A low-cost monitor for simultaneous measurement of fine particulate matter and aerosol optical depth – Part 3: Automation and design improvements

Atmospheric particulate matter smaller than 2.5 μm in diameter (PM2.5) has a negative impact on public health, the environment, and Earth’s climate. Consequently, a need exists for accurate, distributed measurements of surface-level PM2.5 concentrations at a global scale. Existing PM2.5 measurement infrastructure provides broad PM2.5 sampling coverage but does not adequately characterize community-level air pollution at high temporal resolution. This motivates the development of low-cost sensors which can be more practically deployed in spatial and temporal configurations currently lacking proper characterization. Wendt et al. (2019) described the development and validation of a first-generation device for low-cost measurement of AOD and PM2.5: the Aerosol Mass and Optical Depth (AMODv1) sampler. Ford et al. (2019) describe a citizen-science field deployment of the AMODv1 device. In this paper, we present an updated version of the AMOD, known as AMODv2, featuring design improvements and extended validation to address the limitations of the AMODv1 work. The AMODv2 measures AOD and PM2.5 at 20 min time intervals. The sampler includes a motorized Sun tracking system alongside a set of four optically filtered photodiodes for semicontinuous, multiwavelength (current version at 440, 500, 675, and 870 nm) AOD sampling. Also included are a Plantower PMS5003 sensor for time-resolved optical PM2.5 measurements and a pump/cyclone system for time-integrated gravimetric filter measurements of particle mass and composition. AMODv2 samples are configured using a smartphone application, and sample data are made available via data streaming to a companion website (https: //csu-ceams.com/, last access: 16 July 2021). We present the results of a 9 d AOD validation campaign where AMODv2 units were co-located with an AERONET (Aerosol Robotics Network) instrument as the reference method at AOD levels ranging from 0.02± 0.01 to 1.59± 0.01. We observed close agreement between AMODv2s and the reference instrument with mean absolute errors of 0.04, 0.06, 0.03, and 0.03 AOD units at 440, 500, 675, and 870 nm, respectively. We derived empirical relationships relating the reference AOD level to AMODv2 instrument error and found that the mean absolute error in the AMODv2 deviated by less than 0.01 AOD units between clear days and elevated-AOD days and across all wavelengths. We identified bias from individual units, particularly due to calibration drift, as the primary source of error between AMODv2s and reference units. In a test of 15month calibration stability performed on 16 AMOD units, we Published by Copernicus Publications on behalf of the European Geosciences Union. 6024 E. A. Wendt et al.: Automation and design improvements observed median changes to calibration constant values of −7.14 %, −9.64 %, −0.75 %, and −2.80 % at 440, 500, 675, and 870 nm, respectively. We propose annual recalibration to mitigate potential errors from calibration drift. We conducted a trial deployment to assess the reliability and mechanical robustness of AMODv2 units. We found that 75 % of attempted samples were successfully completed in rooftop laboratory testing. We identify several failure modes in the laboratory testing and describe design changes that we have since implemented to reduce failures. We demonstrate that the AMODv2 is an accurate, stable, and low-cost platform for air pollution measurement. We describe how the AMODv2 can be implemented in spatial citizen-science networks where referencegrade sensors are economically impractical and low-cost sensors lack accuracy and stability.

integrated (gravimetric, filter-based) and real-time PM2.5. Here, we describe the design and validation of the AMODv2.

Instrument design
We designed the AMODv2 to sample integrated gravimetric PM2.5 mass concentration, real-time PM2.5 130 mass concentration, and AOD simultaneously. One intended application is large-scale sampling campaigns with the AMODv2 instruments operated by volunteers with little to no background in aerosol or atmospheric science (Ford et al., 2019). Thus, we prioritized a design that is low-cost, mechanically robust, portable, automated, and userfriendly. We provide images of AMODv2 hardware in Fig. 1, highlighting key internal and external components.
The measurement process is fully automated using a solar tracking system (Section 2.3), reducing the potential for misalignment due to user error. Movement in the zenithal plane is achieved using a custom turret module embedded in the interior of the AMODv2 enclosure (Fig. 1a). The module was designed in SolidWorks ® (ANSYS,Inc.,145 Canonsburg, PA, USA) and built using multi-jet fusion printing. The module houses a custom printed circuit board containing the solar tracking sensors and the filtered photodiodes. Light enters the turret through four, 4 mm https://doi.org/10.5194/amt-2021-73 Preprint. Discussion started: 1 April 2021 c Author(s) 2021. CC BY 4.0 License.

AOD measurement and solar tracking 195
The AMODv2 applies the Beer-Lambert-Bouguer law to calculate AOD ( ). This relationship, expressed in terms of measurable parameters, is as follows: where m is the unitless air mass factor, which accounts for the increased air mass that light passes through as the sun approaches the horizon, R is the Earth-sun distance in astronomical units (AU), V is the signal produced by the light 200 detector in volt, τR accounts for Rayleigh scattering by air molecules, p is the pressure at the sensor in Pa, λ is the sensor wavelength in m, τO3 accounts for ozone absorption, and V0 is the extraterrestrial constant in volt, which is the sensor signal if measured at top-of-atmosphere and is determined via calibration. AOD values at 440 nm, 500 nm, 675 nm, and 870 nm are calculated using Eq. (2). The Earth-Sun distance, R, is computed directly from GPS data and the solar positioning algorithm. V is the signal produced by the photodiode and V0 is accessed from on-chip 205 memory. The relative optical air mass factor is computed as a function of solar zenith angle (θ) as follows (Young, The contribution of total optical depth from Rayleigh scattering, τR, is calculated based on wavelength and ambient pressure measured by an ambient pressure sensor mounted on the circuit board with the AOD sensors 210 (Bosch Sensortec BMP 280, Kusterdingen, Germany) (Bodhaine et al., 1999). Ozone concentration is estimated using an empirical model based on time of year and location, and converted to τO3 using wavelength-specific ozone absorption coefficients (Griggs, 1968;Van Heuklon, 1979). With all parameters known, Eq. (2) is applied to calculate AOD.
We implemented automatic solar tracking capabilities using a suite of low cost sensors and a multi-stage calculated solar zenith angle is then compared with the pitch of the AOD turret relative to horizontal. The turret 220 stepper motor rotates the turret in the direction of the sun until the elevation angle of the AOD turret is approximately equal to that of the sun. The base motor rotates counterclockwise in order to achieve approximate azimuth alignment. After every 10 degrees of azimuthal rotation, the total signal of the sun-tracking quadrant photodiode is compared with an empirical threshold. If the threshold is exceeded, the AMODv2 enters closed-loop tracking. If the threshold is not exceeded on the first revolution, the AMODv2 executes a second revolution before 225 ending the search protocol.
In the closed-loop tracking stage, the rotation of the motors is controlled using the zenithal and azimuthal error signals produced by the quadrant photodiode. The quadrant photodiode is mounted in a diamond orientation, with two quadrants forming a vertical axis, and two forming a horizontal axis. The vertical error signal is the ratio of the top and bottom quadrants and the horizontal error signal is the ratio of the right and left quadrants. The stepper 230 motors rotate independently until each error signal is reduced within an experimentally determined threshold. The motors then lock in place while an AOD measurement is recorded. The AMODv2 measures AOD as triplet sets.
Between each measurement, both motors disengage for 30 seconds to conserve power. After 30 seconds, the AMODv2 executes the tracking algorithm and records an AOD measurement. This process is repeated until the triplet set is completed or until 3 minutes have elapsed since the initial measurement request was made by the 235 processor.
Real-time quality control is performed on each measurement triplet. Empty or incomplete triplets are flagged and assigned an error code. Completed triplets are screened for cloud contamination using the AERONET triplet cloud screening algorithm (Smirnov et al., 2000;Giles et al., 2019). The algorithm takes the maximum deviation of any two measurements within a triplet, and applies thresholds to mark triplets as clear or cloud-240 contaminated (Smirnov et al., 2000;Giles et al., 2019). Large deviations of AOD within a triplet are more likely due to cloud contamination than changes in aerosol loading (Smirnov et al., 2000;Giles et al., 2019). Measurements identified as cloud-contaminated are flagged with a unique error code.

AOD calibration procedure
The extraterrestrial constants for all AMODv2s were evaluated via calibration relative to AERONET sun 245 photometers (Cimel CE318, Paris, France) (Holben et al., 1998). AERONET monitors report AOD at 340 nm, 380 nm, 440 nm, 500 nm, 675 nm, 870 nm, 1020 nm, and 1640 nm (Holben et al., 1998). We selected the four AMODv2 AOD wavelengths in part for direct comparison with AERONET monitors. We conducted calibrations at the MAXAR-FUTON site in Fort Lupton, Colorado (40.036 N,104.885 W) between November 2019 and February 2020. AMODv2 units were co-located within 50 m of the AERONET monitor and sampled for 2 to 3 hours at a rate 250 of one sample every 2.5 to 3 minutes. AMODv2 and AERONET level 1.0 measurements concurrent within 60 seconds of each other were included in the calibration data set (Holben et al., 1998). For each set of concurrent measurements, we calculated the extraterrestrial constant by applying Eq. (2) solved for V0, where V was the raw voltage reported by the AMODv2, and τa was the AOD reported by the AERONET monitor. The AMODv2 calibration constants were the average value of V0 for a given instrument and wavelength. 255 https://doi.org/10.5194/amt-2021-73 Preprint. Discussion started: 1 April 2021 c Author(s) 2021. CC BY 4.0 License.

User operation and measurement procedure
We designed the AMODv2 to be operated by individuals without a background in aerosol sampling. We developed a standard procedure that is detailed in a user manual provided as supplementary material. After the initial setup, the AMODv2 requires no operator inputs for the duration of the sample. A flow chart outlining the manual and automatic steps to perform an AMODv2 measurement is provided in Fig. 2. 260 Parallel processes are executed pseudo-simultaneously using a real-time operating system.

265
To initiate a sample, the operator needs only an AMODv2 monitor, a cartridge loaded with a pre-weighed filter, and a smartphone with the AMODv2 control application installed ("CEAMS"; available on the Apple App https://doi.org/10.5194/amt-2021-73 Preprint. Discussion started: 1 April 2021 c Author(s) 2021. CC BY 4.0 License.
is included as a supplement to this work. After executing an initialization routine by selecting "Scan for Device", the operator may connect to their device via Bluetooth TM using the mobile application. The operator can select a 270 wireless network and input the proper credentials to connect the AMODv2 to the internet. The application then prompts the operator to scan the QR code on the back of the filter cartridge to link the filter with the upcoming sample in the data log. After the cartridge is manually loaded into the compartment behind the inlet (Fig. 1b), the AMODv2 should be placed on a flat surface with an unobstructed view of the sun. The operator then starts the sample from the mobile application. After an initial data push, the sample begins at the next 20 minute mark (e.g. 275 12:00, 12:20, or 12:40). The AMODv2 begins sampling air through the inlet at 1 L min -1 and continues to do so for the remainder of the 120-hour sampling period. Real-time PM2.5 and AOD measurements are initiated at each 20 minute mark from the start of the sample. The PM2.5 reported at each 20 minute interval is the average of measurements taken every 10 seconds over a period of 3 minutes. If the sun is less than 10 degrees above the horizon, the motors do not activate and the solar tracking algorithm is not executed. After each AOD and PM2.5 280 measurement is completed, data are uploaded to the affiliated website (csu-ceams.com), where real-time visualizations of AOD and PM2.5 are available. Data reported to the website are accessible with a map-based user interface. Quality-control data are available to research staff via a private administrator portal. A snapshot example of the website is provided in Fig. S7. At the conclusion of a sample, the operator removes the filter cartridge. Upon receipt of the filters, the CEAMS team stored the filters in the refrigerator until mailed to minimize loss of volatile 285 compounds. Complete data files can be downloaded from the website or accessed via a MicroSD card.

Sample deployment and AOD validation studies
We conducted a sample deployment of 10 AMOD units during a wildfire smoke event in Fort Collins, Colorado in October of 2021. We configured the units to sample for approximately 60 hours. The 10 units were colocated and sampled simultaneously. We collected and analysed real-time PM2.5 mass concentrations, AOD, PM2.5 290 to AOD ratio, meteorological data, and quality control data.
We assessed precision and bias of AMODv2 AOD sensors relative to an AERONET monitor at the NEON-CVALLA site in Longmont, Colorado (40.160 N,105.167 W) between June 2020 and December 2020. We colocated our instruments within 50 m on seven separate days with varying atmospheric conditions (e.g. wildfire smoke and clean air) using a total of 14 unique AMODv2s. Each test consisted of 2 to 4 hours of sampling at a rate 295 of one sample every 2.5 to 3 minutes. AMODv2 and AERONET measurements concurrent within 120 seconds were included in the validation data set. The accuracy of AMODv2 AOD measurements was assessed via Deming regression analysis. 3 Results and discussion 305

Sample AMODv2 deployment
The AMODv2 is capable of accurate, real-time, and low-cost measurement of AOD and PM2.5. Here we present results from the sample deployment of 10 units. In Fig. 3, we provide real-time AOD at 500 nm, real-time PM2.5, and the corresponding PM2.5 to AOD ratios. µg m -3 ). This was followed by increases on 17 October 2020 to severe levels (AOD up to 1.5 and PM2.5 up to 300 µg m -3 ) as wildfire smoke swept over the city in the afternoon and gradually subsided over the course of 18 October 2020. We observed reductions in PM2.5:AOD (<10) as ground level PM2.5 decreased to moderate and mild levels (<20 µg m -3 ), while the AOD remained elevated (>0.5) due to the presence of lofted smoke. We then noted the continuation of the trend at ground level with the further reduction of ground-level PM2.5 on 19 October 2020 (5 to 325 15 µg m -3 ). Cloud cover prevented additional AOD measurements on 19 th October, which was automatically screened for using the cloud screening algorithm. The meteorological data was also consistent with cloud cover with lower temperatures and elevated relative humidity reported on that day (Fig. S8).
Data from the sample deployment were accessed from our companion website (csu-ceams.com) in real time. With AOD, PM2.5 and PM2.5:AOD reported every 20 minutes throughout the sample to the website, we could 330 assess the progress of wildfire smoke in Fort Collins remotely in real time. This was not possible with AMODv1, which lacked wireless transmission capabilities. In terms of scalability, the AMODv2 was relatively easy to deploy and maintain owing to its compact and weatherproof design, coupled with its automated measurement protocols. In the sample test, we were able to quickly prepare and deploy units in response to wildfire activity.
We leveraged the data accessibility features of AMODv2 for real-time quality control of incoming sample 335 data. We monitored sample flow rate and total sampled volume to detect potential errors with the gravimetric sample collection. We monitored battery temperature to detect potential overheating of the unit, allowing proper intervention (e.g. temporarily moving the unit into shade) before the instrument reaches a shutoff threshold. We used battery voltage, battery state of charge, and current draw data to identify units unlikely to complete the intended sample duration. Current draw data was also used to identify when the tracking motors were engaged, indicating an 340 attempted AOD measurement at the expected time. Wireless signal strength data were used to identify units with relatively poor connection and move them into areas with better signal. In the sample deployment detailed here, no interventions based on quality control data were warranted. However, in general, these data can be used to remotely identify and address malfunctioning units mid-sample. This feature represents a substantial improvement compared with AMODv1, which provided no sample quality control data in real time, requiring manual data acquisition (via 345 micro SD card) and unit inspection following a failed sample.
We observed close AOD agreement between AMODv2 and AERONET monitors. Correlation plots are provided in Fig. 4 (n = 426 paired measurements per wavelength). Summary statistics are provided for each 355 wavelength in Table 1. A plot of AMODv2 vs. AERONET co-located measurements is provided in Fig. 4.  The bounds defined by Eqs. (4) through (7) contain 85% of the co-located measurement pairs. A logarithmic plot illustrating how the error bounds scale with increasing AOD is provided in Fig. 5. to AERONET for all wavelengths. Existing error between AMODv2 and AERONET measurements was explained primarily by the constant term. 390 AMODv2 bias relative to AERONET was primarily dependent on the specific unit, rather than systemic design uncertainty. A mean-difference plot colored by AMODv2 unit ID is provided in Fig. 6. effective strategies to control emissions and exposures. As a ground monitor measuring both PM2.5 and AOD, AMODv2 measurements can be used to constrain the uncertainty of satellite-based AOD retrievals, as well as the reliability of the conversion between an AOD value, to a ground-level PM2.5 estimate. The spatial resolution of satellite-based PM2.5 retrievals is on the order of kilometers (Salomonson et al., 1989;Diner et al., 1998Diner et al., , 2018Zoogman et al., 2017;Wei et al., 2019). The temporal resolution is often on the order of a full day to a week, 460 depending on the satellite path and period (Salomonson et al., 1989;Diner et al., 1998;Zoogman et al., 2017;Diner et al., 2018;Wei et al., 2019). This leaves variation at lower spatial and temporal scales unaccounted for in the absence of ground monitors. Multiple AMODv2s deployed in a spatially dense network may be used to evaluate the degree to which air pollution varies within the spatial resolution limits of satellite-based retrievals. The relatively high measurement frequency of the AMODv2 ensures a low temporal discrepancy (<10 minutes) between an 465 AMODv2 measurement and a satellite overpass. The remaining AMODv2 data can be used to assess deviations in air pollution between satellite-based retrievals. Information gained through these analyses can be used toward improving the usefulness of satellite measurements for determining surface air quality, upon which many impact assessments and mitigation strategies rely.
Our sample testing has revealed several areas of potential improvement for the AMODv2 design. 470 Individual unit bias was the primary source of error relative to AERONET and relative to other AMODv2 units. We believe instances of individual unit bias highlights potential limitations of the existing calibration protocol. In this https://doi.org/10.5194/amt-2021-73 Preprint.