We present an approach to constrain simulated atmospheric black carbon (BC) using carbon monoxide (CO) observations. The approach uses: (1) the Community Atmosphere Model with Chemistry to simulate the evolution of BC and CO within an ensemble of model simulations; (2) satellite CO retrievals from the MOPITT/Terra instrument to assimilate observed CO into these simulations; (3) the derived sensitivity of BC to CO within these simulations to correct the simulated BC distributions. We demonstrate the performance of this approach through model experiments with and without the BC corrections during the period coinciding with the Intercontinental Chemical Transport Experiment (INTEX-B). Our results show significant improvements (∼50%) in median BC profiles using constraints from MOPITT, based on comparisons with INTEX-B measurements. We find that assimilating MOPITT CO provides considerable impact on simulated BC concentrations, especially over source regions. This approach offers an opportunity to augment our current ability to predict BC distributions.
 Understanding the distribution of black carbon (BC) is important for air quality (AQ) [e.g., Dickerson et al., 2002] and climate research [e.g., Ramanathan and Carmichael, 2008]. A product of incomplete combustion of fossil-fuel/biofuel and of biomass burning, BC in the atmosphere impacts the regional to global atmospheric radiative budget, as well as local AQ with implications for health and agriculture. However, our understanding of its distribution is hindered at present by limited BC measurements. Significant discrepancies in simulated BC still exist within state-of-the-art chemical transport and aerosol models (CTMs) [e.g., Shindell et al., 2008; Koch et al., 2009], precluding a comprehensive and accurate representation of its atmospheric state and variability. The lack of a global climatology for black carbon and the large uncertainty in its emissions hinders our knowledge of its importance to present radiative forcing.
 Currently, direct measurements of BC are limited across a number of surface stations and aircraft flights taken from a number of specific field campaigns. Indirect measurements of BC are mostly in the form of retrievals of aerosol optical depths (AOD) from satellite instruments and ground-based sunphotometer networks. These retrievals represent an integrated quantity that reflects the concentration of BC and other aerosols. As a result, it is difficult to determine precisely the mixing ratios of BC without assumptions of the optical properties and distribution of other aerosol species, which are also inadequately known. At present, model predictions of aerosols are constrained using a combination of these AOD retrievals [Benedetti et al., 2009, and references therein] and improvements in emission inventories [e.g., Bond et al., 2004]. The conventional approach maintains the model-derived relative aerosol fraction and vertical distributions in the assimilated distributions, which are highly uncertain.
 There are, however, extensive observations of chemical species such as CO that can provide additional constraints on model predictions of BC. Significant correlation between CO and BC can exist, especially within source regions where CO and BC emissions are collocated [e.g., Dickerson et al., 2002]. Several studies have used the correlations of ambient BC and CO to extrapolate the emissions of BC given a more well-known CO emission inventory. However, this relationship has not been fully exploited to constrain BC concentrations. This is partly because the correlation of BC and CO varies significantly over space and time, depending on the combustion source and efficiency, governing meteorology and other non-combustion sources of CO and BC [see, e.g., Kondo et al., 2006].
 In this work, we exploit the correlation between CO observations and BC concentrations, using relationships of CO and BC derived from an ensemble of CO and BC model simulations. The ensemble provides the basis for calculating sensitivity between a CO observation and model BC concentrations (see section 2 for detailed description). We use the CO retrievals from the Measurement of Pollution in the Troposphere (MOPITT) instrument on board the NASA Terra satellite [Deeter et al., 2003], together with a global CO and aerosol prediction system to demonstrate the applicability of this approach in the context of the NASA Intercontinental Chemical Transport Experiment (INTEX-B) field campaign during Spring 2006 [Singh et al., 2009].
 The physical basis behind the relationship between CO and BC mass lies in the fact that both species are primarily produced from combustion of carbonaceous fuels. Hence, most emissions of CO and BC are collocated. However, while the ratio of their emissions varies depending on the fuel type and the combustion process, the ratio of their atmospheric concentrations also depends on the processes controlling the removal of CO and BC aerosol, which vary with latitude and altitude. These processes account for the differences in the lifetimes of CO (∼1 month) and BC (∼1 week). Our approach is to model the spatio-temporal relationship of BC and CO through an ensemble of global simulations. Each of the ensemble members simulates the global evolution of BC and CO, with differences between ensemble members in both meteorology (e.g., temperature, wind, humidity, and clouds) and emissions of BC and CO (including chemical oxidation). Differences across the resulting ensemble of simulations represent the expected variability in emissions, meteorology, and removal processes. From this ensemble, we can calculate a local linear sensitivity of BC to CO. The sensitivity is calculated for each model grid and time interval as the ensemble covariance of BC and CO divided by the ensemble variance of CO. The simulated BC can then be corrected using the sensitivity of BC to the model-equivalent of the CO observation multiplied by the correction in CO derived from assimilating CO observations into these simulations (see also Figure S1 of the auxiliary material). To minimize spurious sensitivities due to small ensemble size, the corrections are done for BC concentrations within a local region of the CO observation (i.e., within ∼1200 km in the horizontal and ∼2 km in the vertical).
 We use a prediction system that includes a community data assimilation facility DART (Data Assimilation Research Testbed) and a coupled climate-chemistry model CAM-Chem (Community Atmosphere Model with Chemistry). The modeled meteorology is driven by assimilated meteorological observations. The model also assimilates CO retrievals, which are then used to correct the simulated BC. This system uses an ensemble Kalman filter framework. A description of this system and its initial application is presented by Arellano et al. . See auxiliary material for main references.
 For this application, we treat CO as a trace constituent in CAM-Chem by prescribing the sink of CO through a specified OH distribution (with CH4 lifetime of ∼10.5 yrs) derived from a previous CTM simulation. We specify its emissions from fossil-fuel/biofuel, biomass burning, and chemical oxidation of hydrocarbons as described by Arellano et al. . Injection heights for biomass burning emissions (CO and BC) are based on recommendations from the AeroCom project (http://nansen.ipsl.jussieu.fr/AEROCOM). BC is also treated as a trace constituent in CAM-Chem in a similar manner as in the configuration used by Shindell et al. . It is represented as a bulk aerosol mass. It is emitted as both a hydrophobic (80%) and hydrophilic (20%) species and is converted from hydrophobic to hydrophilic in the atmosphere with an exponential lifetime of 1.15 days. Emissions of BC are taken from Bond et al. . They are removed from the atmosphere through wet deposition via in-cloud and below-cloud scavenging and through dry deposition. The ensemble of model CO and BC concentrations are generated by perturbing the meteorology, emissions and injection heights in CAM-Chem (see auxiliary material for details and references).
2.2. Assimilation Experiments
 Two sets of experiments are carried out for this work. First, we conduct an ensemble simulation (“control”) that assimilates every 6 hours both meteorological and CO observations for April 16 to May 15, 2006. This ensemble simulation constrains the model meteorology and CO concentrations but not the BC concentrations. The second experiment (“constrained”) involves the same ensemble simulation but now with corrections to BC concentrations based on CO observations. The difference between the two experiments can be interpreted as the role of CO in constraining the BC distribution.
 To assimilate the meteorology we use meteorological data from a subset of the data processed at National Centers for Environmental Prediction (NCEP). This includes observations of temperature and horizontal wind from rawinsondes, cloud-drift analysis, and aircraft reports. The assimilation of these observations in CAM-Chem has been shown to effectively constrain the ensemble-mean meteorology. To assimilate CO we use CO retrievals from the MOPITT instrument. MOPITT has a horizontal resolution of 22 km × 22 km and offers near-global coverage within 3-5 days. The retrievals exhibit limited vertical sensitivity that is primarily in the free troposphere [Deeter et al., 2003]. For this application, we use MOPITT v3 L2 retrievals at 700 hPa with a priori contribution <50% and mixing ratios >30 ppbv to ensure that observations are representative of the true CO. Details on NCEP and MOPITT data assimilation are described by Arellano et al. .
 For verification, we use measurements of BC taken during the 2nd phase of the NASA INTEX-B field mission. The 2nd phase of the mission was aimed at sampling pollution outflow from Asia during April and May of 2006. BC measurements were taken from the Droplet Measurement Technologies (DMT) single particle soot photometer (SP2) instrument on board the NCAR C-130 aircraft [Subramanian et al., 2010]. SP2 uses a laser-induced incandescence technique to detect refractory BC particles independent of mixing state and particle morphology. The mass concentration of BC is derived from the incandescent signal and has been multiplied by ∼1.1 to 1.4 (depending on the flight) to correct for possible underestimation of mass outside the size range (145 nm–325 nm). These data have an overall uncertainty of about 22%. BC measurements are available for a total of 10 flights (including transit) during the campaign. A composite of C-130 flight tracks (excluding transit) over Seattle and Northeast Pacific is shown in Figure S2 of the auxiliary material. We focus our verification on science flights and classified the flights as over-ocean or over-land, depending on where the majority of the flight is located. This classification proves to be useful in the succeeding sections to distinguish regions that are constrained by MOPITT.
3. Results and Discussion
3.1. Estimates of BC and CO Relationship
 We first examine the overall relationship between BC mass and CO derived from the ensemble simulation. This relationship is represented in Figure 1 as the ensemble correlation of BC and CO averaged across the study period. The maximum sensitivity of BC to CO over this period is also presented to show potential constraints on BC from available CO observations. A strong relationship between CO and BC is clearly shown in the correlation estimates, mostly as a direct result of covariation in emissions (Figure 1d) and transport to the free troposphere (Figure 1b). Overall, the sensitivities are spatially heterogeneous, with highest sensitivity near the source regions and along transport pathways, strongly indicating the difference in the lifetime of BC and CO. At the surface, there is a large variation in the sensitivity suggesting the large spatial variability in emission ratios (i.e., combustion practices) and/or boundary-layer processes. As shown in Figure 1c, our estimate of BC/CO sensitivity over the southern US range from 2–4 ng m−3 STP/ppbv. This is very similar to ΔBC/ΔCO ratios observed in Maryland for this season [e.g., Park et al., 2005, and references therein]. Over Mexico, western/northern Europe and Tokyo, our estimates (∼2–12 ng m−3 STP/ppbv) appear to be a factor of 1.2 to 3 higher than observed [Subramanian et al., 2010; McMeeking et al., 2010; Kondo et al., 2006], whereas over India [Dickerson et al., 2002], our estimates (∼4–6 ng m−3 STP/ppbv) are lower roughly by a factor of 3. Model-data differences in spatio-temporal scales for which BC is represented and/or on how BC is defined and measured make a precise comparison between the model and data difficult. Nevertheless, our sensitivity and correlation estimates provide evidence of a strong relationship between CO and BC near the source regions, and further suggest the ability of the ensemble approach to simulate this relationship within reasonable limits relative to observed ΔBC/ΔCO ratios available from the literature.
3.2. Modeled and Observed BC Concentrations
 Here, we compare the modeled ensemble-mean BC distribution from both control and constrained experiments with observed BC during INTEX-B. Shown in Figure 2 are median BC profiles for flights classified as over-ocean. We find improvements in modeled BC with corrections from MOPITT for these flights. When corrections are applied, the correlations between modeled and observed BC profiles on April 21, 28, May 1 and 5 change from −0.10 to 0.77, 0.2 to 0.63, −0.04 to 0.65 and 0.51 to 0.82, respectively. The root-mean-square errors (RMSE) change from 14.1 to 8.2, 3.5 to 2.8, 7.9 to 6.5 and 10.9 to 6.6, respectively. In terms of average model skill, which is defined here as a function of correlation and RMSE [see Taylor, 2001, equation 5], the skill improved from ∼0.1 to ∼0.4 when the corrections are applied. This reflects >50% improvement in overall statistics for flights over the ocean. The constrained experiment shows that the magnitude of BC concentrations and its median vertical structure are closer to observed BC than the modeled BC from the control experiment. On one of the flights (April 21, 2006), the constrained experiment appears to be able to capture a large-scale pollution plume of Asian origin over NE Pacific that is consistent with observations including MODIS AOD retrievals. Modest levels of relatively-aged BC are observed downwind of Asia, suggesting transport of BC towards the NE Pacific basin for a period that is well within the lifetime of BC (see Figures 1a and 1b). On the other hand, we find that the modeled BC profiles over Seattle and the west coast of the US (not shown) are not constrained by MOPITT. Small changes in these BC profiles for the constrained experiment are not significant and are within the variability of the observation. The lack of BC constraints appears to coincide with the lack of MOPITT CO observations over this area (see Figure S1), as a result of our choice of data quality control that excludes among others, retrievals over high terrain and cloudy conditions.
3.3. BC Distribution
 Here, we examine the impact of MOPITT CO to the overall ensemble-mean BC distribution. The mean BC distribution as simulated from the constrained experiment and the differences in the mean BC distribution between the constrained and control experiments are presented in Figure 3 (see also Figure S3 of the auxiliary material). These differences are averaged over the study period and reflect the mean MOPITT-based corrections to simulated BC. These corrections represent the combined effect of: (1) the magnitude of model CO corrections, and (2) the ensemble sensitivity of model BC concentrations to MOPITT CO. These corrections are also localized within a fixed distance around the MOPITT pixels. The overall corrections can help elucidate model uncertainties in transport, aerosol removal, and BC emissions highlighted in recent model inter-comparisons [e.g., Shindell et al., 2008]. We find significant corrections in the Asian aerosol plume across the Pacific, in the African plume from biomass burning and over regions of large BC emissions (Figures 3 and S3). At the model surface (Figure 3d), there are corrections >50 ng m−3 over Northeast US, Mexico, China, Southeast Asia, India, northern Europe, and south of Brazil. Based on comparisons with limited surface data, we find that these large BC corrections in the model are qualitatively consistent in sign with surface observations. While at specific measurement locations these corrections are not as large (in magnitude and spatial extent) as some of the model discrepancies reported in the literature [e.g., Shindell et al., 2008; Koch et al., 2009], our analysis suggests that this methodology gives significant improvements to large-scale gridded estimates of the BC distribution. Within Mexico City, the mean BC concentration is ∼700−3000 ng m−3 at the Universidad Nacional Autónoma de México during April 2003 and 2005 [Baumgardner et al., 2007]. The corresponding model mean BC concentration in April 2006 is ∼600 and ∼1000 ng m−3 for control and constrained experiment, respectively. In IMPROVE sites over New York and Massachusetts (http://vista.cira.colostate.edu/improve), the mean BC concentration for April/May is ∼260 ng m−3 compared to ∼450 and ∼350 ng m−3 from control and constrained experiments. In northern Europe, there is an increase in model BC concentration (130 versus 200 ng m−3) consistent with higher observed BC of ∼400 ng m−3 from one of the EMEP campaign sites during April/May 2003 (http://www.emep.int). Over Korea and Japan, the observed mean BC concentration is ∼560 ng m−3 for April 2005 from the Atmospheric Brown Clouds (ABC) observatory [Ramanathan et al., 2007], whereas the model concentration is ∼400 and ∼470 ng m−3 in the control and constrained experiments. We have not compared the large corrections over China, Southeast Asia, India, and Brazil due to lack of accessible data.
 We exploit the observations of CO and the relationship between BC and CO as has been suggested in the literature to constrain model concentrations of BC. Our approach is to simulate this relationship through an ensemble of model simulations and use this to correct simulated BC given CO observations. Our results show significant improvements in the model BC distribution with the assimilation of MOPITT CO. We compared these results with BC measurements during the INTEX-B campaign and corroborated them in a qualitative sense with BC observations from available surface sites and ΔBC/ΔCO ratios from literature. We find that CO observations can be used to constrain BC distribution in places where a strong relationship exists between BC and CO and where CO observations are adequate. We recognize that there are limitations to this approach. As mentioned in section 2.2, the information (i.e., sampling and vertical sensitivity) from MOPITT is not sufficient to fully constrain BC. We also need to conduct sensitivity studies on the impact of our model assumptions, especially with regards to the prescribed BC aging, the relative proportion of hydrophobic/hydrophilic emissions, assumptions of external mixing, and other unrepresented removal mechanisms.
 Nevertheless, we find that the results offer an opportunity to improve our knowledge of the global distribution of BC, given that at present there is a lack of continuous global direct measurements of BC to fully constrain its distribution. It is envisaged that even better constraints can be derived using a combination of CO and AOD observations, especially those that provide better spatio-temporal coverage (e.g., METOP/Infrared Atmospheric Sounding Interferometer CO), better sensitivity to the surface (e.g., MOPITT multispectral CO) and aerosol classification. In addition, there is a potential value in using this approach in constraining other combustion-related and climate/AQ-relevant species like CO2 and Hg.
 We thank G. Kok and R. Subramanian of Droplet Measurement Technologies (DMT) for BC INTEX-B data, NCAR MOPITT team for CO and NCEP for meteorological data. We also thank INTEX-B team for access to other field data, IMPROVE and EMEP for surface BC data, F. Vitt, K. Raeder, J. Anderson, L. Emmons and H. Worden for help and comments. This work was supported by NASA grant NNX07AL57G and in part by NSF ITR grant 115912. NCAR is sponsored by the National Science Foundation.