The Hawaii Group for Environmental Aerosol Research (HiGEAR) operated several in situ instruments on the P-3. These include two Radiance Research particle soot absorption photometers (PSAPs) to measure the aerosol absorption coefficients (at 470, 530, and 660 nm), and two TSI (model 3563) nephelometers (at 450, 550, and 700 nm) to measure the aerosol light scattering coefficients. In addition to the TSI nephelometers, two single wavelength nephelometers (at 540 nm, Radiance Research, M903) were operated concurrently to measure the increase in light scattering as a function of RH. The humidified nephelometer was operated close to 80 % RH, while the dry unit was maintained below 40 % (Howell et al., 2006; Pistone et al., 2019). For in situ SSA or extinction calculation, the measured PSAP absorption was interpolated to nephelometer wavelengths before combining with its scattering data. For both SSA and extinction, we use the corrected or processed data (called SSA_ATP and Exttotal_ATP) that are provided at a time resolution of 1 s within the “merged” files at the ORACLES ESPO Data Archive: https://espo.nasa.gov/oracles/archive/browse/oracles/id8. Here, PSAP absorption corrections were performed according to an updated algorithm (Virkkula, 2010) and nephelometer-measured scattering was corrected according to Anderson and Ogren (1998). The in situ SSA and extinctions are measured and reported at low (<40 %) RH. Additionally, we only use the measurements for which mid-visible dry extinctions are greater than 10 Mm-1 (Pistone et al., 2019).

In addition to dry in situ SSA, we also use the retrievals of column-integrated SSA for ambient conditions from the Spectrometer for Sky-Scanning Sun-Tracking Atmospheric Research (4STAR) instrument, which was also on board the P-3. The 4STAR instrument is an airborne hyperspectral (350–1700 nm) sun photometer that can make direct-beam measurements (sun-tracking mode) for retrievals of partial-column aerosol optical depth (AOD) above flight level (LeBlanc et al., 2020). Under ideal flight and atmospheric conditions, 4STAR can also perform sky scans in either the principal plane or almucantar (sky-scanning mode). The 4STAR sky scans were processed using a modified version of the version 2 Aerosol Robotic Network (AERONET) retrieval algorithm described in Dubovik and King (2000). However, due to suspected stray light contamination within the 4STAR spectrometer around 440 nm (Pistone et al., 2019), the set of input wavelengths for 4STAR were modified to be 400, 500, 675, 870, and 995 nm (as opposed to the standard AERONET input wavelengths). We used the quality-screened data available at the ORACLES ESPO Data Archive and presented in Pistone et al. (2019), but we further limit our analysis to SSA retrievals that had AOD400 nm>0.4 similar to what AERONET uses in its quality control criteria. Additionally, we only consider the SSA retrievals for which flight altitude was greater or equal to 1 km, which is about the typical boundary layer height over the ocean for the ORACLES 2016 region. Since 4STAR retrieves the above-aircraft column SSA, our last screening criterion was used to focus on the SSA of smoke layers above the boundary layer and exclude the influence of marine aerosols within the boundary layer.

We also use the atmospheric profiling provided by the NASA Langley Research Center High Spectral Resolution Lidar (HSRL-2, Hair et al. (2008)) that was deployed on the ER-2 aircraft. Note that flight paths for the P-3 and ER-2 only partly overlapped (Fig. 1a), for example on the 24 September case (Fig. 1b) that we analyze in greater detail to utilize this synergy between the two aircraft (Sect. 3.1). HSRL-2 measures aerosol backscatter and depolarization at 355, 532, and 1064 nm and aerosol extinction at 355 and 532 nm. Out of the standard aerosol data products (Burton et al., 2012) that HSRL-2 provides, we use the aerosol extinction and lidar ratio (extinction-to-backscatter ratio) profiles at 355 and 532 nm for this study, both provided in 60 s averages to improve the instrument signal-to-noise ratio (equates to ∼12 km along-track horizontal resolution of the aircraft).

The Single Particle Soot Photometer (SP2) was deployed on the P-3 as part of HiGEAR to measure the mass of refractory black carbon (rBC) particles by passing a powerful laser beam and heating them to incandescence (Schwarz et al., 2006). The peak value of this incandescence signal has been shown to linearly correlate with the mass of the rBC particle (Stephens et al., 2003). The bulk submicrometer non-refractory aerosol composition, on the other hand, was provided by the time-of-flight (ToF) Aerodyne aerosol mass spectrometer (AMS) in the form of organic aerosol (OA), sulfate (SO4), nitrate (NO3), and ammonium (NH4) mass concentrations (DeCarlo et al., 2008). The AMS-measured mass concentrations are provided at standard temperature (273 K) and pressure (1000 hPa), but we convert them here to ambient conditions using the ideal gas law and measured pressure and temperature information before comparing with the model equivalents.

Particle size distributions used in this study were measured from the P-3 with an ultra-high-sensitivity aerosol spectrometer (UHSAS, Droplet Measurement Technologies, Boulder CO, USA) with a fuselage-mounted inlet. UHSAS is an optical-scattering, laser-based aerosol particle spectrometer that measures particles from 60–1000 nm at 1 s time resolution, thereby covering the entire accumulation mode. The UHSAS measured size distribution is reported as particle number concentrations (in cm-3) for size bins that are approximately logarithmically spaced.

CO was measured from the P-3 with a gas-phase CO / CO2 / H2O analyzer (ABB/Los Gatos Research CO / CO2 / H2O analyzer known as COMA) with an inlet mounted on the aircraft fuselage. It uses off-axis integrated cavity output spectroscopy (ICOS) technology to make stable cavity enhanced absorption measurements of CO, CO2, and H2O in the infrared spectral region (Provencal et al., 2005). The measurements were reported as dry air volume mixing ratios in parts per billion (ppbv).

We use space-based aerosol observations from the Ozone Monitoring Instrument (OMI; Levelt et al., 2006) on board the NASA Aura spacecraft, which flies in a polar orbit with a local Equator-crossing time of 13:30 UTC. OMI is a hyperspectral (270–500 nm), wide swath (∼2600 km) imager that observes the back-scattered solar radiation with a nominal ground pixel size of 13×24 km2 at nadir and 13×150 km2 at the swath edges. Aerosol products are from the OMAERUV algorithm (version 1.8.9.1, after Torres et al., 2007). Fundamental to the OMAERUV retrieval is the so-called aerosol index (AI). The AI is computed from the observed radiances at two channels where ozone absorption is weak (354 and 388 nm for OMAERUV), where the observed spectral contrast is compared to the (easily characterized) spectral contrast expected in a purely molecular atmosphere. Under cloud-free conditions the AI is sensitive to aerosol loading, height, and absorption, where increases in any of those quantities result in a higher positive magnitude value of the AI. The OMI swath ideally provides near-daily global coverage but has been impacted by a “row anomaly” defect since shortly after launch that has effectively degraded its coverage by about 50 % so that OMI now achieves global coverage every two days (Torres et al., 2018). In our comparisons that follow we sample model output only where valid OMI data are collected. Owing to its relatively large pixel size OMI data are frequently cloud contaminated. We restrict our analysis to the best available data, where the algorithm quality-assurance flag is 0.

The Moderate resolution Imaging Spectroradiometer (MODIS) sensors have been flying on two spacecraft, Terra (10:30 LST, local solar time, Equator crossing) since 2000 and Aqua (13:30 LST Equator crossing) since 2002. The MODIS collection 6.1 algorithms (namely, Dark Target Land, Dark Target Ocean, and Deep Blue) retrieve aerosol properties over both ocean and land surface types using the observed spectral reflectance (Levy et al., 2013). However, instead of directly using the MODIS operational retrievals of AOD for model evaluation, we use a bias-corrected AOD dataset, called the MODIS NNR, which was derived initially for use in the Modern-Era Retrospective analysis for Research and Applications, version 2 (MERRA-2, Randles et al., 2017) aerosol reanalysis. As much care as is taken in creating the MODIS standard products, there are nevertheless significant biases related to cloud and land features that must be screened prior to using these data in assimilation systems (e.g., Zhang and Reid, 2006). The NNR refers to a neural net retrieval algorithm that computes AERONET-calibrated AOD from satellite-based radiances, in this case the same MODIS collection 6.1 radiances used in the standard retrieval products. To derive 10 km resolution MODIS NNR AOD, over-ocean predictors include MODIS level 2 multichannel TOA reflectances, glint, solar and sensor angles, cloud fraction (pixels are discarded when cloud fraction >70 %), and albedo derived using GEOS surface wind speeds. Over land, predictors are the same, except a climatological albedo is included for pixels with surface albedo <0.15. The NNR algorithm is trained on the log-transformed AERONET AOD interpolated to 550 nm and co-located with MODIS observations of the predictors. Application of the NNR to the MODIS products is found to produce a higher-quality AOD product compared to (relatively unbiased) AERONET observations (Randles et al., 2017).

Our “smoke age” simulation (see Table 1) has the brown carbon tracer “tagged” in such a way as to determine the day of the week on which its emissions were injected. Effectively, eight instances of the biomass burning organic aerosol are tracked in this simulation (seven instances, one for emissions occurring each day of the week, and an eighth for the cumulative total emissions). The “Monday” tracer, for example, resets the tracer concentration globally to zero when the model clock ticks over to Monday at 00:00 UTC. The differences between the total and individual day-of-the-week tracers allows determination of the smoke age out to smoke emitted 7 d previously. The simulation is otherwise identical to our baseline simulation.

Our “smoke composition” simulation has the biomass burning OA tagged by the vegetation source burned based on the QFED inputs. QFED biomass burning emission sources are classified using the land cover or vegetation type information provided by the MODIS Land Cover Type Product (MCD12C1, IGBP classification scheme, Friedl et al. , 2010) for the year 2016, shown in Fig. 2. The MCD12C1 product supplies global maps of land cover at annual time steps at 0.05° spatial resolution in geographic latitude–longitude projection. Further details on the Collection 6 MODIS Land Cover products, including MCD12C1, are provided in their User Guide: https://lpdaac.usgs.gov/documents/101/MCD12_User_ Guide_ V6.pdf, last access: 5 December 2023. We define six major vegetation or land cover (LC) types over central and southern Africa that could contribute to the smoke over ORACLES region: (1) savannas (LC type 9), (2) woody savannas (LC type 8), (3) grassland (LC type 10), (4) tropical forest (LC type 2, 4, 5), (5) shrubland (LC type 6,7), and (6) croplands (LC type 12, 14) (Fig. 2). We then define seven instances of brown carbon tracers (six are “tagged” based on their emission source vegetation type, plus one for the total) within the model to be able to investigate the contributions of each of these vegetation types towards the smoke observed over the ORACLES region.

We employ a radiative transfer code to simulate the OMI aerosol index (Buchard et al., 2015; Colarco et al., 2017). Similar to the computation of the model AOD, the AI simulator takes the GEOS-simulated aerosol mass distributions and meteorological fields as input, and – subject to the GEOS aerosol optical property assumptions – simulated OMI radiances are calculated at the OMI viewing conditions (viewing geometry, terrain height, and surface reflectance). AI is then calculated as in Colarco et al. (2017). Because we do not explicitly simulate the impact of modeled cloud fields on the simulated radiance, we restrict our comparisons to the highest-quality OMI retrievals (formally, QA flag = 0) to eliminate cloudy pixels in the satellite product impacting our comparisons as much as possible.

To understand the vertical variations in SSA, we first analyze the smoke composition of this 24 September profile by comparing the model-simulated aerosol mass concentrations to the corresponding measurements from the AMS and SP2 in situ instruments (Fig. 6). Organic aerosols (from all sources) contribute about 65 %–70 % of the total aerosol mass, followed by nitrates, sulfate, and BC in both the model and the observations. Model-simulated aerosol composition agrees well with in situ observations, especially for the top plume centered around 4.5 km (where both show peak OA mass concentration ∼17 µg m-3). OA is overestimated by the model for altitudes below 2.5 km (Fig. 6a), while nitrates are underestimated for the same altitudes (Fig. 6c). Sulfate (Fig. 6d) and BC (Fig. 6b) show a similar profile shape to the simulated OA but are present at lower concentrations than the observations suggest, particularly for the lower-altitude layer. Sulfate is about half the concentration in the model as observed. BC has similar concentration to the observations for the higher smoke plume (∼1 µg m-3) but is only about half the concentration of the observations in the lower layer. Except for nitrate, the observed multi-layer structure for this profile is evident in the model species.

SSA strongly depends on the relative amounts of scattering to absorbing aerosols. BC is the primary absorbing component at 550 nm in the smoke plumes simulated here, while OA, nitrates, and sulfates are mostly scattering. Given that the SSA is changing with altitude (Fig. 5d) and multiple plumes are observed within the vertical profile (Fig. 5a), we investigate the following two things: (1) the smoke age in the vertical and (2) how the relative aerosol composition varies with altitude and smoke age. Figure 7a shows the model-simulated, extinction-weighted mean smoke age for our 24 September profile using our smoke age tagged tracer run (described in Table 1 and Sect. 2.4.1). Fig. 7a also shows the smoke age derived using the WRF-AAM (Weather Research and Aerosol Aware Microphysics) model (Saide et al., 2016) that was used for forecasting and flight planning during the ORACLES campaign (Redemann et al., 2021). The GEOS and WRF-AAM models have a similar overall structure of simulated plume age, showing minimal smoke age around the center altitude of the upper smoke plume (∼4 d at 4.5–5 km altitude) and higher smoke age at lower altitudes, with a local maximum of about 5–6 d between 3–4 km and increasing to 7–8 d below 2 km. We also show the smoke age distribution from our GEOS smoke age run at three different altitudes of the profile to clarify the interpretation of the mean smoke age number and to note some of the limitations of model-calculated smoke age (Fig. 7b). For example, for the youngest smoke plume centered at 4.5 km, even though the mean smoke age is calculated as 4.1 d, about 40 % of the smoke is 1–2 d old, while about 20 % of the smoke is older than 7 d, suggesting quite a wide distribution. At 3 km altitude the lower mode of the smoke age distribution peaks at 2–3 d old, and at 2 km it peaks at 3–5 d old. We should also be mindful that different age distributions could lead to the same mean age value. Finally, for GEOS, we are restricted by the way we track the smoke age that we can only resolve smoke age up to 7 d, and older smoke is lumped into this last bin of our histogram so that the effective age computed is slightly younger than if we could account for all possible smoke ages. Nevertheless, the simulation shows the presence of generally younger smoke at higher altitudes and older smoke at lower altitudes.

Next, we investigate how the relative composition of aerosols changes with smoke age and if we can further use that relation to explain the vertical variations of SSA as well. Figure 8a shows the simulated and observed change in SSA with the GEOS-simulated mean smoke age for the profile shown in Fig. 5d, and this variation is almost identical to the SSA variation in the vertical since smoke age and altitude show a strong negative correlation. The initial drop in SSA between 4–5.5 d estimated smoke age is correlated with a simultaneous decrease in OA and nitrates with respect to BC (Fig. 8b, c). After 5.5 d age the nitrates are essentially constant in time, and thus a further decrease in SSA with smoke age is more clearly correlated with a continued decrease in OA : BC ratio (Fig. 8b). The model captures the variation in NO3 : BC and SO4 : BC ratio well for most parts of the profile (Fig. 8c, d). The major discrepancy between the model and observations lies in the inability of the model to simulate the observed OA : BC variation, which changes in the observations from about 15 (for 4 d old smoke) to about 5 (for 7 d old smoke), whereas the model is essentially constant at a OA : BC ratio of 15. Figure 8e and f show, respectively, the correlation of OA : BC with SSA (r=0.89) and NO3 : BC with SSA (r=0.88).

The model's failure to simulate the observed OA : BC ratio variability with age is correlated with its failure to simulate the observed variability in SSA. In the following we consider two hypotheses to explain the age-related variation in the OA : BC ratio. In the first we consider the possibility that smoke of different ages may be originating from different source regions with different emissions of OA and BC. In the second we consider the possibility that there is some non-simulated mechanism for the loss of OA during transport that is related to its age.

The first hypothesis we examine is whether smoke of different ages is originating from different source regions with different characteristics in the composition of emitted species or different proportions of the flaming and smoldering phases of the combustion products. We use three approaches to examine this hypothesis: (1) we generate ensembles of back-trajectories to track the origin of smoke layers at different vertical levels, (2) we utilize the vegetation type smoke composition tracer run of our GEOS model (Table 1, Sect. 2.4.2) to quantify the contribution of each vegetation type towards OA amounts (or extinction), and (3) we examine the observed black-carbon-to-carbon-monoxide ratios as an indicator of whether the smoke arises from flaming versus smoldering combustion. Our analysis is for the same 24 September aerosol profile discussed previously.

GEOS uses biomass burning emissions from the QFED inventory based on the MODIS fire radiative power products (Sect. 2.4). QFED provides daily gridded biomass burning emission fluxes of relevant species, such as OA, BC, and SO2. Figure 9a shows the QFED emission locations and cumulative amounts of OA emitted over the 7 d prior to 24 September. Figure 9b and c show the HYSPLIT-model-generated ensemble back trajectories (multi-colored lines) initialized from two different vertical levels (4.5 km versus 2 km) that represent the location of the two distinct smoke plumes evident in the profile plots (e.g., Fig. 6a). For the higher-altitude smoke layer (centered around 4.5 km), the clustering of the trajectories and their intersections with the surface (where they would entrain smoke) occurs only a few days before intercepting our profile location (the black star in Fig. 9b, at 12° S, 11° E), consistent with the smoke being young (about 4 d old, Fig. 7). By contrast, the lower-altitude initialized back trajectories (originating at 2 km, Fig. 9c) intercept the surface several days back and further to the east of the profile location, consistent with the smoke at these levels being older (about 6–7 d old, Fig. 7).

The trajectory information alone does not tell us that the composition of the smoke in the two different layers is similar. Our smoke composition simulation (see Table 1 in Sect. 2.4) is used to distinguish the contributions of individual vegetation types to the total smoke composition. QFED distinguishes among several different vegetation types according to land use datasets, and vegetation-dependent emission factors are used to scale from biomass burned to emissions of specific species (i.e., OA and BC). In the smoke composition simulation we “tag” the smoke emissions from each vegetation type so we can track its evolution separately. In Fig. 10a we quantify the contributions of emissions from individual vegetation types to the smoke composition at our profile location separately for altitudes of 4.5 and 2 km, which are the central altitudes of the two smoke plumes in the profile. Our results indicate that the major emission sources for both smoke layers are savannas, woody savannas, and grasslands. The contributions from other vegetation types (tropical forests, shrublands, and croplands) amount to only about 20 % of the total emissions. More importantly, the contributions from all vegetation types, except grasslands and woody savannas, are comparable for both the higher- and lower-level smoke plumes. The higher grassland contributions (relative to woody savannas) for the 2 km level are possibly due to smoke originating from the grassland-dominated regions south of 15° S and east of 20° E based on the 2 km back trajectories (Fig. 9c) and the land cover map (Fig. 2). However, Akagi et al. (2011) and Andreae (2019), which are the key databases that most global models use for their assumptions of emission factors of different biomass burning species, do not differentiate between grasslands and savannas as fuel types, assigning them both the same emission factors. After savannas and grasslands, the next most prevalent vegetation type contributing emissions is forest, which has a higher OA : BC emission ratio than grasslands and savannas but contributes only about 10 % to the total smoke load and is similar in contribution for both plumes. Crop and agricultural residue have distinct OA : BC ratios (that are highly uncertain; see Table 2) but are an even smaller contribution to the total aerosol load and are also a similar contribution to each layer.

Finally, differences in the OA : BC ratio could also be due to different combustion characteristics of the source fires (i.e., flaming versus smoldering), wherein flaming fires are known to generate a lower concentration of organic carbon particles but a higher concentration of BC than smoldering fires (Christian et al., 2003; Yokelson et al., 2009). We note here that we currently do not make the distinction in emission ratios based on fire characteristics within QFED or GEOS. Vakkari et al. (2018) suggests that the excess-BC-to-excess-carbon-monoxide ratio (ΔBC / ΔCO or BC : CO here onwards) can be used as a reliable marker to assess the combustion characteristics, especially for diluted and aged plumes. An increasing ΔBC / ΔCO value is indicative of increasing flaming fraction during the burning (Yokelson et al., 2009). Precisely, Vakkari et al. (2018) distributed the BC : CO ratios into bins of 0.005 to represent fires of similar characteristics. Further, based on their measurements over southern African savanna and grassland regions, they found that BC : CO <0.005 represents predominantly smoldering conditions, while BC : CO >0.010 represents predominantly flaming conditions. We consider background CO to be 100 ppbv based on the values outside the plume (found at altitudes >6 km, see Fig. 10b) for the 24 September case. Figure 10c shows that the excess BC : CO ratio has very little variation in the vertical, and the values are within the range of 0.005–0.010, thereby suggesting an approximately equal mix of smoldering and flaming combustion conditions at altitudes between 1 and 5.5 km.

Based on the analyses shown here we conclude it is unlikely that the observed differences in smoke composition (i.e., the OA : BC ratio, as in Fig. 8b) at different vertical levels are due to differences in either the burning source vegetation type or combustion conditions.

The second hypothesis we consider is that there is a chemical or microphysical loss of OA during transport that is related in some way to the smoke age. OA from biomass burning sources is composed of primary organic aerosol (POA) directly emitted from the burning biomass and SOA formed via oxidative processing of organic gases. As the biomass burning plume ages in the atmosphere, the oxygenation of OA (reflected in O / C ratio) is reported to increase significantly in almost all studies, suggesting strong chemical transformation of OA with aging and formation of more oxidized OA that is compositionally different from POA (Capes et al., 2008; Jolleys et al., 2012; Cubison et al., 2011; Forrister et al., 2015; Zhou et al., 2017). OA mass is increased by SOA formation but can also be balanced out by dilution and subsequent evaporation to the gas phase of semivolatile components, resulting in loss of POA mass (Cubison et al., 2011; May et al., 2015; Zhou et al., 2017). In addition, the volatilities of organic species are affected by atmospheric oxidation reactions. In the gas phase, two main processes can affect the volatilities of organics during atmospheric oxidation: fragmentation and functionalization. Fragmentation refers to the loss of carbon from the organic particles, whereas functionalization refers to an increase in particulate oxygen due to the addition of polar functional groups. Therefore, fragmentation leads to an increase in vapor pressure (i.e., the organic compounds become more volatile), while functionalization leads to lowering of vapor pressure of the organic compounds. In the condensed phase, additional bimolecular processes, such as accretion and oligomerization reactions can also affect volatility (Kroll et al., 2009). Changes in OA composition with aging inevitably lead to changes in OA optical and hygroscopic properties and thus have climate implications. The reasons for the variability across studies may be related to variations in fuels burned and combustion conditions; variation in key co-emitted species such as NOx; and differences in environmental conditions such as dilution rate, temperature, and humidity (Shrivastava et al., 2017).

Within GEOS, the OA : BC ratio is prescribed at the point of emission based on the fuel type and related emission factors (see discussion in Sect. 3.1.2). Further, OA is emitted as 50 % hydrophobic and hydrophilic, while BC is emitted as 80 % hydrophobic and 20 % hydrophilic. Once emitted, both OA and BC undergo hydrophobic-to-hydrophilic conversion with a conversion rate or e-folding time of 1–2 d (Colarco et al., 2010). This hydrophobic-to-hydrophilic conversion can change the OA : BC ratio somewhat from what it is at emission due to wet scavenging, but for older smoke (>2 d old) we do not account for any additional OA or SOA loss due to any microphysical or chemical processes. Indeed, our simulated OA : BC ratio remains mostly constant over time, as in Fig. 8b.

With the present model limitations in mind, we introduce an ad hoc loss process for the hydrophilic OA in our model. We additionally increase our overall emissions of biomass burning produced OA to compensate for this added loss channel to approximately preserve the overall regional extinction and AOD. Sensitivity studies suggest a 6 d e-folding time applied to the biomass burning hydrophilic OA, with a corresponding increase of about 60 % in emissions of OA and about 15 % in the BC emissions from biomass burning, greatly improves our simulated OA profile (Fig. 11a). When these factors are applied to our simulations (“OA-loss” simulation, Table 1), we can simulate the observed SSA variation with altitude well (Fig. 11b). Simultaneously, these changes resulted in overall more absorbing aerosol mixtures. Overall, there is better agreement in the SSA between 4STAR and the OA-loss simulation than 4STAR and the baseline, except at the shortest wavelength (Fig. 11c, where at 400 nm we have SSA4STAR=0.88, SSABaseline=0.87, and SSAOA-loss=0.86, and at longer wavelengths the SSAOA-loss is closer to 4STAR than the baseline by a magnitude of 0.02; see also Table 3).

Satellite-based observations are used in this section to assess the model performance over the broader SE Atlantic and the source region over southern and central Africa. Figure 16a shows the September 2016 monthly mean MODIS NNR AOD. We sample the model AOD based on the MODIS swath, overpass time, and data availability, and the resulting monthly mean AOD from the model baseline and the final “OA-loss + updated optics” simulations are shown in Fig. 16b, d, respectively. The last column (Fig. 16c, e) shows the differences between the model and the observations. Broadly, as depicted in the AOD difference plots, both model simulations show a good agreement with the NNR retrievals, especially over the ocean and the smoke outflow region compared to the source region over the continent. There is a significant underestimation in terms of model AOD over the southeastern parts of the continent that persists in both the model simulations. This is a shortcoming of the GEOS model that has persisted in previous model versions as well, possibly due to missing biomass burning or coarse-mode aerosol sources (Das et al., 2017). Otherwise, there is a high bias in the model AOD in the areas around the western coast and 10° S in the baseline simulation (Fig. 16c). In our final simulation, however, this high bias in model AOD is reduced from +0.02 to -0.02 (Fig. 16d, e). Closer to the burning sources over land, this decrease in model AOD in our final simulation is due to the adjusted hygroscopic growth for the OA particles (see Sect. 3.2), whereas over the ocean, as the plumes move away from the continental burning sources, the decrease in model AOD is due to the accounting of OA loss with increasing smoke age. Finally, we also present the performance metrics (mean bias; root-mean-square error, RMSE; and Pearson correlation coefficient, r) to quantify the agreement between the model simulations and the observations with a focus over the ORACLES-I region (Fig. 16c, e). The statistical comparisons further emphasize that inclusion of OA loss processes and adjustments to model OA optics does not deteriorate the model performance in terms of AOD simulation, but that it instead makes it marginally better compared to baseline run by lowering the RMSE (from 0.11 to 0.08) and increasing the r values (from 0.91 to 0.93). Note that we chose to show only the Aqua MODIS results here since Aqua has a closer overpass time with OMI (Aura) than Terra, but the comparisons of model AOD simulations with respect to Terra MODIS retrievals (not shown here) also showed similar results.

In Fig. 17 we show the GEOS-simulated OMI AI for September 2016 in comparison to the OMI retrievals. The OMI retrieved AI is shown in Fig. 17a. To minimize the impact of clouds on the comparison we retain only OMI pixels with QA values of 0 (low probability of cloud contamination). For the model-calculated AI, the GEOS-simulated aerosol profiles are sampled at the OMI footprints. The model optical property assumptions are applied to the simulated aerosol profiles, and with the OMI observation geometry and retrieved surface reflectance they are input to the AI calculator, which simulates the OMI radiances. The simulated radiances only include terms for the aerosol, molecular background, and surface.

Two cases are shown in the comparison: the GEOS baseline run (Fig. 17b) and the final run that includes OA loss with smoke age, increased biomass burning emissions, and ORACLES-informed optical properties for OA (OA-loss + updated optics, Fig. 17d). The baseline run, with its more highly absorbing OA assumptions and high bias in aerosol loading especially near the coast of southwestern Africa (10° S, Fig. 16c), results in the model overestimating the AI relative to OMI, with this high bias especially apparent over both land and ocean where the AOD was overestimated in the model (Fig. 17c). By contrast, the GEOS model run with the ORACLES-informed adjustment to the optical property assumptions, along with the reduced overall burden of OA, results in a much more favorable comparison of the simulated to retrieved AI (Fig. 17e). The apparent discontinuity in the AI magnitude between land and sea along the western coast of southern Africa is a sampling artifact brought on by cloud screening, with nearly 3× as many OMI pixels retained over land as for over the coastal ocean. For this reason, we restrict a quantitative assessment of the model performance to the boxed region over land shown in Fig. 17c, e, and statistics are also reported in Fig. 17. In the indicated region there is an improvement in the bias of the modeled AI from +0.28 to -0.14 and a reduction in the RMSE from 0.42 to 0.30 in moving to the updated aerosol optics. This improvement and our assessment of the model performance with the independent OMI dataset shows the overall consistency between the optical properties that were derived in situ from the ORACLES measurements and what can be retrieved from space-based remote sensing, with the GEOS model as the interpolator between the two datasets.