Correction for a measurement artifact of the Multi-Angle Absorption Photometer ( MAAP ) at high black carbon mass concentration levels

Introduction Conclusions References


Introduction
A widely used method for measuring atmospheric black carbon (BC) mass concentration involves the determination of absorption of an aerosol sample collected on an appropriate filter matrix.The most common instruments utilized Published by Copernicus Publications on behalf of the European Geosciences Union.
today for this purpose are the filter-tape-based Aethalometer (Hansen et al., 1984), Multi-Angle Absorption Photometer (MAAP) (Petzold et al., 2002;Petzold and Schönlinner, 2004), and the single-filter-based Particle Soot Absorption Photometer (PSAP) (e.g., Bond et al., 1999).Since BC by definition cannot be unambiguously measured with these instruments, it is customary to refer to the measured carbonaceous light absorbing aerosol constituent as equivalent BC (BCe) or light-absorbing carbon (LAC).For the sake of simplicity, we use the term BC throughout.For a detailed discussion of the nomenclature used for black carbon or light-absorbing carbon components of the atmospheric aerosol, see, e.g., Bond and Bergstrom (2006) and Andreae and Gelencsér (2006).
It is well known that filter-based BC measurements suffer from several artifacts.These include the filter loading effect that causes a decrease in the measured BC concentration with increasing filter load, and the sample matrix effect that causes scattering aerosols on the filter to increase the measured BC concentration.These artifacts can be corrected to some extent by using different numerical methods (e.g., ond et al., 1999;Weingartner et al., 2003;Arnott et al., 2005;Virkkula et al., 2007;Collaud Coen et al., 2010).All of the correction schemes have their advantages and disadvantages under field conditions.Thus far, the MAAP has been deemed as the most reliable filter-based instrument for measurement of BC, since the instrument design and software take the typical filter-related artifact effects into account.
We have conducted aerosol field measurements in Gual Pahari (India) from December 2007 to January 2010, including BC measurements with the MAAP (Hyvärinen et al., 2010).During this campaign, we observed that at high BC concentrations the MAAP is not free of measurement artifacts.The observed artifact is different from those seen with other filter-based BC instruments, and to our knowledge has not been reported before in the literature.Here, we quantify this artifact with the assistance of laboratory measurements utilizing two MAAPs operating at different flow rates.The focus of this paper is to raise awareness of the MAAP artifact within the aerosol community, and to demonstrate how the artifact can be circumvented by logging the reflected photodetector signal.In addition, we present a method for correcting the results from the typical instrument print formats in the post-processing phase.The correction is applied to three MAAP datasets: Gual Pahari (India) (Hyvärinen et al., 2010); Beijing (Garland et al., 2009) (China), and Welgegund (South Africa) (Beukes et al., 2012;www.welgegund.org).In Beijing, the results could be compared against a photoacoustic spectrometer (PAS; Garland et al., 2009).

Multi-Angle Absorption Photometer
The Thermo Scientific Model 5012 MAAP measures the aerosol BC mass concentration at a single nominal wavelength 670 nm.However the true wavelength has later been measured to be 637 nm (Müller et al., 2011).The typical filter-loading-related artifacts are already taken into account in the design and the internal programming of the instrument.In inter-comparison tests, the MAAP has been found to give reliable results of light absorption by aerosols (e.g., Sheridan et al., 2005;Petzold et al., 2005).
The principle of measuring the absorption coefficient (b AP ) using multi-angle absorption photometry has been well documented during instrument development (Petzold and Schönlinner, 2004).The key principle is that, in addition to the typical transmission measurement, the signals scattered to angles at 130 and 165 • are also measured (Fig. 1).Additionally, radiative processes are modeled by a radiative transfer scheme for the particle-loaded filter, the aerosolfilter layer alone, and the blank filter alone (Hänel, 1987).The final output of the radiative model is the single scattering albedo (ω FILTER ) and the optical depth (τ FILTER ) of the aerosol loaded filter layer that match the measured transmitted and reflected signals.From these values, the absorption coefficient of the MAAP (b AP,MAAP ) is given by the following equation (Petzold et al., 2005): where b ATN,MAAP is a method-dependent coefficient related to absorption, A is the filter spot area, and V is the sampled volume.
Since the values of ω FILTER and τ FILTER at a given time t always refer to the initial values of the particle-free filter, the incremental increase in b AP,MAAP during time interval The air flow is drawn through a glass fiber filter tape, and the ambient aerosol is collected on a sampling spot of A = 2 cm 2 area.The sample volume flow through the instrument is measured continuously by the pressure drop across an orifice.For default instrument settings, the filter tape is moved forward to the next blank sampling spot after the transmission of the particle-loaded spot has decreased below 20 %.The initial signals at detection angles 0, 130 and 165 • for the particlefree sample spot are determined after the filter spot change during a zeroing procedure, while respective values for the particle-loaded filter are measured at distinct time intervals during aerosol sampling.
In the commercial software version provided by Thermo Instruments, the attenuation of light by the deposited aerosol is measured in time steps of 1 min during continuous aerosol sampling.Values referring to longer time intervals are calculated as averages from the basic 1 min data.
In addition to the MAAP method that utilizes the instrument's internal algorithm, the absorption coefficient can also be determined from the photodetector raw signals in a postprocessing procedure.The typically used print formats of the MAAP do not give the raw signals as an output, and have to be logged, e.g., by using the scientific print format 12.These raw signals include the photodetector response signal at 0 • (transmittance), and signals at 130 and 165 • (reflectance).From the 0 and 165 • signals, the light attenuation (ATN) by the sample can be determined as where (T 0 /T ) and (R 0 /R) are the ratios of photodetector signals at 0 and 165 • for a particle loaded and a particlefree filter, respectively.The factor 0.5 in the reflectance measurement results from the fact that the light passes through the layer of sampled aerosol twice before reaching the photodetector.
The measured properties b ATN,TRANS and b ATN,REFL cannot be used directly to obtain the BC mass concentration, because typical filter loading artifacts affect the measured signals.Petzold et al. (2005) where ω 0 is the single scattering albedo of the aerosol.Finally, the absorption coefficients obtained with different methods (MAAP, TRANS and REFL) are directly proportional to the BC mass concentration (BC METHOD ) by a factor of 1/σ BC , where σ BC = 6.6 m 2 g −1 , the mass-specific absorption cross section of BC at a wavelength of 637 nm.
The increase of BC mass deposited on the filter spot during one time interval of sampling, or BC mass accumulation rate BC in µg cm −2 , is calculated accordingly as The value of σ BC was determined from a series of lab and ambient measurements (Petzold and Schönlinner, 2004) against gravimetric and thermal reference methods.It is in close agreement with the value of 7.5 ± 1.2 m 2 g −1 at 550 nm for "fresh" carbonaceous aerosol (Bond and Bergstrom, 2006) taking into account scaling of σ BC with the inverse wavelength.

Initial observations
The artifact was first noticed during measurements at the EU-CAARI station in Gual Pahari (India).The station was located about 25 km south of New Delhi.Typical for the Indo-Gangetic plains, the region is heavily polluted.For the measurement campaign period of December 2007-January 2010, the PM 10 and PM 2.5 average values were 216 µg m −3 and 126 µg m −3 , respectively (Hyvärinen et al., 2010).The average BC mass concentration was found to be 12.3 µg m −3 .The BC measurements were conducted utilizing a PM 10 inlet, and the aerosol was dried by a diffusion drier prior to entering the instrument.The MAAP (Thermo Scientific Model 5012) was run at a nominal flow rate of 8 L min −1 .The data were mostly saved as 5 or 1 min averages.During high concentration periods, an artificial decrease in concentration was observed in the MAAP data at Gual Pahari (Fig. 2).The artifact is related to the filter spot change conducted by the instrument; i.e., when a new blank filter spot is moved into the sampling head, the observed BC concentration decreases.The artifact is clearly not related to the typical filter loading effects, which result in concentration increase rather than decrease after the filter spot change (e.g., Virkkula et al., 2007;Petzold et al., 2005).
For a better understanding of the artifact, we analyzed the instrument raw signals from the MAAP at Gual Pahari.Using Eqs.(1)-( 5) and the absorption cross section σ BC , the whole dataset was converted to three different BC values: BC MAAP , BC TRANS and BC REFL .The ω 0 in Eq. ( 4) for calculating b AP,TRANS was estimated to be 0.9, which is typical for ambient aerosols.This choice was found to affect the magnitude but not the shape of the transmitted signal.
When comparing the results obtained with the three different methods (Fig. 2), it becomes apparent that the artifact is already present in the raw photodetector signals, especially in the transmitted 0 • signal.BC TRANS almost mimics the behavior of BC MAAP , although showing somewhat higher concentrations in general.BC REFL exhibits a much weaker, yet visible artifact effect.It is also notable in Fig. 2 that, during the few minutes when the filter spot change takes place, there is a strong scattering in the raw signal, which is related to stabilization of the instrument.The exact mechanism causing the artifact is not known, but it seems to be related to erroneous dark counts in the photodetector measuring the transmitted light during the filter zeroing procedure, in combination with an instrument internal averaging procedure of the photodetector raw signals.

Quantification of the artifact
A laboratory test was conducted to accurately quantify the artifact.The basic assumption for designing the quantification experiment was that the observed artifact is closely related to the BC mass accumulation rate BC in µg min −1 .Hence, two MAAP instruments were set up to sample from the same aerosol but operated on different volume flow rates.This in turn resulted in different values for BC.
Test aerosol was produced by atomizing a water solution of "Aquadaq", a soot reference standard (Baumgardner et al., 2012), into a mixing volume of ∼ 5 L. Two MAAP instruments sampled from this mixing volume -one (s/n 145) with a high flow rate (16-20 L min −1 ), and the other (s/n 87) with a low flow rate (7-10 L min −1 ).Make-up air was taken from the lab through a HEPA filter.The concentration of BC was changed by changing the flow rate of the atomizer.MAAP flow rates were calibrated against a Gilibrator bubble flow meter.The concentration-and flow rate-ratios were controlled so that the artifact occurred only in the high flow rate MAAP, thus utilizing the low flow rate MAAP as a reference instrument.
As the photodetector response (and thus the mass accumulation rate) depends both on the BC mass concentration and the flow rate Q of the instrument, we express the laboratory results in terms of BC = BC × Q, i.e., accumulation rates.We prefer to use this notation, as both BC and Q can be logged with the MAAP standard print formats.Cumulatively, this also becomes the mass on the filter spot, m.The results indicate that at BC mass accumulation rates >∼ 0.04 µg min −1 the artifact can be identified from the data.The observed artifact is very systematic.After an initial drop, the BC signal increases with a rate proportional to the prevailing BC concentration.The artifact can be roughly divided in three distinct regions (Fig. 3a): 1.At moderately high mass accumulation rates (0.04 µg min −1 < BC × Q < 0.08 µg min −1 ), only the very first minutes experience a decrease of the signal, followed by a prominent increase above the initial signal level, before stabilization back to the correct level occurs.
2. At high mass accumulation rates (0.08 µg min −1 < BC × Q < 0.14 µg min −1 ), the initial decrease of the signal becomes more apparent compared to the following overestimation.However, the point where the concentration stabilizes back to the correct level is still very distinct, as it is characterized by 1 or 2 min of clearly higher concentrations.
3. At very high mass accumulation rates (BC × Q > 0.14 µg min −1 ), the initial signal decrease is so strong that the signal never recovers to the correct level before the next filter spot change, leading to an inevitable underestimation of the BC concentration.This reveals that an erroneous temporal response is not the only outcome of the artifact.At high enough BC concentrations, the MAAP underestimates BC values entirely (Fig. 3b).

Correction algorithm
To compile a correction algorithm, we addressed the two principal problems present in the original data, i.e., (1) the overall concentrations, which are underestimated when the rising MAAP signal cannot reach the true concentration (Fig. 3b), and (2) the temporal response of the erroneous concentrations (Fig. 3a).
1.The overall correction was made by correlating the smoothed data from the high flow rate MAAP with the data from the reference low flow rate MAAP.The smoothed concentrations, BC smooth , could simply be the last few minutes before the next filter spot change, by removing the other data suffering from the artifact.The correlation can be described with a third level polynomial resulting in the corrected BC CORR × Q: BC CORR × Q = 5.665 ± 0.25 In conditions with changing concentrations, only considering data from a few minutes before each spot change might be misleading due to the poor time resolution.
2. In order to smooth the temporal response, we assumed that the real concentration is a sum of the extracted artifact signal and real changes in the concentrations, and may thus be expressed as where BC ini is the concentration before the spot change, BC meas the measured non-corrected concentration and BC artifact the artifact signal dependent on the initial concentration described below.
The shape of the MAAP artifact signal can be described with the so-called Hill function as a function of the mass of BC on the filter spot, m: where BC max × Q is the maximum plateau value simulated by the Hill function, and k and n are parameters describing the slope of the rising mass accumulation rate.
The measured accumulated mass was chosen as the base for characterizing the artifact, as it is a reasonable assumption that the artifact is dependent on both the initial mass accumulation rate and the change of accumulated mass on the filter spot.The laboratory cases were fitted with this function, and the parameters BC max × Q, k and n were optimized (Supplement).The parameters can be expressed with the following functions and constants: where BC ini × Q is the mass accumulation rate in µg min −1 before the spot change.As noted earlier, when BC × Q < 0.14 µg min −1 , the artifact signal eventually makes an abrupt decrease back to the correct level (see Fig. 3a).This point was found to follow the relation: The algorithm is not able to predict this decrease back to the real signal level, and should not be applied if the accumulated mass from the spot change is greater than m d (in µg).
An overall representation of the laboratory results is illustrated in Fig. 4. The smoothed data (BC smooth ) correct for the temporal response, but not the overall underestimation.However, the polynomial regression (Eq.7) brings the two datasets to a 1 : 1 ratio (R 2 = 0.99).Equation ( 7 0-0.14 µg min −1 with an average relative difference of 3.8 % from a linear 1 : 1 correlation.The upper limit here is restricted by the artifact arising in the reference MAAP.The low flow rate data were used up to BC × Q = 0.14 µg min −1 so that only the artifact-free data were chosen for the correlation.The average absolute deviation of the corrected measurement points of the high flow rate MAAP compared to the low flow rate MAAP was 0.49 µg m −3 , and the corresponding average relative deviation was 5.4 %.The highest individual deviations typically occur during the first 2-3 min after the filter spot change, due to the very steep signal increase of the artifact.

Correction from the raw reflectance signal
Similarly to field observations in Gual Pahari, the laboratory dataset was also converted to BC REFL using Eqs.
(3) and ( 5).We see that these data follow the reference concentrations closely (Fig. 4), although with more scatter (R 2 = 0.96).Obviously the reflectance signal is not affected by the measurement artifact.This observation may indicate that the measurement artifact is occurring only in the processed transmission signal, which then affects also the final MAAP output signal while the reflectance signal only, with appropriately applied corrections of the filter matrix effect (Petzold et al., 2005), is not affected by the artifact and reports accurate BC mass concentration values.This presents an opportunity to utilize the reflected signal from the MAAP when mass accumulation rates are high enough that the artifact appears.However, as most of the print formats of the MAAP do not give the raw signals as an output, the algorithm is a useful way for correcting the results.

Application of correction to ambient measurements
Three ambient datasets were chosen for testing the correction algorithm: Gual Pahari (India) (Hyvärinen et al., 2010); Beijing (Garland et al., 2009) (China), and Welgegund (South Africa) (Beukes et al., 2012;www.welgegund.org).All these locations suffer from such high BC concentrations that the measurement artifact could be observed from the data.In Gual Pahari, the MAAP was run at 8 L min −1 .The correction algorithm was applied to the full dataset from 14 December 2007 to 19 January 2010.In Beijing, the flow rate of the MAAP was measured to be 9.2 L min −1 .The algorithm was applied to a short-term dataset from 10 August 2006 to 9 September 2006, for which we could additionally utilize a PAS as reference method for the absorption of BC.Finally, in Welgegund (South Africa) a MAAP was run at a flow rate of 16.7 L min −1 .These data covered the period from 1 June 2010 to 31 August 2010.
Using the algorithm (Eqs.7-10) clearly improves the temporal response of the MAAP signal, removing most of the signal decreases observed in the uncorrected data from all three locations (Fig. 5a, c and e).On occasion, the first points after a filter spot change still show a concentration decrease for BC CORR , which is due to the steep shape of the artifact: a small uncertainty in m (accumulated mass) can lead to a large uncertainty in BC.In addition, problematic situations may occur when the true concentration exhibits a high gradient during the filter spot change.This can happen especially if there are short-term pollution episodes taking place.In such cases the assumption that the last value of BC on the previous filter spot equals the initial concentration is not valid.This has direct consequences on both the modeled new concentration and the length of the artifact effect.If there is a BC concentration decrease during the filter spot change, the concentrations would be underestimated and the last point of the artifact, m d , would be overestimated.For an increase, the effect is the opposite.As seen in the example figures, BC REFL may occasionally show values lower (Fig. 5a) or higher (Fig. 5e) than BC CORR .However, the overall trends produced by the two methods are very similar.
In order to better evaluate the performance of the suggested correction algorithm, we compared the original BC MAAP and the corrected BC CORR against BC REFL from all three ambient locations (Fig. 5b, d and f).The erroneous temporal response of BC MAAP is again evident from the downward "tails" appearing in the figures.In addition, BC MAAP underestimates the higher concentrations, similarly to what was seen in the laboratory experiments.Linear regressions fitted to the data (Table 1) do not indicate a substantial difference between the relations BC MAAP vs. BC REFL and BC CORR vs. BC REFL .However, this is mostly because the median concentration is well below 10 µg m −3 at all the locations.The regressions do reveal that the correlation at all locations is improved by the correction.For Welgegund, the correlation is worst (R 2 = 0.94), and in general BC REFL is higher than BC CORR .At this location, high BC concentrations are related to pollution episodes, and 90 % of data points are below ∼ 3 µg m −3 .It is possible that the poorer correlation in Welgegund is related to the assumptions made in the determination of BC REFL .Also in Gual Pahari, BC REFL is generally higher than BC CORR .Opposite to Welgegund, here the difference is probably related to the very high overall concentrations (90th percentile = 24.5 µg m −3 , maximum 73.1 µg m −3 ).At mass accumulation rates higher than the applicability of the correction function (0.4 µg min −1 ), concentrations may be underestimated.While we cannot say with certainty which method produces the most accurate results, it has to be kept in mind that BC REFL is corrected from the raw photomultiplier signal by an empirical function based on test aerosols (Eq.5).In ambient conditions, especially of high loading with strongly scattering aerosol, Eq. ( 5) may not be valid.
The results do, however, indicate that the artifact correction based on the laboratory experiment may be applied to different ambient environments.Finally, we were able to compare the results from a PAS against those derived from the MAAP in Beijing.The absorption coefficient values from MAAP at 637 nm were converted to those at 532 nm (PAS wavelength) by assuming an absorption Ångström exponent of 1.In addition, the MAAP in Beijing was run with 1 : 1 dilution, while the PAS sampled without dilution, so the MAAP absorption coefficients were further multiplied by 2. Although the results from the MAAP are slightly lower (by 15 %) than those reported from the PAS (Fig. 5d and Table 1), the difference can be considered acceptable for a correction scheme.Similar and consistent results compared to a PAS were obtained, when the algorithm was applied to MAAP data from another megacity region, Guangzhou, China (Garland et al., 2008;R. M. Garland, private communication, 2012).This agreement further confirms our laboratory findings.

Conclusions
We have observed a measurement artifact in the MAAP at high BC concentrations.The artifact is related to the filter spot change -as mass is accumulated on a fresh filter spot, the photodetector response of the transmitted 0 • light is lower than anticipated.However, the 165 • photodetector signal is not compromised.The artifact seems to be related to erroneous dark counts in the transmitted light photodetector, in combination with an instrument internal averaging procedure of the photodetector raw signals.The artifact behavior however appears to be entirely related to the currently implemented data inversion algorithm, but not to any unknown physical processes.Using raw data on a 1 Hz basis and post-processing the data independently by an algorithm similar to that described by Petzold and Schönlinner (2004) shows no artifacts as described here (T.Onasch, private communication, 2011).The artifact can be observed if the BC mass accumulation rate BC × Q exceeds 0.04 µg min −1 .At the typical flow rate of 1 m 3 h −1 , this relates to a BC concentration of ∼ 3 µg m −3 .Overall concentrations of uncorrected MAAP data are underestimated if BC × Q exceeds 0.14 µg min −1 .With increasing BC accumulation rate, the underestimation may be several tens of percent.
We compiled an algorithm to correct the BC estimation from the typically most commonly used print formats of the MAAP.The algorithm is not dependent on the saving interval of the data and takes the instrument flow rate into account.The algorithm was tested on data originating from three different ambient environments, and was found to improve all the datasets considerably.In principle, the artifact can also be avoided by diluting the sampled air, but this will result in a loss of accuracy at lower concentrations.The MAAPreflected signal may also be used to derive correct concentration levels.Therefore, it is strongly recommended to log the raw reflectance signals of MAAPs in highly polluted environments.However, utilizing solely the reflected signal may result in an increased noise in the data.
An updated version of the MAAP firmware is currently in preparation for distribution.However, the correction algorithm as described here is urgently needed for correcting data from worldwide operated MAAP instruments.

Fig. 1 .
Fig. 1.Schematic of the MAAP.The transmitted light is measured with the photodetector at θ 0 = 0 • , and the reflected light with the photodetectors at θ 1 = 130 • and θ 2 = 165 • .(Figure from: "Aerosol Science and Technology: Evaluation of Multiangle Absorption Photometry for Measuring Aerosol Light Absorption", 39, 40-51, Copyright 2005, Mount Laurel, NJ, reprinted with permission.) determined relationships for the b AP and b ATN,TRANS as well as b AT N,REF L by utilizing test aerosols.These test aerosols consisted of pure black aerosol samples from kerosene flame particles, and externally mixed gray and black aerosols of varying single scattering albedo.The correction functions are b AP,TRANS = b ATN,TRANS × (0.654

Fig. 2 .
Fig. 2. Comparison of BC measured with the MAAP by multi-angle photometry BC MAAP , transmission BC TRANS and reflectance BC REFL methods during a high concentration episode in Gual Pahari, India, 12 August 2009.

Fig. 3 .
Fig. 3. (a) Examples of the artifact temporal response.Symbol lines indicate data from the high flow rate MAAP, and solid lines data from the reference MAAP.Regimes 1, 2 and 3 correspond to accumulation rates (BC × Q) of 0.04-0.08µg min −1 , 0.08-0.14µg min −1 and > 0.14 µg min −1 , respectively.See text for details.(b) Comparison of the low flow rate reference (BC ref ) and high flow rate (BC MAAP ) BC concentrations.

Fig. 4 .
Fig. 4. High flow rate MAAP BC concentrations as a function of the reference (low flow rate) MAAP concentrations.BC MAAP is the original signal; BC smooth is the smoothed signal without an overall concentration correction.BC CORR is the smoothed signal with the overall concentration correction.BC REFL is the original signal determined with the reflectance method.The 1 : 1 ratio line is also shown.

Fig. 5 .
Fig. 5. Examples of the correction algorithm applied to the datasets from Gual Pahari (a, b), Beijing (c, d), and Welgegund (e, f).In panels (b), (d), and (f) the datasets from different methods are compared against the reflected BC signal, BC REFL .Lines in these figures are linear fits to data.BC in Beijing was measured with 1 : 1 dilution.For Beijing, the absorption coefficients b METHOD at 532 nm from the PAS and those derived from the MAAP (see text for details) are also shown.

Table 1 .
Linear regression analysis of BC MAAP , BC CORR and BC PAS against BC REFL at three ambient locations.10th percentile, median and 90th percentile of the concentrations (in µg m −3 ) are also shown.