arrays. Sens. Actuators, B 2014, 199, 398−402. (d) Shevade, A.; Ryan, M.; Homer, M. L.; Manfreda, A.; Zhou, H.; Manatt, K. Molecular Modeling of Polymer Composite−analyte Interactions in Electronic Nose Sensors. Sens. Actuators, B 2003, 93, 84−91. (11) (a) Joo, B.; Huh, J.; Lee, D. Fabrication of polymer SAW sensor array to classify chemical warfare agents. Sens. Actuators, B 2007, 121 (1), 47−53. (b) Matatagui, D.; Martí, J.; Fernández, M. J.; Fontecha, J. L.; Gutiérrez, J.; Gràcia, I.; Cané, C.; Horrillo, M. C. Chemical warfare agents simulants detection with an optimized SAW sensor array. Sens. Actuators, B 2011, 154 (2), 199−205. (c) Alizadeh, T.; Zeynali, S. Electronic nose based on the polymer coated SAW sensors array for the warfare agent simulants classification. Sens. Actuators, B 2008, 129 (1), 412−423. (d) Matatagui, D.; Fernández, M. J.; Fontecha, J.; Santos, J. P.; Gràcia, I.; Cané, C.; Horrillo, M. C. Love-wave sensor array to detect, discriminate and classify chemical warfare agent simulants. Sens. Actuators, B 2012, 175, 173−178. (12) Mlsna, T. E.; Cemalovic, S.; Warburton, M.; Hobson, S. T.; Mlsna, D. A.; Patel, S. V. Chemicapacitive microsensors for chemical warfare agent and toxic industrial chemical detection. Sens. Actuators, B 2006, 116, 192−201. (13) (a) Collins, G.; Buckley, L. Conductive polymer-coated fabrics for chemical sensing. Synth. Met. 1996, 78, 93−101. (b) Kwon, O.; Park, S.; Lee, J.; Park, E.; Kim, T.; Park, H.-W.; You, S.; Yoon, H.; Jang, J. Multidimensional conducting polymer nanotubes for ultrasensitive chemical nerve agent sensing. Nano Lett. 2012, 12, 2797− 2802. (14) Hopkins, A. R.; Lewis, N. S. Detection and Classification Characteristics of Arrays of Carbon Black/Organic Polymer Composite Chemiresistive Vapor Detectors for the Nerve Agent Simulants Dimethylmethylphosphonate and Diisopropylmethylphosponate. Anal. Chem. 2001, 73 (5), 884−892. (15) Fennell, J. F.; Hamaguchi, H.; Yoon, B.; Swager, T. M. Chemiresistor Devices for Chemical Warfare Agent Detection Based on Polymer Wrapped Single-Walled Carbon Nanotubes. Sensors 2017, 17 (5), 982. (16) (a) Nallon, E.; Schnee, V.; Bright, C.; Polcha, M.; Li, Q. Chemical Discrimination with an Unmodified Graphene Chemical Sensor. ACS Sensors 2016, 1, 26−31. (b) Rumyantsev, S.; Liu, G.; Shur, M. S.; Potyrailo, R.; Balandin, A. Selective Gas Sensing with a Single Pristine Graphene Transistor. Nano Lett. 2012, 12, 2294−2298. (17) Yavari, F.; Koratkar, N. Graphene-Based Chemical Sensors. J. Phys. Chem. Lett. 2012, 3, 1746−1753. (18) (a) Park, J.; Yan, M. Covalent Functionalization of Graphene with Reactive Intermediates. Acc. Chem. Res. 2013, 46, 181−189. (b) Paulus, G.; Wang, Q.; Strano, M. Covalent Electron Transfer Chemistry of Graphene with Diazonium Salts. Acc. Chem. Res. 2013, 46, 160−170. (c) Liu, L.-H.; Yan, M. Functionalization of Pristine Graphene with Perfluorophenyl Azides. J. Mater. Chem. 2011, 21, 3273−3276. (d) Foley, B.; Hernández, S.; Duda, J.; Robinson, J.; Walton, S.; Hopkins, P. Modifying Surface Energy of Graphene via Plasma-Based Chemical Functionalization to Tune Thermal and Electrical Transport at Metal Interfaces. Nano Lett. 2015, 15, 4876− 4882. (19) (a) Nallon, E. Graphene-Based Chemical Vapor Sensors for Electronic Nose Applications; Dissertation, George Mason University, 2016. (b) Schnee, V. P.; Nallon, E.; Bright, C. In Discrimination of Complex Odors Using a Polymer Modified Graphene Resistor Array; 2017 ISOCS/IEEE International Symposium on Olfaction and Electronic Nose (ISOEN), Montreal, QC, Canada, Marco, S., Ed.; Montreal, QC, Canada, 2017. (20) Zurutuza, A.; Marinelli, C. Challenges and opportunities in graphene commercialization. Nat. Nanotechnol. 2014, 9, 730−734. (21) Pearce, T. Handbook of Machine Olfaction Electronic Nose Technology; Wiley-VCH: Weinheim, Germany, 2003.

AUTHOR INFORMATION

Corresponding Author

*E-mail: [email protected]. Phone: +1-(508)-2334018. ORCID

Michael S. Wiederoder: 0000-0003-2294-7142 Notes

The authors declare no competing financial interest.

■

ACKNOWLEDGMENTS Funding was provided by In-House Laboratory Independent Research (ILIR) and Section 219 Innovation Funding at the U.S. Army Natick Soldier Research Development and Engineering Center in Natick, MA. This work was supported by a collaboration with and performed at the Night Vision and Electronic Sensors Directorate in the CommunicationsElectronics Research Development and Engineering Center (CERDEC) in Ft. Belvoir, VA.

■

REFERENCES

(1) Pacsial-Ong, E. J.; Aguilar, Z. P. Chemical warfare agent detection: a review of current trends and future perspective. Front. Biosci., Scholar Ed. 2013, S5, 516−543. (2) (a) Che Harun, F. K.; Covington, J. A.; Gardner, J. W. Mimicking the biological olfactory system: a Portable electronic Mucosa. IET Nanobiotechnol. 2012, 6 (2), 45−51. (b) Albert, K. J.; Lewis, N. S.; Schauer, C. L.; Sotzing, G. A.; Stitzel, S. E.; Vaid, T. P.; Walt, D. R. Cross-Reactive Chemical Sensor Arrays. Chem. Rev. 2000, 100 (7), 2595−2626. (3) Diehl, K. L.; Anslyn, E. V. Array sensing using optical methods for detection of chemical and biological hazards. Chem. Soc. Rev. 2013, 42, 8596−8611. (4) García, J. M.; Pablos, J. L.; García, F. C.; Serna, F. Sensory Polymers for Detecting Explosives and Chemical Warfare Agents. In Industrial Applications for Intelligent Polymers and Coatings, Hosseini, M.; Makhlouf, A., Eds.; Springer: Cham, 2016; pp 553−576. (5) (a) Hierlemann, A. Fundamentals of Chemical Sensing, In Integrated Chemical Microsensor Systems in CMOS Technology; Springer, 2005; pp 9−14. (b) Abraham, M. H.; Ibrahim, A.; Zissimos, A. M. Determination of sets of solute descriptors from chromatographic measurements. Journal of Chromatography A 2004, 1037, 29−47. (6) (a) Grate, J. Acoustic Wave Microsensor Arrays for Vapor Sensing. Chem. Rev. 2000, 100, 2627−2648. (b) Ricco, A.; Crooks, R. M.; Osbourn, G. Surface Acoustic Wave Chemical Sensor Arrays: New Chemically Sensitive Interfaces Combined with Novel Cluster Analysis to Detect Volatile Organic Compounds and Mixtures. Acc. Chem. Res. 1998, 31, 289−296. (7) (a) Woodka, M.; Schnee, V.; Polcha, M. Fluorescent Polymer Sensor Array for Detection and Discrimination of Explosives in Water. Anal. Chem. 2010, 82, 9917−9924. (b) Levitsky, I.; Grate, J. W.; Krivoshlykov, S. G. Rational Design of a Nile Red/Polymer Composite Film for Fluorescence Sensing of Organophosphonate Vapors Using Hydrogen Bond Acidic Polymers. Anal. Chem. 2001, 73, 3441−3448. (8) (a) Bright, C. J.; Nallon, E. C.; Polcha, M. P.; Schnee, V. P. Quantum Dot and Polymer Composite Cross-Reactive Array for Chemical Vapor Detection. Anal. Chem. 2015, 87 (24), 12270−5. (b) Peveler, W.; Roldan, A.; Hollingsworth, N.; Porter, M.; Parkin, I. Multichannel Detection and Differentiation of Explosives with a Quantum Dot Array. ACS Nano 2016, 10, 1139−1146. (9) McQuade, D.; Pullen, A.; Swager, T. Conjugated polymer-based chemical sensors. Chem. Rev. 2000, 100, 2537−2574. (10) (a) Gao, T.; Woodka, M.; Brunschwig, B.; Lewis, N. Chemiresistors for Array-Based Vapor Sensing Using Composites of Carbon Black with Low Volatility Organic Molecules. Chem. Mater. 2006, 18, 5193−5202. (b) Severin, E.; Sanner, R. D.; Doleman, B.; Lewis, N. Differential Detection of Enantiomeric Gaseous Analytes I

DOI: 10.1021/acssensors.7b00550 ACS Sens. XXXX, XXX, XXX−XXX

Article

ACS Sensors (22) Bartelt-Hunt, S. L.; Knappe, D. R.; Barlaz, M. A. A Review of Chemical Warfare Agent Simulants for the Study of Environmental Behavior. Crit. Rev. Environ. Sci. Technol. 2008, 38 (2), 112−136. (23) Ryan, M. A.; Shevade, A. V.; Taylor, C. J.; Homer, M. L.; Blanco, M.; Stetter, J. R. Computational Methods for Sensor Material Selection; Springer: New York, NY, 2010. (24) Nallon, E. C.; Schnee, V. P.; Bright, C. J.; Polcha, M. P.; Li, Q. Discrimination Enhancement with Transient Feature Analysis of a Graphene Chemical Sensor. Anal. Chem. 2016, 88 (2), 1401−1406. (25) (a) Marco, S.; Gutierrez-Galvez, A. Signal and data processing for machine olfaction and chemical sensing: A review. IEEE Sens. J. 2012, 12 (11), 3189−3214. (b) Hierlemann, A. Higher-Order Chemical Sensing. Chem. Rev. 2008, 108, 563−613. (26) Kalman, R. E. A New Approach to Linear Filtering and Prediction Problems. J. Basic Eng. 1960, 82, 35−45. (27) (a) Weiss, M.; Wiederoder, M. S.; Paffenroth, R.; Nallon, E. C.; Bright, C. J.; Schnee, V. P.; Polcha, M. P.; McGraw, S. K.; Uzarski, J. R. Applications of the Kalman Filter to Chemical Sensors for Downstream Machine Learning; Unpublished work, Worcester Polytechnic University, 2017. (b) Althouse, M. L.; Chang, C.-I. Chemical vapor detection with a multispectral thermal imager. Opt. Eng. 1991, 30 (11), 1725−1733. (c) Wenzel, M. J.; Mensah-Brown, A.; Josse, F.; Yaz, E. E. Online drift compensation for chemical sensors using estimation theory. IEEE Sens. J. 2011, 11 (1), 225−232.

J

DOI: 10.1021/acssensors.7b00550 ACS Sens. XXXX, XXX, XXX−XXX