Performance of microAethalometers: Real-world Field Intercomparisons from Multiple Mobile Measurement Campaigns in Different Atmospheric Environments

Small aethalometers are frequently used to measure equivalent black carbon (eBC) mass concentrations in the context of personal exposure and air pollution mapping through mobile measurements (MM). The most widely used is the microAethalometer (AE51). Its performance in the laboratory and field is well documented, however, there is not sufficient data in the context of its performance in different environments. In this investigation, we present the characterization of the performance of the AE51 through field unit-to-unit intercomparisons (IC), and against a reference absorption photometer from three MM campaigns conducted in drastically different environments. Five IC parameters were considered: i) study area, ii) location of IC, iii) time of day, iv) duration of IC, and v) correction for the filter-loading effect. We can conclude that it is crucial where and how long the IC have been performed in terms of the correlation between the mobile and reference instruments. Better correlations (R > 0.8, slope = 0.8) are achieved for IC performed in rural, and background areas for more than 10 minutes. In locations with more homogenous atmosphere, the correction of the loading effect improved the correlation between the mobile and reference instruments. In addition, a newer microAethalometer model (MA200) was characterized in the field under extreme cold conditions and correlated against another MA200 (R > 0.8, slope ≈ 1.0), AE51(R > 0.9, slope ≈ 0.9), and a stationary Aethalometer (AE33) across all wavelengths (R > 0.8, slope ≈ 0.7). For MA200, the loading effect was more pronounced, especially at the lower wavelengths, hence the correction of the loading effect is essential to improve the correlation against the AE33. The MA200 and AE51 proved to be robust and dependable portable instruments for MM applications. Real-world quality assurance of these instruments should be performed through field IC against reference instruments with longer durations in areas of slowly changing eBC concentration.

of the performance of the AE51 through field unit-to-unit intercomparisons (IC), and against a 23 reference absorption photometer from three MM campaigns conducted in drastically different 24 environments. Five IC parameters were considered: i) study area, ii) location of IC, iii) time of 25 day, iv) duration of IC, and v) correction for the filter-loading effect. We can conclude that it is 26 crucial where and how long the IC have been performed in terms of the correlation between the 27 mobile and reference instruments. Better correlations (R 2 > 0.8, slope = 0.8) are achieved for IC 28 performed in rural, and background areas for more than 10 minutes. In locations with more 29 homogenous atmosphere, the correction of the loading effect improved the correlation between 30 the mobile and reference instruments. In addition, a newer microAethalometer model (MA200) 31 was characterized in the field under extreme cold conditions and correlated against another 32 MA200 (R 2 > 0.8, slope ≈ 1.0), AE51(R 2 > 0.9, slope ≈ 0.9), and a stationary Aethalometer 33 (AE33) across all wavelengths (R 2 > 0.8, slope ≈ 0.7). For MA200, the loading effect was more 34 pronounced, especially at the lower wavelengths, hence the correction of the loading effect is 35 essential to improve the correlation against the AE33. The MA200 and AE51 proved to be robust 36 and dependable portable instruments for MM applications. Real-world quality assurance of these 37 instruments should be performed through field IC against reference instruments with longer 38 durations in areas of slowly changing eBC concentration. The microAethalometer (microAeth® AE51 model, Aethlabs, San Francisco, CA) is 54 currently the most widely used portable absorption photometer for measurements of BC mass 55 concentrations aboard mobile platforms. The AE51 measures the attenuation of light (880 nm) 56 through a particle-loaded filter (T60 Teflon coated glass fiber) and converts this to an equivalent 57 black carbon (eBC; (Petzold et al., 2013)) mass concentration using a fixed mass attenuation 58 coefficient (MAC). The time resolution can be set from 300 seconds down to 1 second. It is also 59 small enough to fit in a pocket, making it extremely portable. This instrument is the most 60 characterized portable instrument for eBC measurements in terms of filter loading effect (Cheng 61 and Lin, 2013; Good et al., 2017), and sensitivity to sudden changes in relative humidity, 62 temperature (Cai et al., 2013), and vibration (Apte et al., 2011). To investigate the performance 63 of any instrument, it is often compared against "reference" instruments with operating principles 64 considered as standard method. While there is no standard method to measure BC, there are three

A C C E P T E D M A N U S C R I P T
3 potential candidate methods to measure aerosol absorption ab-initio. These methods are capable 66 of measuring the absorption of particles suspended in the air, rather than collected on a filter. The 67 first method measures extinction and scattering, and calculates the absorption as the difference 68 between them. This method works for single-scattering albedo values below about 0.8. The 69 second is the photoacoustic method which measures the pressure waves generated by modulated 70 absorption of light by aerosols, subsequent heating and change in the density of the air. Usually, a 71 resonant cavity is employed to amplify the signal. This resonance needs to be tracked and is 72 sensitive to changes in ambient conditions. The signal is also dependent on the losses to latent 73 heat as the coating of the particles changes phase. The third is the photothermal interferometry 74 which employs a similar heating of the sample, measuring the change in the refractive index 75 following the change in density. All methods require drying of the sample. Filter photometers 76 are commonly used in the field. Light from an LED source passes through a particle-loaded filter, 77 and is detected by a photodiode. The amount of light attenuated by the light-absorbing particles 78 trapped in the filter is proportional to the concentration of these particles. The light absorption 79 coefficient of the particles is converted to mass concentration of light absorbing carbon by 80 dividing it with the MAC. The calculation assumes the filter properties in the derivation of the 81 absorption coefficient introducing uncertainties into the reported parameters. This method 82 provides "mass equivalent black carbon" or "eBC" as recommended by Petzold et al., 2013. 9   away from the backpacks and with ~ 3 m height difference between their inlets. For IC at the  176   urban background site, there were two instances: IC against the reference instruments on the 4th  177 floor of a building (as mentioned above), and IC against the aerosol container when it was moved 178 near the same building for a week. The inlet of the of building site was ~55 m from the inlets of 179 the backpacks during IC periods. IC periods at the urban street and urban background were done 180 within one run. Alas et al., 2018 have demonstrated that the difference in concentration between 181 the building site and the aerosol container when they were in the same location was not 182 significant. The mean eBC mass concentrations (standard deviation) at the building site was 6.9 183 (4.8) µg m -³, while at the ground site with the aerosol container was 7.6 (4.9) µg m -³ showing that 184 particles in this location, specifically eBC particles, were spatially homogenous. 185

A C C E P T E D M A N U S C R I P T
For the IC performed at the urban street canyon site, the same aerosol container was used 186 but was mounted on 1-m cement blocks. Therefore, the vertical distance between the inlets of the 187 aerosol container and the backpacks were ~4 m.

A C C E P T E D M A N U S C R I P T
11 within one run (one filter). For the IC at the rural village, the backpacks' inlets were ~8 m 212 horizontally and ~ 2m vertically away from the inlets of the fixed station. For the IC at the rural 213 background, the horizontal distance between the inlets was ~2.5 m and the vertical distance was 214 ~3 m. 215

Data processing 216
The loading effect in filter photometers is a bias, which reduces the apparent 217 concentrations relative to the ambient ones. The apparent reduction depends on the loading of the 218 spot. The filter-loading effect (FLE) is a non-linearity due to the saturation of the attenuation 219 (ATN) as the amount of the sample on the filter in the photometer continually increasesthe 220 eBC mass should depend only on the change of attenuation in time, but due to saturation, an 221 photoacoustic extinctionmeter (PAX) was used as a reference instrument, which, being not filter-228 based, is not susceptible to FLE. However, as this instrument was not used in any of the

A C C E P T E D M A N U S C R I P T
12 campaigns in this study, the eBC mass concentrations measured by the AE51 and MA200 were 230 instead compared against the AE33 which has a real-time FLE correction, and the MAAP which, 231 compared to the AE51, is less susceptible to the FLE (Petzold et al., 2005). 232 For the Manila and Rome datasets, three approaches to assess the FLE were performed:

A C C E P T E D M A N U S C R I P T
13 plots only for eBC values below the 95 th percentile of ATN. If the fit has a negative slope, the 248 eBC is decreasing with increasing ATN, hence, there is a loading effect. Normally, the loading 249 parameter to correct the AE51 raw concentrations can be derived from the slope and intercept of 250 the regression line. 251 The deviation (ATN) approach follows that of Masey et al., 2020 to assess FLE during IC periods. 252 The raw eBC mass concentrations measured by the AE51 during the IC periods were taken and 253 aggregated to 1-min averages. Two statistical parameters were used to investigate deviation of 254 the measurements between the AE51 and reference instruments as a function of the ATN of the 255 AE51: the ratio (AE51/reference) and the difference (AE51reference). Similar to the first 256 approach, the slope of the linear fit indicates the FLE.

Unit-to-unit comparability of AE51 295
For campaigns in Rome and Loški Potok, two aerosol backpacks with identical instrumentation 296 were used to explore the unit-to-unit variability of two AE51 units in real-world MM. The 297 models used were the AE51 S5 and AE51 S6, where the former is an older model. 298 Figure 1 shows the correlation analyses (RMA) between the two models during both campaigns. 299 It must be noted that exactly the same models were used for both campaigns. The correlation of 300 the two units is slightly lower in the Rome campaign compared to the Loški Potok campaign. In

A C C E P T E D M A N U S C R I P T
16 study area in Rome, which was in an urban area with higher variabilities of sources. Nonetheless, 303 the unit-to-unit variability of the two AE51 units during MM is low at around 5% at 10-second 304 time resolution. 305

Intercomparability of mobile devices to reference instruments in different environments 306
In this section, we explore how the AE51 performed in different environments using data from 307 three different campaigns in comparison to rack-mounted, widely used absorption photometers 308 (MAAP 5012 and AE33), which are considered as reference instruments. The AE51 309 measurements were aggregated to 1-minute averages to compare against the reference 310 instruments with 1-minute time resolution. The entire IC dataset for each campaign was used for 311 this correlation analysis and the results are shown in Figure  at an urban background site, one at a street side, and one at a street canyon. For Loški Potok, one 320 was done at a rural background region (up a hill) and one at a street side of a rural village. We 321 performed the correlation analysis again, this time not only as a function of the study area, but 322 also of where the IC was performed (Fig. 3). One can see now that, for Manila ( Fig. 3(a)), the 323 low correlation (R 2 < 0.5, and slope = 0.75 and 1.5) between the AE51 and MAAP was due to the 324

A C C E P T E D M A N U S C R I P T
17 results are the same as in Fig. 2(b). Fig. 3(c) also shows good correlation between the mobile and 327 reference instruments, indicating that the area of Loški Potok has a homogenous distribution of 328 eBC particles. The poor correlation at street side IC in Manila is due to the higher variabilities 329 that arise from passing of vehicles, turbulence, and other local sources as well as the vertical and 330 horizontal distance between the inlets of the aerosol backpack and aerosol container. Although, 331 this does not mean that the AE51 do not perform well in areas with high spatial variabilities, it is 332 simply difficult to conduct an IC in such locations due to rapidly fluctuating concentrations. This 333 could be improved by connecting the backpack to the same inlet as the reference instrument, but 334 this would disrupt the MM. Therefore, to harmonize mobile instruments during mobile 335 measurement campaigns, IC done at atmospherically homogenous areas work best. 336 337 The filter-loading effect 338 All datasets were analyzed for filter-loading effect. From the three datasets, the measurements 339 from both AE51 and MA200 of the Loški Potok campaign were corrected for the filter-loading 340 effect. For the Manila and Rome datasets, the results of the three assessment approaches are 341 presented and discussed here. From the first approach, the BC(ATN) plots showed a dependency 342 on the route ( Fig. A1 and A3), indicating that a single loading parameter cannot be derived, 343 because the area being studied has a very inhomogeneous atmosphere and specific areas with 344 different sources have to be analyzed separately. Unfortunately, there isn't sufficiently large data 345 set to derive an empirical k (Fig. A2). The deviation (ATN) approach showed similar results (

A C C E P T E D M A N U S C R I P T
18 slopes as well (0.0028 and 0.0039). However, this approach may not be suitable for this study: 351 the IC periods (co-located AE51and reference measurements) were performed in the middle of 352 the runthis means, that we would have only a fraction of the ATN range to analyze. To be able 353 to apply results from this approach, we would need eBC mass concentration data that is evenly 354 distributed over the whole ATN range, otherwise, it would be misleading to use a loading 355 parameter derived from this and apply it to the whole measurement route in urban areas.  Figure A5 shows that correcting for the FLE with the given k did not 359 significantly improve the eBC mass concentrations of the AE51 (3-8% increase). 360 Also, owing to the inhomogeneity of the study area, correcting the whole dataset with a single 361 loading parameter may cause an over/underestimation in specific parts of the route. Unlike in 362 Loški Potok, which is a rural area, the eBC levels vary widely and rapidly in urban environments 363 due to micrometeorology and high spatial variation of sources and their strengths. These 364 variations are also greater than the possible error caused by the loading effect. Hence, correcting 365 for it will not lead to any significant improvement of the AE51 eBC measurements. 366 Results of the three approaches suggest that there are no significant detectable FLE in the Manila 367 and Rome datasets. Dedicated experiments are necessary to develop methods that would lead to 368 derivation of a loading parameter appropriate for data obtained from MM in urban areas. 369 In this section, the impact of the FLE correction on the Loški Potok data is discussed. Figure 4  370 shows the IC between the AE51 and the AE33 in the two lC locations in Loški Potok for both 371 uncorrected and corrected AE51 eBC data. The correlations between uncorrected eBC measured 372 by microAethalometers (AE51_S5 and AE51_S6) and the reference instrument AE33 were good

A C C E P T E D M A N U S C R I P T
19 uncorrected eBC measurements for both mobile instruments was higher at the urban background 375 station (0.88) than for the village station. This can be explained by taking a closer look into the 376 course of each run, since every single run started at the rural background, where the attenuation 377 of the filter was low, continuing towards the village, where filter attenuation was already high. 378 This leads to increased loading effect and consequently lower slope, when comparing to the 379 reference AE33 in the village: 0.81 and 0.84 for the AE51 S5 and AE51 S6, respectively (Fig. 4

Time of IC 388
The time of the day when the IC was performed was also investigated (Fig. 5) to determine if the 389 intercomparability of the AE51 and reference instruments is affected by the variability of the 390 meteorological conditions and sources within a day. The time of IC was segregated to morning, 391 afternoon, and evening as proxy to variations in incoming solar radiation, temperature, and height 392 of boundary layer. Fig. 5 shows that there is no obvious dependence of the intercomparability to 393 the time of IC. In all IC locations, the AE51s were able to capture the eBC mass concentrations 394 regardless of the variabilities within the day. 395

A C C E P T E D M A N U S C R I P T
20 Lastly, the duration of IC was investigated. Only data from rural background, rural village, 398 background and urban background regions were used for this analysis which is shown in Figure.  399 6. The duration of IC increases from Fig. 6 (a) to (d) and it shows that longer IC durations lead to 400 better correlation and harmonization of the mobile with the reference instruments. Longer 401 durations provided more time for the mobile instruments to adjust to its surroundings as they are 402 not in the same inlet as the reference instruments nor are they on the same height from the ground. 403 Therefore, IC should be done for more than 10 minutes in atmospherically homogenous areas to 404 achieve better harmonization between the mobile and reference instruments.  Table 3. The data from AE33 were already corrected online. 431 The IC of measurements obtained with the MA200 and the reference AE33 for five different 432 wavelengths (UV, Blue, Green, Red and IR) at two stations in Loški Potok, showed a more 433 pronounced filter-loading effect in MA200 instruments as compared to the AE51. Moreover, 434 higher loading effect is seen for the lower wavelengths (Fig. 9). 435 Correcting for the filter-loading effect in MA200 makes a significant improvement of the 436 correlation against AE33 for all wavelengths, particularly for lower wavelengths. Less loading 437 effect was observed for the rural background station, due to low filter attenuation at the beginning 438 of each run. The slope between datasets for the UV wavelength increased after compensation 439 from >0.29 to >0.78, with an increase of the R2 from >0.80 to >0.93. For the IR wavelength, the 440 improvement of correlation with corrected data was the smallest, yet with an important increase 441 of the slope at the rural village site from >0.74 to >0.87. The loading parameter k_MA200 differs 442 from the one featured in other Aethalometer instruments due to a completely different filter 443 materialit is not fibrous but rather a membrane. Loading effect for Teflon coated glass fiber

A C C E P T E D M A N U S C R I P T
22 MA200 instruments is evaluated. As observed during the Loški Potok campaign, MA200 446 instrument experiences much stronger loading effect than the AE51. Therefore, loading 447 compensation should be applied to the raw data especially with high filter loading and when 448 AAE is calculated from the multi-wavelength data, since stronger loading effect in low 449 wavelengths leads to biased values of AAE. 450 451 CONCLUSIONS 452 453 microAethalometers, despite being widely-used for mobile measurements of eBC mass 454 concentrations, have hardly been assessed in real-world environments. In this study, two models 455 (AE51 and MA200) were assessed to determine how well they perform in the field during mobile 456 measurements when compared against a reference absorption photometer. 457 Data from three mobile measurement campaigns were used in this study: a highly urbanized 458 megacity during the summer (Manila, Philippines), a touristic but urbanized city in winter (Rome,  459 Italy), and a rural village in winter (Loški Potok, Slovenia). The assessment was in terms of its 460 comparability against another unit of the same model, and a reference absorption photometer. 461 The AE51 showed a unit-to-unit variability of 5% in urban areas, and lower in rural areas. This 462 was also reflected by the intercomparison (IC) against the reference instruments, where R 2 are 463 higher and slopes closer to unity for IC's done at the rural background, rural village, background, 464 and urban background locations than at urbanstreet and urban street canyon. The 465 intercomparability of the AE51 to the reference instruments showed dependence on the location

A C C E P T E D M A N U S C R I P T
23 of the IC, filter-loading effect correction, and duration of IC, but not on the time of day when the 467 IC was done. This implies that the AE51 performs well in different environments and can capture 468 the variabilities of the eBC mass concentrations within the day which are caused by the varying 469 strength of sources and meteorological conditions. Also, for mobile measurements, 470 harmonization of the AE51 with the reference instruments should be done in an atmospherically 471 homogenous environment at longer duration (10-30 minutes) where the spatial variabilities are 472 much lesser than at the street side. 473 In addition, the field performance of a newer microAethalometer with 5-wavelengths (MA200) 474 was also assessed in terms of its intercomparability against another MA200, the AE51, and a 7-475 wavelength Aethalometer. The MA200 has low unit-to-unit variability (~2%) across all 476 wavelengths as determined at the rural sites. The variability is greater at the rural village, 477 especially at lower wavelengths (UV = 15-22%, blue = 12-18%, green = 11-15%, red = 0-8%, 478 and red = 0-3%). The MA200s (880 nm channel) showed good agreement with the AE51s. In the 479 environments with similar conditions as in Loški Potok, where biomass burning is an important 480 source of eBC, correcting the raw data for filter-loading effect is of exceptional importance for 481 reliable data interpretation. In the study in Loški Potok, Slovenia, compensation parameter k was 482 determined for each wavelength and applied with the post-processing method (Virkkula et al., 483

A C C E P T E D M A N U S C R I P T
24 can introduce systematic errors in the data, which can be avoided by determining highly changing 485 k values using the online algorithms (e. g. Drinovec  Here, the details of the BC(ATN) approach performed to assess the FLE for the Manila and 521 Rome datasets are presented. For the other two approaches, the information is provided in the 522 Methods section of the main manuscript. They require the same data preparation as below. 523

A C C E P T E D M A N U S C R I P T
26 when the new filter is inserted at the start of each run, we deducted the initial value for each run 526 (ATN at t=0, ATN0) from the ATN values during the measurements: the corrected ATN 527 (ATNcorr) was calculated as the difference between the ATN measured at the next point in time 528 (ATNt=1) and ATN0. ATN does not start at 0 when the filter is inserted due to ununiform 529 illumination of the sample and reference spots in the filter photometers. Then, the BC mass percentile of the ATN. If the fit featured a negative slope, and BC is decreasing with increasing 538 ATN, we interpreted this as the presence of the loading effect, that is the dependence of BC on 539 ATN rather than just on the change of ATN in time. Normally, the loading parameter to correct 540 the AE51 raw concentrations can be derived from the slope and intercept of the regression line. 541

Investigation of the filter-loading effect on AE51 data from Manila and Rome campaigns 542
In this section, the loading effect on the AE51 measurements from the Manila and Rome 543 campaign was investigated following three approaches presented in the manuscript. The 544 prerequisites for applying the filter loading effect correction using a loading parameter derived 545 from a single period of analysis are having sufficient measurement data points and homogenous 546 sources of particles. 547 For the BC(ATN) approach, again, the whole datasets (not just the data points during the 548 intercomparison (IC) period) were used for a complete loading effect assessment. The results are 549 shown in Figure A1. The blue and red dots represent the median and mean eBC mass 550 concentration per ATN bin, respectively, while the error bars represent the standard deviation. To 551 detect if there is a loading effect, a linear fit was performed over the whole ATN range and the 552 ratio of the slope and the intercept represents the loading parameter k. If the slope of the fit is 553 negative and its absolute value is greater than 0, then there is a loading effect. 554 However, Fig. A1 shows a positive slope which could be a statistical artifact (Drinovec et al., 555 2015). Hence, to determine an appropriate range of ATN for fitting, the frequency distribution of 556 the number of measurements per ATN bin was plotted and are shown in Figure A2. From here, 557 the ATN range for fitting was adjusted to include only everything below the 95th percentile of 558 the ATN as the frequency of the measurement decreases towards higher ATN. 559 The BC(ATN) was plotted again, this time fitting within the range of ATN reflecting 0-95th 560 percentile of the data ( Figure A3). For the Taft and Rome routes, the slopes are still positive. 561 Refitting with ATN range down to < 85th percentile still resulted to positive slopes (not shown). 562 For the Katipunan route, fitting the median values for an ATN range covering up to 95th and up 563 to 85th percentile of the data gave negative slopes which could indicate a loading effect. 564 However, from these plots, it can be observed that the dependency of BC on ATN seem to be 565 affected by the route itself. both, indicating a possible FLE. However, it must be noted that the number of datapoints used for

A C C E P T E D M A N U S C R I P T
29 this analysis is quite low (222 for Katipunan, and 383 for Taft) with the IC periods of less than 5 578 minutes each. This is evident in the figures with ratios much greater than 1 and large differences. 579 As mentioned in the manuscript, the IC periods occurred in the middle of a run, hence, this 580 analysis do not cover a uniform dataset over the whole ATN range. Deriving a loading parameter 581 from this analysis would also be misleading as we do not expect that the loading parameter in one 582 point in space would be representative of the rest of the route in inhomogeneous atmospheres. 583 The loading parameter depends on the whole collected sample on the spot. 584 As a last attempt, a fixed k value of 0.005 was used to correct the Manila and Rome datasets (as 585 was done for the Loški Potok AE51 data). This value represents the loading effect of a diesel 586 exhaust dominated atmosphere as well as from fresh ambient wood burning (Drinovec et al., 587 2017). The corrected eBC was then plotted against the uncorrected eBC and is show in Fig. A5. 588 This shows that the correction did not change the eBC measurements substantially (6%, 8%, and 589 3% overall differences between corrected and uncorrected measurements for the Katipunan, Taft,  590 and Rome routes, respectively). As a result, no filter-loading effect correction was applied on the 591 Manila and Rome datasets. As for the Manila dataset, the discrepancy between the mobile AE51 592 and the reference instrument is due to the high variabilities of different factors (wind, sources, 593 etc.) characteristic of an urban area. 594

A C C E P T E D M A N U S C R I P T
30 Appendix B 597 <Tables of regression results> 598 599 600 10 a "am" -morning; "nn" -noon to afternoon; "pm"evening 708 709

A C C E P T E D M A N U S C R I P T
36

A C C E P T E D M A N U S C R I P T
37

A C C E P T E D M A N U S C R I P T
38

A C C E P T E D M A N U S C R I P T
39

A C C E P T E D M A N U S C R I P T
40    and (c) Loški Potok). This is basically the same as in Fig. 2, but now also segregated into 744 different IC locations. For IC details, see Tables 2, A1, A2, B1, and B2.      Katipunan Route, b) Taft route, and c) Rome city route. Data were taken from the raw AE51 770 measurements (1-s resolution) from all the runs performed in each location (see Table A2), 771 wherein a new filter was used for each run. The duration of a run is 1 hour for the Katipunan and 772 Taft Route, and 2.5 hours for the Rome route. The blue and red dots represent the median and 773 mean eBC mass concentration per bin, respectively, with the error bars as standard deviation. 774 The solid lines are the linear fit for each statistic. The whole ATN range was used for linear 775 fitting. 776