Efficacy of Paired Electrochemical Sensors for Measuring Ozone Concentrations

Abstract Typical low-cost electrochemical sensors for ozone (O3) are also highly responsive to nitrogen dioxide (NO2). Consequently, a single sensor’s response to O3 is indistinguishable from its response to NO2. Recently, a method for quantifying O3 concentrations became commercially available (Alphasense Ltd., Essex, UK): collocating a pair of sensors, a typical oxidative gas sensor that responds to both O3 and NO2 (model OX-B431) and a second similar sensor that filters O3 and responds only to NO2 (model NO2-B43F). By pairing the two sensors, O3 concentrations can be calculated. We calibrated samples of three NO2-B43F sensors and three OX-B431 sensors with NO2 and O3 exclusively and conducted mixture experiments over a range of 0–1.0 ppm NO2 and 0–125 ppb O3 to evaluate the ability of the paired sensors to quantify NO2 and O3 concentrations in mixture. Although the slopes of the response among our samples of three sensors of each type varied by as much as 37%, the individual response of the NO2-B43F sensors to NO2 and OX-B431 sensors to NO2 and O3 were highly linear over the concentrations studied (R2 ≥ 0.99). The NO2-B43F sensors responded minimally to O3 gas with statistically non-significant slopes of response. In mixtures of NO2 and O3, quantification of NO2 was generally accurate with overestimates up to 29%, compared to O3, which was generally underestimated by as much as 187%. We observed changes in sensor baseline over 4 days of experiments equivalent to 34 ppb O3, prompting an alternate method of calculating concentrations by baseline-correcting sensor signal. The baseline-correction method resulted in underestimates of NO2 up to 44% and decreases in the underestimation of O3 up to 107% for O3. Both methods for calculating gas concentrations progressively underestimated O3 concentrations as the ratio of NO2 signal to O3 signal increased. Our results suggest that paired NO2-B43F and OX-B431 sensors permit quantification of NO2 and O3 in mixture, but that O3 concentration estimates are less accurate and precise than those for NO2.


Introduction
Low-cost sensor networks are playing a profound role in the lower-accuracy/larger sample measurement paradigm emerging in environmental health. [1][2][3][4][5] Each node within such networks is commonly equipped with sensors that produce an electrical signal proportional to the concentration of a target gas. [6,7] Reference instruments for gas pollutants commonly utilize technologies such as optical (UV) spectroscopy (e.g., fluorescence, chemiluminescence, and absorption), but these technologies have a number of disadvantages for producing highly resolved temporospatial measurements in the environment, including high initial costs, the need for skilled operators, and designs geared towards benchtop, laboratory or regulatory applications. [2,[4][5][6] The low cost, small size, portability, and low power consumption, of gas sensors present an opportunity to overcome some of the disadvantages of reference instruments. [3,4,8,9] However, compared to reference instruments, gas sensors require thorough laboratory/field calibration by end users, have lower sensitivity/specificity, exhibit greater crosssensitivity with non-target species, are subject to larger amounts of signal baseline drift over time, and produce data of lower quality. [2][3][4][5]10] Electrochemical gas sensors are capable of quantifying a range of target gases including carbon monoxide, ozone, oxides of nitrogen, hydrogen sulfide, chlorine, and sulfur dioxide at part-per-million and -billion concentrations. [7,11] The principle of operation of electrochemical sensors relies on a chemical reaction, typically an oxidation or reduction reaction, taking place between an electrode and the target gas that produces an electrical signal proportional to the gas concentration. The electrode composition depends on the gas of interest and the chemical reaction that must take place to detect the target gas. [7,[12][13][14] The reaction creates a difference in electric potential between the sensor's working and counter electrodes, which generates an electric current that constitutes the sensor's output signal. [7,[12][13][14][15] Electrochemical sensors are typically paired with a potentiostatic circuit, which processes the sensor signal from a current to a voltage. [16] Despite lower performance compared to reference instruments, electrochemical sensors may demonstrate sufficient selectivity, accuracy, linearity, and repeatability for many applications, and in conjunction with their low power consumption, they are widely used in many settings. [3] The disadvantages of electrochemical sensors include electrolyte loss, a lifespan limited to 2 years or less (especially in low relative humidity or high concentration environments), sensitivity to electromagnetic frequencies, and crosssensitivity with interfering gases. [7,9] Although electrochemical sensors can be customized for particular applications and configurations, the need for laboratory set up, calibration and a characterization of cross-sensitivities is well recognized and inhibits their ease of use. [3,5,13] For example, many existing commercial electrochemical sensors for O 3 and NO 2 respond to both gases simultaneously, without discrimination, due to the fact that NO 2 and O 3 are both reducible at similar potentials on carbon and gold electrodes. [17] These are in effect "oxidative gas" sensors, and their response is proportional to the combined concentration of O 3 and NO 2 . This nonselectivity is particularly challenging when NO 2 and O 3 are both present, for example, in manufacturing facilities where these gases are produced by welding and other combustion processes, or in the ambient environment where both gases are associated with traffic-related air pollution. Previous studies have characterized the response of oxidative gas sensors to their target and interfering gases, [5,14] while others have attempted to differentiate sensor response Figure 1. Sensor setup. Oxidative gas and NO 2 sensors were mounted onto Individual Sensor Boards. A custom circuit board connected the sensor-ISB assembly to a microcontroller. between target and interfering gas using statistical modelling techniques such as linear regression and artificial neural networks. [15] To address the simultaneous quantification of O 3 and NO 2 concentrations, Alphasense Ltd. (Essex, UK) has proposed utilizing a pair of collocated electrochemical sensors: one that responds to NO 2 and O 3 (model OX-B431; a typical "oxidative gas" sensor) and one sensor that only responds to NO 2 (model NO2-B43F). The NO2-B43F sensor is equipped with a manganese dioxide (MnO 2 ) filter which catalyzes O 3 into oxygen (O 2 ), thereby preventing sensor response to O 3 in the environment. The response of the oxidative gas sensor to O 3 is calculated by subtracting the response to NO 2 . This paired sensor method of quantifying O 3 was previously introduced and the differential response (in nA) of one pair of sensors was demonstrated for concentrations of 1 ppm O 3 , 1 ppm NO 2 and a mixture of 1 ppm O 3 and 1 ppm NO 2 ; however, the authors did not evaluate how well the paired sensor method quantified NO 2 and O 3 concentrations. [17] We have previously reported on an earlier generation of the OX-B431 sensor (model OX-B421) only, and its response to NO 2 and O 3 exclusively. [18] This study examines the ability of the paired sensors to distinguish NO 2 and O 3 concentrations simultaneously. We also assessed the bias and precision of the paired sensor method for quantifying NO 2 and O 3 concentrations at atmospherically relevant concentrations. The response of the sensors to NO 2 and O 3 gas individually was used to create calibration curves and those calibration curves were used to calculate the concentration of each gas in mixtures over a range of NO 2 and O 3 concentrations. We outline the practical aspects of setting up, calibrating and using the paired sensor method for quantifying NO 2 and O 3 . We also present two separate methods to calculate gas concentrations from sensor signal and have described the utility of each method, depending on the end user's application.

Sensor configuration
We mounted three pairs of new NO 2 (model NO2-B43F, Alphasense Ltd., Essex, UK) and oxidative gas (model OX-B431, Alphasense Ltd., Essex, UK) sensors onto Individual Sensor Boards (000-0ISB-02) produced by the same manufacturer ( Figure 1). These assemblies were connected to a microcontroller (model Seeeduino Cloud, Seeed Technology Inc., Shenzhen, China) through a customized circuit board, two sensors to each board. The working electrode and reference electrode signals were each amplified by a factor of 2 with signal amplifiers (model MCP6002, Microchip Technology Inc., Chandler, AZ) and fed into a 10-bit analog-to-digital convertor on the microcontroller. The 5-volt power for the device was smoothed with a 5-volt LM7805 linear regulator to reduce signal noise. Voltage outputs from each sensor were calculated by taking the difference between the working and reference electrode values and were transmitted over a serial channel approximately every 2 sec to a computer.

Experimental setup
The sensors were opened and installed on their assemblies for 50 days prior to the start of experiments and data collection, during which time we ran sensors in ambient laboratory air, conducted preliminary test, and made hardware and software adjustments on the sensor boards. Over the course of 4 days, we We exposed sensors to NO 2 and O 3 in a 22 cm x 15 cm x 24 cm (7.92 L) acrylic chamber ( Figure 2). A small vent in the chamber allowed gas to escape and the chamber to operate at a slightly positive pressure with respect to the room. A digital thermometer/hygrometer (model Hygrochron iButton, Maxim Integrated Inc., San Jose, CA) monitored the chamber's temperature and relative humidity. Both NO 2 and O 3 concentrations in the chamber were measured with highly specific reference instruments (NO 2 : model 42c, Thermo Environmental Instruments Inc., Franklin, MA; O 3 : model Personal Ozone Monitor, "POM," 2B Technologies Inc., Boulder, CO). The Thermo 42c chemiluminescent analyzer is a designated federal reference method (FRM) for NO 2 (with a precision of ± 0.4 ppb) and the POM UV absorption instrument is a designated federal equivalent method (FEM) for O 3 (with a precision of ± 2 ppb). Both instruments were calibrated before use, and experimental conditions were within the instruments' operating ranges. Nitrogen dioxide was supplied to the chamber with a dynamic gas calibrator (model 146i, Thermo Environmental Instruments Inc., Franklin, MA) by diluting high-concentration (500 ppm) NO 2 from a tank with zero-air. Ozone was supplied to the chamber by an O 3 generator (model 146c, Thermo Environmental Instruments Inc., Franklin, MA). Airflow from both the dynamic gas calibrator and ozone generator were supplied to the chamber at 5.0 L/min during all experiments (including at gas concentrations equal to zero), and concentrations of NO 2 and O 3 were adjusted to achieve the target gas concentrations in the chamber. We maintained temperature between 24.6-27.8 C and relative humidity (RH) between 36.3-51.8% by circulating chamber air through a bubbler filled with water at a flowrate of 25 L/min with a vacuum pump (MEDO VP0435A, Roselle, IL).
Although the manufacturer provides calibration slopes and intercepts for each sensor, we first conducted experiments to develop sensor-specific calibration curves for O 3 and NO 2 exclusively with the sensor setup and configuration used in this study. We also performed experiments to assess how well the sensor pairs quantified concentrations of NO 2 and O 3 in mixtures with one another. For each target concentration of 0.1, 0.25, 0.5, and 1.0 ppm NO 2 , the chamber was first flushed with zero air for 10 min during which the sensor baseline response was recorded. Then, the steady-state NO 2 concentration for each experiment was established, for 10 min before the introduction of O 3 . Then O 3 was introduced at varying levels to maintain concentrations of approximately 65, 125, 30, 95, and 0 ppb for 10 min each without altering the NO 2 concentration. This procedure is exemplified by the time series of the reference instrument measurements, shown in Figure SA1 of the Supplemental Materials during the 0.1 ppm NO 2 experiment (panel a) and the 0.25 ppm NO 2 experiment (panel b). Similarly, time series for one of the sensor pairs' (sensor pair 1) responses for the same experiments are shown in Figure SA2 of the Supplemental Materials during the 0.1 ppm NO 2 experiment (panel a) and the 0.25 ppm NO 2 experiment (panel b). A 10-min average of the sensors' 2sec voltage output from each experimental condition was used to establish the response of the sensor to the target concentration(s).

Calculating nitrogen dioxide and ozone concentrations with low-cost sensors
The OX-B431 sensors respond to both NO 2 and O 3 , and the NO2-B43F sensors respond only to NO 2 , although the sensitivities of both types of sensors to each of the gases differs. Consequently, separate calibration curves for the OX-B431 sensors to NO 2 , OX-B431 sensors to O 3 and NO2-B43F sensors to NO 2 were first determined. Here, calibration curves for each sensor were developed by applying least-squares linear regression to sensor signal in response to NO 2 and O 3 exclusively. To measure O 3 in mixture with NO 2 , the NO2-B43F and OX-B431 sensors must be collocated and the NO 2 contribution to the OX-B431 sensor response is subtracted by first calculating the NO 2 concentration with the NO2-B43F sensor. To test this procedure, we conducted experiments mixing NO 2 and O 3 , and measuring both gases with electrochemical sensor pairs. We then calculated NO 2 and O 3 concentrations for experimental conditions using two different methods in response to an observed change in sensor baseline values over the course of experiments.
Additionally, we estimated the limit of detection for each sensor from data collected during the first 10 min of each experiment when the sensors were exposed to zero air according to (Equation (1)): where LOD g is a sensor's limit of detection for NO 2 or O 3 gas, rm V is the standard deviation of each sensor's baseline response calculated from 1-min averages of sensor baseline in millivolts (mV), and m g is the calibration slope or sensitivity of the sensor for either NO 2 or O 3 .

Method 1: Applying calibration slope and intercept
In the first method to calculate gas concentration, subsequently referred to as "Method 1," the slopes and intercepts of each of the sensors determined by our calibration experiments with a single gas were applied to the sensor response. Method 1 would reasonably be employed by typical end users for a low-cost sensor network deployment, where frequent recalibration or baseline correction is not practical. The calibration curve derived for the NO2-B43F (Equation (2)) was rearranged to solve for the NO 2 concentration (Equation (3)): where mV NO2-B43F is the response of the NO2-B43F sensor in mV, [NO 2 ] is the concentration of NO 2 , m NO2-B43F is the slope of the calibration curve of the NO2-B43F sensor, and b NO2-B43F is the intercept of the calibration curve for the NO2-B43F sensor. In Method 1, we approached the signal from the OX-B431 sensor in a similar fashion as the NO2-B43F sensor, including terms for the OX-B431 sensor calibration slope to NO 2 and O 3 and the calibration intercept for O 3 (Equation (4)): where mV OX-B431 is the response of the OX-B431 sensor in mV, m OX-B431,NO2 is the slope of the calibration curve of the OX-B431 sensor to  (5)): Method 2: Baseline-correcting sensor response and applying calibration slope In an alternate analysis, prompted by an observed change in sensor baseline values and subsequently referred to as "Method 2," we applied only the slopes determined in the calibration experiments to the baseline-corrected sensor response. This approach, while not necessarily practical in the context of a sensor network deployment, may be useful for other applications where sensor baseline can be more frequently corrected for, and offers insight into the sources of error for these paired electrochemical sensors. The baseline response for each sensor was recorded at a concentration of 0 ppm NO 2 Table SA3 of the Supplemental Materials. The relationship between sensors' response and gas concentrations were therefore: where mV NO2-B43F,baseline-corrected and mV OX-B431,baseline-corrected are baseline-corrected signals from the NO2-B43F and OX-B431 sensors, respectively, and all other terms remain the same. Method 2 provides a strategy to manage transient changes in sensor baseline, which is comparable to the calibration intercept, but assumes that sensor calibration slope is constant for the dataset.

Bias and precision of NO 2 and O 3 concentrations estimated by electrochemical sensors
To quantify the accuracy of sensor concentration estimates, measurement error was taken as the percent bias of each NO 2 and O 3 concentration estimate for each sensor pair and the average of the three sensor pairs compared to the reference instruments. Bias was calculated according to: where Sensor is the concentration estimated from the electrochemical sensors and Reference is the concentration according to the reference instruments. We estimated the concentration of NO 2 and O 3 using each sensor pair and then took the mean concentration of the three sensor pairs. This mean concentration estimate was evaluated against the reference instruments to calculate the mean bias of NO 2 and O 3 concentration estimates for each experimental condition. Bias was compared to guidance values from NIOSH and the EPA for direct reading monitors and air sensors. NIOSH specifies that percent bias should be within ±10%, [19] whereas the EPA recommends that bias be within 20-50%, depending on the application area, including Education and Information (<50%), Hotspot Identification and Characterization (<30%) Supplemental Monitoring (<20%), and Personal exposure (<30%). [20] Similarly, the mean absolute percent error (MAPE) was calculated to summarize the measurement error of NO 2 and O 3 concentration estimates of mixture experiments for each sensor pair and the average of three sensor pairs, according to: where n is the number of NO 2 or O 3 experimental concentrations studied, Sensor is the NO 2 or O 3 concentration measured by the electrochemical sensors, and Reference is the O 3 or NO 2 concentration measured by the reference instrument. Percent bias and MAPE were not calculated for the lowest concentrations of NO 2 and O 3 where the target concentration was zero. Here, bias provides a measure of error at each experimental condition within mixture experiments, and MAPE summarizes the bias observed across the range of conditions for each mixture experiment.
To characterize the precision of gas concentration estimates, we calculated the coefficient of variation (CV) by dividing the standard deviation of the three sensor concentration estimates, r conc. , by the absolute value of the mean concentration estimate of the sensors, l, and expressed it as a percent: Taking the absolute value of the mean of sensor concentration estimates allowed for the calculation of precision when sensor signals produced concentration estimates that were negative. Negative estimates of gas concentration are an artifact of processing sensor signal to gas concentration, particularly at low concentrations. A higher CV indicates more variability and more imprecision in concentration estimates. As with bias, we compared calculated CVs to guidance values from the EPA which recommends precision between 20 and 50% depending on the application. [20] NIOSH does not provide a recommended value for precision. All data were analyzed with MATLAB R2017a (Natick, MA).

Sensor response to NO 2 or O 3 exclusively
The results of the linear regression on voltage output from each sensor with respect to NO 2 and O 3 exclusively are shown in Table 1. Among the three NO2-B43F sensors, the mean of the slopes of the response observed to NO 2 was 283 mV/ppm with a standard deviation of 27 mV/ppm (9% of the mean). Individual sensors' response to NO 2 was highly statistically significant (p < 0.00001) and linear (R 2 ¼ 1.00). In contrast, the slopes of the response of the NO2-B43F sensors to O 3 were low (mean slope ¼ 13 mV/ppm), not statistically significant (p ! 0.165), and nonlinear (R 2 0.53), consistent with the expectation that these sensors do not respond to O 3 .
The mean slope of the response of the OX-B431 sensors to NO 2 was 382 mV/ppm and to O 3 was 431 mV/ppm. The standard deviations of the mean slopes were 56 mV/ppm NO 2 (15% of the mean) and 81 mV/ppm for O 3 (19% of the mean). Individual OX-B431sensor response to NO 2 was highly statistically significant (p < 0.00001) and linear (R 2 ¼ 1.00). Similarly, the individual OX-B431 sensor response to O 3 was highly statistically significant (p < 0.0005) and linear (R 2 ¼ 0.99). Of note, the mean OX-B431 sensor response to NO 2 gas was 1.35 times greater than the NO2-B43F sensors (382 vs. 283 mV/ppm) and the mean OX-B431 sensor response to O 3 gas was 1.13 times greater than to NO 2 gas on a concentration basis (431 mV/ppm vs. 382 mV/ppm). The mean LOD for NO 2 was 5 ppb for the NO2-B43F sensors and 3 ppb for the OX-B431 sensor. The LOD for O 3 was 4 ppb for the OX-B431 sensor.

Calculating NO 2 and O 3 concentrations by applying calibration slope and intercept: Method 1
The mean bias of NO 2 and O 3 concentration estimates for Method 1 are shown in Figure 3, Panels (a) and (c). The mean bias of NO 2 and O 3 concentration estimates for each experimental condition was calculated using the mean gas concentration estimate of the three sensor pairs. The mean bias points shown in Figure 3 are colored based on this value. The contour plot was created by linear interpolation of the overlying mean bias points to describe the bias between experimental conditions. Bias is an indicator of accuracy and values closer to zero represent closer agreement of the electrochemical sensors to the reference instrument. For Method 1, the mean bias for of NO 2 ranged from -8 to 29%, with bias of a larger magnitude observed at higher NO 2 concentrations ( Figure  3a Table SA4 of the Supplemental Materials. The MAPE is interpreted here as a summary measure of experimental biases and shown in Table 3. For NO 2 concentration estimates, the overall MAPE was equal to 8%, less than NIOSH's bias criterion of ± 10% (Table 3). For O 3 concentration estimates, the overall MAPE was equal to 71% (Table 3) and was greater than the largest EPA criterion for bias of ± 50%.
The mean CV in NO 2 and O 3 concentrations estimated with the three sensor pairs at each experimental condition for Method 1 is shown in Figure 3, Panels (b) and (d). We observed generally uniform CV in concentration estimates between 1 and 7% (median ¼ 5%) for NO 2 (Figure 3b), but strongly increasing CV in ozone ranging from 6 to 146% (median ¼ 44%) for O 3 that increased with increasing NO 2 and decreasing O 3 concentrations (Figure 3d). The NO 2 concentrations estimated via Method 1 met the most stringent EPA guidelines for precision (<20%), whereas 6 out of 16 (38%) of the O 3 concentration estimates met the same guideline.
Calculating NO 2 and O 3 concentrations by applying calibration slope to baseline-corrected sensor response: Method 2 We observed decreases in sensor response to zero air (0 ppm NO 2 and 0 ppb O 3 ) ranging from 12-22 mV over the 4 days of experiments that were unrelated to temperature or relative humidity differences (Table 2). These baseline voltages decreased over the 4 days by as much as 107% for the OX-B431 and 92% for the NO2-B43F sensors comparing the first day of experiments to the last day. Of particular note was the observed change in sensor baseline compared to the magnitude of the sensor response to target gas. For example, the OX-B431 sensor with the largest absolute change in sensor baseline among the OX-B431 sensors had a change of 17 mV, corresponding to   For Method 2, we observed higher levels of bias for NO 2 concentration estimates but lower levels of bias for O 3 concentration estimates compared to Method 1 (Figure 4). For Method 2, the mean bias for NO 2 ranged from -44 to 17%, with the magnitude of the bias higher at lower NO 2 concentrations (Figure 4a). For O 3 concentration estimates, the mean bias for Method 2 was between -107 and 5% and displayed a pattern of bias similar to Method 1 with higher bias observed at lower O 3 and higher NO 2 concentrations (Figure 4c). For Method 2, 10 out of 20 (50%) NO 2 concentration estimates and 2 out of 16 (13%) O 3 concentration estimates met the NIOSH criterion of bias ± 10%. The bias of individual sensor pair concentration estimates of NO 2 and O 3 for Method 2 are presented in Table SA5 of the Supplemental Materials. The overall MAPE of NO 2 concentration estimates was equal to 14% and of O 3 concentration estimates was 30% (Table 3), which were greater than the NIOSH criterion of 10%, but for NO 2 within the most stringent limits suggested by the EPA for supplemental monitoring activities (±20%) and hotspot identification and characterization (±30%). CV for Method 2 was calculated the same way as in Method 1, and we observed comparable CV in concentration estimates between 2 and 6% (median ¼ 5%) for NO 2 ( Figure  4b). For O 3 , CV was between 3 and 1753% (median ¼ 20%) and increased with increasing NO 2 and decreasing O 3 concentrations (Figure 4d). NO 2 concentrations estimated via Method 2 met the most stringent EPA guidelines for precision (<20%), whereas 9 out of 16 (56%) of the O 3 concentration estimates met the same guideline.
For Methods 1 and 2 overall, the bias and CV of NO 2 concentration estimates were less than for O 3 (Figures 3 and 4), indicating that concentration estimates of NO 2 were more accurate and precise compared to those for O 3 . We observed a larger overall error of O 3 concentration estimates for Method 1 compared to Method 2 (MAPE ¼ 71 vs. 30%) and a larger overall error of NO 2 concentration estimates for Method 2 compared to Method 1 (MAPE ¼ 14 vs. 8%) ( Table 3). Method 1 produced NO 2 concentration estimates with MAPE between 3 and 12% with generally smaller error at low NO 2 concentrations, and Method 2 produced concentration estimates between 4 and 34% with smaller error at high NO 2 concentrations (Table 3). In addition, while the CV of concentration estimates with both methods was comparable for NO 2 , for O 3, the CV increased as the concentration of NO 2 increased and O 3 decreased and was generally smaller for Method 2 except for one outlier. (Figures 3d and 4d).

Discussion
Our calibration experiments demonstrated that the NO2-B43F sensors had a highly linear response to NO 2 , and that the OX-B431 sensors had a highly linear response to NO 2 and O 3 , comparable with previous studies. [21] The variability we observed in the calibration slope among the three sensors of each type (Table 1) is consistent with the variability in the sensor-specific calibration slopes provided by the manufacturer. In our sample of three sensors of each type, the standard deviation of the mean of the calibration slopes among the three NO2-B43F sensors to NO 2 was equal to 27 mV/ppm, and for the OX-B431 sensors to O 3 was equal to 81 mV/ppm, and to NO 2 was equal to 56 mV/ppm. These results suggest it is appropriate to use sensor-specific calibration curves rather than a common curve for each type of sensor, although our sample size was small.
In the case of NO 2 and O 3 , the strategy of paired sensors works, although with decreasing accuracy and precision when the signal from NO 2 obscures the signal from O 3 . On an individual sensor pair level, the accuracy of NO 2 and O 3 concentration estimates varied across the three sensor pairs studied, with one of the sensor pairs out-performing the other two according to the MAPE. This also suggests that sensor pairs should be calibrated and tested in the laboratory prior to deployment in the field to identify sensor pairs with unacceptable levels of measurement error.
We observed the specificity of the NO2-B43F sensors, and although they responded slightly to increasing concentrations of O 3 gas, on average, their response to NO 2 gas was over 20 times greater than the response to O 3 . Furthermore, least-squares regression of the response among the three NO2-B43F sensors to O 3 produced p-values 0. 16 p 0.63, indicating that O 3 concentration was not a significant predictor of NO2-B43F sensor response. These results provide further evidence that the MnO 2 filter on the NO2-B43F sensor is effective at excluding O 3 from the sensor and are consistent with prior studies. [17] Over the course of the 4 days in which we carried out experiments, we observed changes in sensor baseline that affected the quantification of NO 2 and O 3 and the measurement error associated with each sensor concentration estimate. We investigated whether temperature or relative humidity could explain this drift, but ultimately the reasons for the changes in baseline are not known. Compared to the NO2-B43F sensor, a unit of the OX-B431 sensor signal is associated with a greater concentration of gas, making concentration estimates more vulnerable to errors given a change in sensor baseline. This is an especially acute problem given that O 3 concentrations of interest are often less than 100 ppb, compared to NO 2 concentrations which are often greater than 100 ppb. Our strategy to correct for changes in sensor baseline resulted in a differential change in error associated with NO 2 and O 3 concentration estimates between the two methods. For Method 1 compared to Method 2, we observed a higher degree of accuracy in NO 2 concentration estimates (MAPE ¼ 8 vs. 14%), but worse accuracy for O 3 concentration estimates (MAPE ¼ 71 vs. 30%). In this laboratory study were easily able to accommodate changes in sensor baseline with Method 2, however, a comparable methodology in the field on the day-to-day timescale may be impractical.
One reason there is more error estimating O 3 concentration in a mixture with NO 2 with paired electrochemical sensors compared to estimating NO 2 concentration with a single sensor is that the error associated with two sensors is propagated through the subtraction procedure. Additionally, if using a common sensor calibration slope for sensors of the same type, the mean response of the OX-B431 sensors to O 3 may be smaller in magnitude than the variability of the OX-B431sensors' response to NO 2 . For example, we observed the range of response across the 3 OX-B431 sensors exposed to 0.5 ppm NO 2 equal to 57 mV which is equivalent to the mean OX-B431 sensor response to 132 ppb O 3 . Another challenge is that the changes in sensor baseline are large relative compared to the response of the sensor to O 3 at typical ambient and occupational concentrations. Here we observed maximum changes of 17 mV with the OX-B431 sensor associated with 34 ppb O 3 and 22 mV with the NO2-B43F associated with 0.070 ppm NO 2 .
These dynamics make accurate O 3 concentration estimates in a mixture with NO 2 challenging with pairs of electrochemical sensors, especially if the end-user chooses to use a common calibration curve for sensors of each type. For these reasons, when measuring O 3 concentrations with paired electrochemical sensors, we caution against using single calibration curves for each sensor type without previously examining individual sensor response to target gas. This conclusion may not be consistent with previous evaluations of an earlier-generation oxidative gas sensor (model: OX-B421, Alphasense Ltd., Essex, UK) where a single calibration curve for NO 2 and a single calibration curve for O 3 adequately characterized the response of a sample of three sensors. [18] Our evaluation of these sensors occurred over a stable and controlled range of temperature and RH for each experimental condition (mean temperature ± SD: 27 ± 1 C, mean RH ± SD: 39 ± 5% RH) to reduce their influence on sensor response. Purposefully characterizing sensor response under a larger range of temperature and RH or applying temperature and RH correction factors from the manufacturer would be particularly important for deployment in environments where temperature and RH are highly variable.
A limitation of this study is that we did not examine other gases that interfere with the quantification of NO 2 and O 3 with electrochemical sensors, such as nitrogen monoxide (NO) and carbon dioxide (CO 2 ). In a study using previous generations of the sensors used here at ambient concentrations of CO 2 , NO, NO 2 , and O 3 , the impact observed on O 3 concentration estimates by the OX-B421 sensor was 20.6% for NO and 365.8% for CO 2 , whereas the impact on NO 2 concentration estimates by the NO2-B4 sensor was -20.6% for NO and 118.9% for CO 2 . [5] These interfering gases co-occur with O 3 and NO 2 in ambient and occupational environments and would decrease the accuracy of concentration estimates or may completely obscure target gas signals if present in high concentrations. The present study with NO 2 and O 3 provides evidence that the strategy of filtering out cross sensitive gases and deploying collocated sensors could be successfully developed for other target gases with known interferents depending on the required accuracy of the application.

Conclusions
We evaluated a method for measuring NO 2 and O 3 in mixture using paired electrochemical sensors: one sensor that responds to O 3 and NO 2 (OX-B431) and another that responds to NO 2 only (NO2-B43F). We observed the strategy works over a range of concentrations and mixtures of the two gases, but the precision and accuracy of O 3 concentration estimates declined as NO 2 concentration increased. We observed substantial variability in the concentration estimates of O 3 in a sample of three sensor pairs. Over the course of the 4 days of experiments, we also observed a change in sensor baseline, complicating the calculation of O 3 , and prompting an alternate method of baseline-correcting sensor signal to calculate concentration. Although the paired sensor method has potential to improve the specificity of O 3 concentration estimates compared to a single oxidative gas sensor, concentrations of NO 2 and O 3 where the ratio of NO 2 signal to O 3 signal is large may still challenge their performance, performance among sensor pairs is variable, sensor baseline voltage is subject to drift, and the cost to measure O 3 effectively doubles. Increases in target gas specificity will ameliorate a major drawback and improve the utility of electrochemical sensors and has the potential to provide higher-quality data for environmental and occupational sensor networks.