Description of the NASA GEOS Composition Forecast Modeling System GEOS‐CF v1.0

Abstract

The Goddard Earth Observing System composition forecast (GEOS‐CF) system is a high‐resolution (0.25°) global constituent prediction system from NASA's Global Modeling and Assimilation Office (GMAO). GEOS‐CF offers a new tool for atmospheric chemistry research, with the goal to supplement NASA's broad range of space‐based and in‐situ observations. GEOS‐CF expands on the GEOS weather and aerosol modeling system by introducing the GEOS‐Chem chemistry module to provide hindcasts and 5‐days forecasts of atmospheric constituents including ozone (O₃), carbon monoxide (CO), nitrogen dioxide (NO₂), sulfur dioxide (SO₂), and fine particulate matter (PM_2.5). The chemistry module integrated in GEOS‐CF is identical to the offline GEOS‐Chem model and readily benefits from the innovations provided by the GEOS‐Chem community. Evaluation of GEOS‐CF against satellite, ozonesonde and surface observations for years 2018–2019 show realistic simulated concentrations of O₃, NO₂, and CO, with normalized mean biases of −0.1 to 0.3, normalized root mean square errors between 0.1–0.4, and correlations between 0.3–0.8. Comparisons against surface observations highlight the successful representation of air pollutants in many regions of the world and during all seasons, yet also highlight current limitations, such as a global high bias in SO₂ and an overprediction of summertime O₃ over the Southeast United States. GEOS‐CF v1.0 generally overestimates aerosols by 20%–50% due to known issues in GEOS‐Chem v12.0.1 that have been addressed in later versions. The 5‐days forecasts have skill scores comparable to the 1‐day hindcast. Model skills can be improved significantly by applying a bias‐correction to the surface model output using a machine‐learning approach.

Key Points

• GEOS‐CF is a new modeling system that produces global forecasts of atmospheric composition at 25 km² horizontal resolution

• GEOS‐CF model output is freely available and offers a new tool for academic researchers, air quality managers, and the public

1. Introduction

Near real‐time information of global atmospheric composition is invaluable for a wide range of applications, including academic research, airborne and satellite mission support, air quality forecasting, disaster management, and ecosystem monitoring. However, the numerical simulation of atmospheric chemistry is computationally expensive because it involves hundreds of species that interact with each other on time scales from milliseconds to years, and the species are also influenced by dynamics across a wide range of spatiotemporal scales. This precludes the inclusion of detailed aerosol and reactive trace gases in standard operational numerical weather prediction (NWP) systems. Instead, real‐time simulation of atmospheric composition is typically done within a simplified system in order to reduce the computational burden, for example, by running the model at reduced horizontal resolution or over a regional domain only, using a simplified representation of atmospheric composition, or by coupling a weather model with an offline chemical transport model (CTM) (e.g., Bhattacharjee et al., 2018; Emmons et al., 2020; Flemming et al., 2009; Marécal et al., 2015).

Here we present the NASA Goddard Earth Observing System (GEOS) composition forecast modeling system, GEOS‐CF v1.0, which provides global near real‐time estimates (“hindcasts”) and forecasts of atmospheric composition such as ozone (O₃), carbon monoxide (CO), nitrogen dioxide (NO₂), sulfur dioxide (SO₂), and fine particulate matter (PM_2.5) in near real‐time at a horizontal resolution of approximately 25 × 25 km². As summarized by the World Meteorological Organization (WMO) Global to Local Air Quality Forecast Inventory (https://community.wmo.int/activity-areas/gaw/science-for-services/gafis/air-quality-forecast-inventory), GEOS‐CF is one of only a few global forecasting systems conducting gaseous and aerosol composition simulations in near real‐time. The European Center for Medium‐Range Weather Forecasts (ECMWF) offers 5‐days global forecasts of aerosols and trace gases at approximately 40 × 40 km² horizontal resolution through the Copernicus Atmosphere Monitoring Service (CAMS, https://atmosphere.copernicus.eu/global-forecast-plots). The US National Center for Atmospheric Research (NCAR) conducts 10‐days global forecasts at approximately 100 × 100 km² horizontal resolution based on offline simulations of the Model for Ozone and Related chemical Tracers (MOZART) Chemistry Mechanism in the Community Earth System Model Version 2 (CESM2) (Emmons et al., 2020) driven by GEOS meteorological forecasts (https://www2.acom.ucar.edu/acresp/forecasts-and-near-real-time-nrt-products). Finally, the Finnish Meteorological Institute provides daily 4‐days global forecasts of atmospheric composition at approximately 35 × 35 km² resolution using the System for Integrated modeling of Atmospheric coMposition (SILAM v5.7, http://silam.fmi.fi/). The atmospheric composition forecasts from these models can vary considerably due to differences in the underlying meteorological fields, observational constraints, chemical mechanisms, or assumptions about pollutant emissions (see, e.g., Huijnen et al., 2019). The uncertainties associated with these processes can be difficult to quantify from a single model simulation alone, and the availability of multiple, independently developed models offers great potential to provide better air quality information through combination of a suite of models (Marécal et al., 2015). This is recognized by the Global Air Quality Forecasting and Information System (GAFIS) project, a new initiative by the WMO Global Atmospheric Watch (GAW) program with the aim to provide globally harmonized air quality forecasts tailored to the needs of end users. GEOS‐CF can support such efforts by offering an independent but complementary composition forecast to the existing suite of global composition forecasting systems in the form of global 5‐days forecasts of atmospheric composition based on the GEOS‐Chem atmospheric chemistry module (http://www.geos-chem.org) within the GEOS Earth System Model (ESM).

One of the key aspects of the GEOS‐CF system is the full integration of the offline GEOS‐Chem CTM in the GEOS system, which allows for the simulation of reactive gases and aerosols at the same temporal and spatial resolution as the meteorology (Hu et al., 2018; Long et al., 2015). GEOS‐Chem is a global CTM driven by archived meteorological inputs produced by GEOS. It is actively evaluated and developed by a large international research community. The chemical components of GEOS‐Chem (including emission, deposition, and aerosol‐oxidant chemistry) can also be used online in an ESM such as GEOS, bypassing the need for the offline transport scheme used by the CTM. Instead, meteorology and species transport is calculated directly within the ESM. The scientific validity of the GEOS‐Chem chemistry module within GEOS has been demonstrated by Hu et al. (2018), who show that a one‐year global simulation of atmospheric composition at 12.5 km² produces results consistent with the offline GEOS‐Chem model. The online application has advantages over the CTM as it does not introduce transport errors related to temporal and spatial averaging of the archived meteorological input fields (Yu et al., 2018), but comes at the cost of increased complexity and computing requirements. GEOS‐CF is an application of the online GEOS‐Chem module embedded in GEOS, with a focus on daily operation and forecast capabilities. The GEOS‐Chem chemistry module incorporated in GEOS‐CF uses the exact same codebase as the offline CTM. This enables seamless integration of scientific updates provided by the GEOS‐Chem CTM community into GEOS‐CF without the need to make any modifications to the source code.

The GEOS‐CF modeling system is the latest in a series of research and applications products generated by the NASA Global Modeling and Assimilation Office (GMAO), including the GEOS forward processing weather and aerosol system (GEOS FP; Lucchesi, 2017), GEOS FP for instrument teams (GEOS FP‐IT; Lucchesi, 2015), the second Modern‐Era Retrospective analysis for Research and Applications (MERRA‐2; Gelaro et al., 2017), and the Seasonal to Subseasonal Forecasting System (GEOS‐S2S; Borovikov et al., 2019; Molod et al., 2020). It leverages GMAO's model infrastructure and directly builds on a number of development activities centered around the GEOS model, with the goal to extend these forecasting capabilities toward (short‐lived) trace gases and aerosols.

The GEOS ESM is a General Circulation Model (GCM) and Data Assimilation System (DAS) consisting of a suite of model components that can be connected in a modular manner through the Earth System Modeling Framework (ESMF, Hill et al., 2004) and Modeling Analysis and Prediction Layer (MAPL, Suarez et al., 2007) software interface. The model can be configured to run with fully interactive chemistry so that the chemical constituents feed back to the dynamics (“online”), or as an offline model where external meteorological fields are used as input. A hybrid approach is the “replay” feature, where the model dynamics are nudged toward pre‐computed analysis fields (from a previous DAS simulation) in a way that is consistent with the internal physics of the model (Orbe et al., 2017). This approach is particularly useful for chemistry simulations as it sidesteps the need to conduct a computationally costly meteorological data assimilation cycle. Several chemistry and aerosol modules of varying complexity are available in GEOS (Nielsen et al., 2017), enabling a wide range of applications including but not limited to near real‐time simulation of aerosols using the GOCART module (Buchard et al., 2017; Colarco et al., 2010; Randles et al., 2017), computationally efficient analysis of stratospheric O₃ using parameterized chemistry in combination with 3D‐Variational assimilation of satellite observations (Wargan et al., 2015); multi‐decade simulation of tropospheric and stratospheric chemistry using the Global Modeling Initiative (GMI) chemistry module (Douglass et al., 2004; Duncan et al., 2007; Strahan et al., 2007); and fully coupled simulation of gas‐phase and aerosol chemistry using the chemistry module of the GEOS‐Chem CTM embedded in GEOS (Long et al., 2015). The latter is used as the principal driver of chemistry in GEOS‐CF.

In this paper we describe the configuration of GEOS‐CF version 1.0 (Section 2) and demonstrate the validity of the produced hindcasts through comparison against independent observations for years 2018 and 2019 (Sections 3 and 4). Model forecasts over the same time period are evaluated in Section 5 and we discuss how model skill scores can be improved by applying a bias‐correction to the surface observations using a machine‐learning approach. As described in detail below, the current version of GEOS‐CF constitutes a hybrid model between an online weather and chemistry assimilation system and an offline CTM application, with a development pathway toward a fully coupled forecasting system with integrated trace gases and aerosols. Many of its design features were guided by practical considerations as well as computational limitations related to the simulation of atmospheric chemistry.

2. Model Description

2.1. General Description

A schematic of the GEOS‐CF v1.0 modeling system is provided in Figure 1. Since 2018, GEOS‐CF is operated in near real‐time, producing a 5‐days forecast once a day. The forecast initial conditions are provided by a 1‐day replay simulation (referred to as “hindcast”) constrained by pre‐computed meteorological analysis fields. In the v1.0 configuration, the meteorological forcings that GEOS‐CF uses during the hindcast step are obtained from the GEOS FP‐IT system. GEOS FP‐IT is a “frozen” model system that is comparable to MERRA‐2 but—unlike MERRA‐2—is available in near real‐time to support retrievals by satellite instrument teams. The GEOS model in general, and the version of the GEOS GCM used by GEOS FP‐IT (v5.12.4) in particular, has shown to be well suited for atmospheric chemistry applications as it realistically captures features critical to atmospheric composition, such as the seasonal climate of moisture and temperature and large‐scale transport of constituents (e.g., Douglass et al., 2014; Gelaro et al., 2017; Molod et al., 2015; Oman and Douglass, 2014; Pawson et al., 2007).

Figure 1.

Schematic of the GEOS‐CF modeling system approach, consisting of 1 day hindcast and 5‐days forecast. This combination of simulations is conducted on a daily basis.

In GEOS FP‐IT, the GOCART module provides aerosol feedback on meteorology. To ensure consistent model physics between GEOS FP‐IT and GEOS‐CF, the physics component in GEOS‐CF is coupled to the GOCART module, with GOCART aerosol optical depth (AOD) constrained by satellite observations as further described below. The GEOS‐Chem chemistry module, which contains its own aerosol module coupled to the gas‐phase chemistry, is run “passively” to provide coupled aerosol‐oxidant chemistry in the troposphere and stratosphere (Hu et al., 2018). That is, the GEOS‐Chem aerosols and trace gases are advected online (including turbulence), while emissions, dry and wet deposition, and chemistry are handled by the GEOS‐Chem chemical module driven by instantaneous meteorological information from the GEOS physics modules. Deep convective transport of chemical species is performed by the GEOS‐Chem convection scheme driven by instantaneous model diagnostics to capture scavenging in the updrafts. The GEOS‐Chem aerosols and trace gases are currently integrated without observational constraints, with the exception of stratospheric ozone which is nudged to precomputed ozone fields from the GEOS FP system (described in more detail below).

A full snapshot of the model conditions at the end of the simulated hindcast day are saved and serve as input for the next day's hindcast step (“restart files”). This preserves the model chemical and meteorological state from one forecast cycle to the next, leading to a continuous atmospheric composition archive since January 2018.

GEOS‐CF v1.0 uses a model physics package that is similar to MERRA‐2 and GEOS FP‐IT, consisting of the GEOS atmospheric model, version 5, described in Rienecker et al. (2008) with updates as described in Molod et al. (2012) and Molod et al. (2015). The model uses the finite‐volume dynamical core of Lin (2004) with a cubed sphere grid discretization to avoid grid‐cell singularities (Putman and Lin, 2007). It is run at cubed‐sphere c360 horizontal resolution (roughly equivalent to 0.25° × 0.25°) and 72 hybrid‐eta levels from the surface to 0.01 hPa. This results in 30–35 model levels in the troposphere and a surface model level depth of approximately 130 m. Model physics includes parameterizations for moist processes, radiation, turbulent mixing, land‐surface processes, and gravity wave drag. The moist module contains parameterization of convection using the Relaxed Arakawa‐Schubert scheme (Moorthi & Suarez, 1992), and the single‐moment parameterization for large‐scale precipitation and cloud cover described in Bacmeister et al. (2006). Note that in GEOS‐CF v1.0, the moist processes in GEOS are not directly applied to the GEOS‐Chem species but provide the diagnostic variables needed by GEOS‐Chem to perform these tasks (Hu et al., 2018; Yu et al., 2018). The radiation module includes parameterization for long‐wave (Chou 1990, 1992) and short‐wave radiation processes (Chou and Suarez 1994). Turbulence is parameterized using the gradient Richardson number in the free atmosphere and the Lock scheme (Lock et al., 2000) interfaced with the scheme of Louis and Geleyn (1982) in the boundary layer. Exchange of heat, moisture and momentum between land, atmosphere, and ocean or sea ice surfaces are parameterized using Monin‐Obukhov similarity theory (Helfand and Schubert, 1995; Molod et al., 2013), and the gravity wave drag parameterization contains orographic (McFarlane, 1987) and non‐orographic (Garcia & Boville, 1994) waves.

In the GEOS‐CF system, the GEOS physics components are coupled to the GOCART aerosol component to provide consistent physics with the GEOS FP‐IT meteorology. Specifically, the GEOS‐Chem chemistry module is run alongside the GOCART module, with no feedbacks to model physics.The GOCART aerosols, constrained during the hindcast step by AOD observations from the Moderate Resolution Imaging Spectroradiometer (MODIS) aboard the Terra and Aqua satellites (Randles et al., 2017), are used to compute the feedback between aerosols and dynamic while the GEOS‐Chem chemistry module, run alongside the GOCART module, does not feedback to model physics. Currently, there is no chemical coupling between GOCART and GEOS‐Chem, and no observations are directly assimilated into GEOS‐CF. However, stratospheric O₃ in GEOS‐Chem is nudged toward O₃ produced by GEOS FP, which is constrained by ozone measurements from the Microwave Limb Sounder (MLS), Ozone Monitoring Instrument (OMI), and NASA's Ozone Mapping and Profiler Suite (OMPS) and produces a realistic analysis of O₃ in the stratosphere (Wargan et al., 2015, 2020). In addition, near real‐time MODIS observations of fire radiative power are used to constrain fire emissions, as produced by the Quick Fire Emissions Dataset (QFED) (Darmenov and Da Silva, 2015). For the forecasts, the latest available QFED estimates are assumed to persist over the entire forecasting period.

All computations are conducted on the Discover supercomputing cluster of the NASA Center for Climate Simulation (NCCS). Run on 3510 Intel Xeon Haswell processor cores, the 1‐day hindcast and 5‐days forecast takes approximately 8.5 wall‐clock hours in total. GEOS‐CF hindcast and forecast output includes chemistry and meteorology “surface” output every 15 min as well as hourly‐average and instantaneous fields for surface, column‐average, and 3‐dimensional model output. The model output is publicly available at https://gmao.gsfc.nasa.gov/weather_prediction/GEOS-CF/data_access/in the form of on‐demand figures or through access to the model output (in netCDF data format) via Hypertext Transfer Protocol (HTTP) file download or through the Open‐source Project for a Network Data Access Protocol (OPeNDAP) remote access tool. The full 5‐days model forecast output is publicly available for a duration of 14 days. Given the growing interest in air quality forecasting applications, the model forecasts for a selection of surface air pollutants are made available on the public portal indefinitely. Full details on available output and data access are available in the GEOS‐CF File Specification document (Knowland et al., 2020).

2.2. Chemistry

GEOS‐CF v1.0 uses the continually updated standard version of the GEOS‐Chem chemistry module to simulate coupled aerosol‐oxidant chemistry in the troposphere and stratosphere. The results presented here use GEOS‐Chem version 12.0.1 in order to have a two‐year record for comparison to observations, while the latest version 13.0 has been implemented into GEOS for a future version of GEOS‐CF. GEOS‐Chem is ESMF‐compliant and its chemistry module is implemented here as an ESMF gridded component of GEOS, as described in Long et al. (2015) and Hu et al. (2018).

The gas‐phase chemistry scheme includes detailed tropospheric chemistry of HO_x, NO_x, BrO_x, volatile organic compounds (VOC), and O₃, as originally described by Bey et al. (2001), with addition of halogen chemistry by Parrella et al. (2012) and Sherwen et al. (2016) plus updates to isoprene oxidation as described by Mao et al. (2013) and Miller et al. (2017). Stratospheric chemistry is fully coupled with tropospheric chemistry through the Unified tropospheric‐stratospheric Chemistry eXtension (UCX; Eastham et al., 2014) and extends to the top of the atmosphere. Photolysis rates are computed by GEOS‐Chem using the Fast‐JX code (Bian and Prather, 2002). The gas‐phase mechanism comprises 250 chemical species and 725 reactions and is solved using the Kinetic PreProcessor (KPP) Rosenbrock solver (Sandu and Sander, 2006). The aerosol simulation includes sulphate‐nitrate‐ammonia chemistry (Park et al., 2004), black carbon (Park et al., 2003; Wang et al., 2014), organic aerosols (Marais et al., 2016), mineral dust (Fairlie et al., 2007; Ridley et al., 2012), and sea salt aerosols (Jaeglé et al., 2011). Aerosol and gas‐phase chemistry interact through gas‐aerosol partitioning (Fountoukis and Nenes, 2007; Pye et al., 2009), heterogeneous chemistry on aerosol surface (Evans and Jacob, 2005; Mao et al., 2013), and aerosol impacts on photolysis (Martin et al., 2003). Methane concentrations are prescribed as monthly mean surface concentrations, spatially interpolated from NOAA GLOBALVIEW flask data (Dlugokencky et al., 1994).

2.3. Emissions and Deposition

The dry deposition scheme in GEOS‐Chem is based on the resistance‐in‐series model of Wesely (1989), as implemented by Wang et al. (1998). Wet deposition of aerosols and soluble gases includes scavenging in convective updrafts, in‐cloud rainout, and below‐cloud washout (Amos et al., 2012; Liu et al., 2001). While these processes are not fully coupled to GEOS, they are all driven by instantaneous GEOS model diagnostics and thus directly respond to changes in the model environment.

All emission calculations are done using the Harmonic Emissions Component HEMCO v2.1.009 (Keller et al., 2014). Table 1 summarizes the emission configuration used by GEOS‐CF v1.0. The emission inputs used by GEOS‐CF are either calculated online based on current conditions (e.g., lightning NO_x, dust emissions, sea salt), or obtained from pre‐computed emission inventories. For anthropogenic emissions, the Hemispheric Transport of Air Pollution emissions inventory (HTAP v2.2; Janssens‐Maenhout et al., 2015) is used as the primary source due to its high spatial resolution (0.1°) and the availability of monthly emissions data (as opposed to annual emission averages). HTAP v2.2 provides harmonized global emission maps combining global emissions data from the Emission Database for Global Atmospheric Research (EDGAR) (Crippa et al., 2018) with regional emission inventories over Europe, Asia, and the US. For Africa, HTAP v2.2 anthropogenic emissions are overwritten with the DICE emissions inventory (Marais and Wiedinmyer, 2016) to take full advantage of local information. Anthropogenic emissions are further broken down into hourly values using sector‐specific day‐of‐week and diurnal scale factors (Denier van der Gon et al., 2011). For CO, the HTAP v2.2 base emissions, which are representative for year 2010, are adjusted to more recent conditions by applying an annual gridded scale factor based on the latest available Open‐source Data Inventory for Anthropogenic CO₂ (ODIAC; Oda & Maksyutov, 2017), assuming the same relative change in anthropogenic CO emissions as for CO₂. This results in an increase in anthropogenic CO emissions by a factor of 1.2–1.4 over China, India, and South America and decreases by 5%–10% over Europe and North America for year 2018 (relative to 2010). For other anthropogenic emissions, no attempts are currently made in GEOS‐CF to adjust the HTAP v2.2 base emissions. This is an obvious simplification, especially for species and regions undergoing rapid emission changes, such as SO₂ and NO₂ over China or India over the last decade. Incorporation of real‐time information into the emissions calculation is currently under development and slated to be included in future versions of GEOS‐CF.

Table 1.

Emissions Used by GEOS‐CF

Description	Reference	Comments
Offline inventories
Anthropogenic NO, CO, black carbon (BC), organic carbon (OC), Ammonia (NH3)	HTAP v2.2 (Janssens‐Maenhout et al., 2015)	Global except Africa
Anthropogenic SO2	OMI‐HTAP (Liu et al., 2018)	Global except Africa
Anthropogenic VOCs	RETRO (Schultz et al., 2008)	Global except Africa
Anthropogenic NO, CO, SO2, BC, OC, NH3, VOCs	DICE‐Africa (Marais and Wiedinmyer, 2016)	Africa
Arctic seabird NH3	Croft et al. (2016)
Volcanic SO2	Carn (2019)	5% of the sulfur emitted as SO4
Aircraft NOx (=NO + NO2), CO, SO2, VOCs, BC, OC	AEIC (Stettler et al., 2011)
Emissions calculated online based on real‐time environment
Biogenic VOCs	MEGAN v2.1 (Guenther et al., 2012)
Biomass burning (wildfires) NOx, CO, SO2, VOCs, BC, OC	QFED v2.5 (Darmenov and da Silva, 2015)	35% emitted between 3.5 and 5.5 km altitude (Fischer et al., 2014).
Lightning NOx	Murray et al., 2012
Soil NOx	Hudman et al., 2012
Soil dust	Zender et al., 2003
Sea salt aerosols	Gong, 2003; Jaeglé et al., 2011
Oceanic DMS, CH2O, C3H6O	Johnson, 2010; Nightingale et al., 2000
Oceanic iodine	Carpenter et al., 2013

2.4. Observations Used for Model Evaluation

GEOS‐CF is intended to supplement NASA's broad range of space‐based and in‐situ observations, providing a new tool for researchers, government scientists, and air quality managers. We therefore focus our evaluation on the species most pertinent to these applications, including O₃, NO₂, CO, SO₂, and aerosols. These species are also constantly evaluated by the broader GEOS‐Chem community using the standard CTM simulations (e.g., Hu et al., 2017, 2018; Travis & Jacob, 2019). However, it should be noted that GEOS‐CF simulations can differ from standard GEOS‐Chem simulations due to the higher horizontal resolution (Hu et al., 2018), the use of online meteorological information instead of time‐averaged meteorological fields (Yu et al., 2018), and differences in the emission inputs, in particular wildfire emissions from QFED versus the Global Fire Emissions Database (GFED) typically used in the GEOS‐Chem CTM (Carter et al., 2020).

We first evaluate the ability of the GEOS‐CF hindcast to provide a realistic representation of atmospheric composition based on the hourly‐average hindcast fields (Section 4). The validation covers the two‐year time period from the start of the GEOS‐CF record on 1 Jan 2018 to 31 Dec 2019. Differences between the 5‐days model forecasts and the model hindcast are discussed in Section 5. All aerosol comparisons are for the aerosol fields calculated by the GEOS‐Chem module, thus providing aerosol information that is consistent with the gas‐phase chemistry. A detailed comparison of GOCART AOD and PM_2.5 is given in Buchard et al. (2017).

The data sets used for model validation are summarized in Table 2. Briefly, we evaluate the global tropospheric distribution of O₃ against ozonesonde observations obtained from the World Ozone and Ultraviolet Data Center (WOUDC, http://www.woudc.org) and the NOAA Earth System Research Laboratory—Global Monitoring Division (ftp://ftp.cmdl.noaa.gov/ozww/Ozonesonde/), NO₂ against the OMI NASA standard tropospheric column NO₂ product v4.0 (Lamsal et al., 2021) and CO total columns against satellite retrievals from the Measurements Of Pollution In The Troposphere (MOPITT) V8 (Deeter et al., 2019). All satellite comparisons use model output sampled at satellite overpass time and with averaging kernels applied to it. For AOD, we compare the model simulation against the Aqua MODIS combined dark target and deep blue aerosol level 2.0 product (https://modaps.modaps.eosdis.nasa.gov/services/about/products/c6/MYD04_L2.html; Remer et al., 2005; Levy et al., 2010, 2015) as well as Aerosol Robotic Network (AERONET) level 2.0 daily average data from the version three data product (https://aeronet.gsfc.nasa.gov/; Giles et al., 2019). Simulated surface concentrations are compared against in‐situ observations from the Global Atmospheric Watch (GAW) World Data Center for Greenhouse Gases (WDCGG, https://gaw.kishou.go.jp/) and World Data Center for Reactive Gases (WDCRG, https://www.gaw-wdcrg.org/), as well as observations from the OpenAQ database (https://openaq.org). We only include OpenAQ sites with at least 80% of data coverage during the analyzed 2‐year period. The analysis of surface concentrations is based on daily mean values. For NO₂, we omit mountainous GAW sites given that the model resolution of 25 × 25 km² is not high enough to resolve the fine‐scale vertical gradients around mountain slopes. On average, the depth of the GEOS‐CF model surface layer is 130 meters and we use this value without attempting to adjust for sub‐grid vertical gradients. Accounting for the subgrid vertical gradient can improve model‐observation comparisons for ozone (Travis & Jacob, 2019) but is more challenging to implement for species with both sources and sinks, such as NO₂ or PM_2.5. All aerosol evaluation is based on the GEOS‐Chem aerosols, and model PM_2.5 is calculated for a relative humidity (RH) of 35%, following standard practice in GEOS‐Chem.

Table 2.

Overview of Observation Data Sets Used for GEOS‐CF Model Validation

Description	Species	# Sites	Reference
Ozonesonde	O3	24	Tilmes et al., 2012; Thompson et al., 2017
NASA OMI NO2 v4.0	Tropospheric column NO2	global	Lamsal et al., 2021
MOPITT v8	Total column CO	global	Deeter et al., 2019
MODIS	AOD at 550 nm	Global	Remer et al., 2005; Levy et al., 2010, 2015
AERONET	AOD at 550 nm	195	Giles et al., 2019
GAW WDCGG	CO	54	https://gaw.kishou.go.jp/
GAW WDCRG	O3, NO2, SO2	48 (O3), 6 (NO2), 9 (SO2)	https://www.gaw-wdcrg.org/
OpenAQ	O3, NO2, SO2, PM2.5	3151 (O3), 2789 (NO2), 1221 (SO2), 2667 (PM2.5)	https://openaq.org

Figure 2 shows the global distribution of all in‐situ observations used in the evaluation. The GAW observation sites (WDCGG and WDCRG) are located away from the major pollution sources, representing the global background concentrations. In contrast, most observations available on the OpenAQ platform are located in densely populated areas, providing information about local air pollution. With a model grid size of 25 × 25 km², GEOS‐CF is not expected to resolve many of the local features observed at monitoring sites available on OpenAQ. However, the use of model forecasts for local air quality applications is a popular request and the OpenAQ comparisons shown here are intended to serve as a benchmark for such types of applications. We discuss in Section 5 how bias correction methods can be used to overcome some of the model representativeness errors. It should also be emphasized that the observations provided on OpenAQ are reported in near real‐time and thus do not undergo the same level of quality control as the GAW observations. While this makes OpenAQ an excellent platform to monitor a system in near real‐time, it can introduce analysis errors related to observation errors that can be difficult to quantify. In an ad‐hoc attempt to exclude obviously errorenous observations (such as ozone concentrations of several 1000 ppbv), we remove all OpenAQ observations that are 3 standard deviations beyond the mean monthly value of that site. This results in a removal of less than 0.1% of all OpenAQ observations.

Figure 2.

Overview of observation sites used for model validation. Black boxes show the six regions used for aggregation of OpenAQ and Aeronet observations.

The OpenAQ and AERONET observations are further grouped into six geographical regions, as shown in Figure 2. This subset of observations provides good coverage for North America, Europe, and Eastern Asia (especially China) but a limited view of other regions such as South America or Africa. For those under‐represented regions, our analysis relies heavily on the global daily satellite observations.

3. Evaluation of Model Hindcast

3.1. Model Skill Scores

We first highlight the model skill scores against all surface observations before discussing the individual species in more detail below. Figure 3 shows monthly model skill scores for O₃, NO₂, SO₂, CO, and PM_2.5 and/or AOD, aggregated by observation type across all measurement sites. The skill scores used are the normalized mean bias (NMB), normalized root mean square error (NRMSE), and Pearson correlation coefficient (R):

$$\text{NMB}=\frac{{\mathrm{\Sigma }}_{n=1}^N\left(y_n-o_n\right)}{\overline{o}}$$ (1)

$$\text{NRMSE}=\frac{\sqrt{\frac{1}{N}{\mathrm{\Sigma }}_{n=1}^N{\left(y_n-o_n\right)}^2}}{o_{0.95}-o_{0.05}}$$ (2)

$$R=\frac{{\mathrm{\Sigma }}_{n=1}^N\left(y_n-\overline{y}\right)\left(o_n-\overline{o}\right)}{\sqrt{{\mathrm{\Sigma }}_{n=1}^N{\left(y_n-\overline{y}\right)}^2}\sqrt{{\mathrm{\Sigma }}_{n=1}^N{\left(o_n-\overline{o}\right)}^2}}$$ (3)

where y _n denotes an indidivual model estimate, $\overline{y}$ is the model average, o _n is the observation associated with y _n, and $\overline{o}$ is the observation average; o _0.05 and o _0.95 are the 5th and 95th percentile, respectively, of the observations sample, and N is the total number of observation‐model pairs. All skill scores are calculated using daily average concentrations. This removes the impact of the diurnal cycle, which otherwise can dominate the correlation signal. Differences between night‐time and day‐time skill scores are small on average (not shown).

Figure 3.

Overview of NMB, NRMSE and R for daily mean O₃, NO₂, SO₂, CO, and PM_2.5 (OpenAQ only) and AOD (Aeronet only). Shown are the monthly skill score means, as calculated from daily mean concentrations, aggregated by observation type. Note. the different scale for SO₂, PM_2.5 and AOD.

For daily average surface O₃, the model shows a normalized mean annual bias of 0.011 compared to the GAW sites, with little seasonal variability (Figure 3, first column). The GAW sites are typically located in remote locations and the small model bias indicates that the model adequately captures atmospheric background O₃ throughout the year. This is confirmed by the low NRMSE of 0.13 and the correlation coefficient of 0.71, again with small seasonal variability. Compared to the GAW sites, the simulated O₃ shows weaker skill scores relative to the OpenAQ observations with an NMB of 0.14, an NRMSE of 0.61 and R of 0.45. A lower model skill against the observations from the OpenAQ data portal is not surprising given that OpenAQ sites are typically located close to polluted sites, and the model representativeness error is therefore expected to be higher. As further discussed below, the OpenAQ comparisons also highlight a pronounced model overestimation of surface O₃ during the summer months.

For NO₂ (second column of Figure 3), the normalized model bias is 0.11 compared against the GAW sites, and −0.32 against the OpenAQ sites. As for O₃, the larger differences between GEOS‐CF and OpenAQ observations (compared to GEOS‐CF vs. GAW) can likely be attributed to the spatial representation mismatch between the 25 × 25 km² model output and the OpenAQ observations. The NMB at the GAW sites varies strongly from month‐to‐month, reflecting the small sample size (6 stations). The OpenAQ bias shows the strongest negative bias between November through January, with a reduction of the negative bias from −0.4 to −0.25 during June–August. Given that the OpenAQ sites are predominantly located in the NH (Figure 2), this reflects a stronger model underestimation during winter. The NRMSE is 0.15 at the GAW sites and 0.64 at the OpenAQ sites, similar to O₃. The R values for NO₂ are 0.6 and 0.34 at the GAW and OpenAQ sites, respectively. Analysis of NO₂ is complicated by the potential NO₂ observation interference with other nitrogen compounds (Grosjean and Harrison, 1985; Steinbacher et al., 2007; Winer et al., 1974), as well as the short atmospheric lifetime of NO₂, which makes it challenging for the model to capture the spatiotemporal variability in surface NO₂ in full detail. This explains some of the low model bias for NO₂ relative to OpenAQ observations, which are commonly located in urban locations close to combustion sources.

For SO₂, the model—observation comparison shows a large model bias of up to a factor of 4, with an annual mean bias of 2.1 at GAW sites and 1.5 at OpenAQ locations. The NRMSE is 0.47 for the GAW observations and 0.37 for OpenAQ data, and the correlations are similar with 0.29 and 0.25 for GAW and OpenAQ, respectively. An important factor for the model overestimation is the use of outdated anthropogenic SO₂ emissions data, as further discussed below.

The simulated CO compares well against the 54 GAW sites, with an NMB of −0.042, an NRMSE of 0.096 and R of 0.73. The model error is largest during the NH summer months, with a decrease in NMB to −0.14, an increase in NRMSE to 0.2, and a reduction of R to 0.5.

The model comparisons against AERONET AOD and OpenAQ PM_2.5 observations reveal a systematic model overestimation of aerosol concentrations in GEOS‐CF, with an average NMB of 0.48 for AOD and 1.5 for surface PM_2.5. Excessive sulphate aerosol as well as known model issues in the treatment of model emissions and wet scavenging contribute to these biases, as will be discussed in more detail below. In addition, some PM_2.5 observations on the OpenAQ platform represent dry particulate matter while model PM_2.5 assumes 35% RH, which also contributes to the mismatch. At the AERONET sites, the model shows a good correlation score of 0.71 and the NMRSE of 0.15 is comparable to that of other species at background sites, indicating that the model captures the relative changes in AOD well despite the high model bias. For surface PM_2.5, the average model NRMSE is 1.65 and R skill score is 0.46.

3.2. Ozone

Comparisons of GEOS‐CF against ozonesondes and GAW surface observations demonstrate that GEOS‐CF reproduces well‐known features of background surface O₃, such as the local maxima in the NH during spring.

Figure 4 shows annual mean (2018–2019) vertical O₃ profiles as observed by ozonesondes at 24 locations across the world (purple crosses in Figure 2), compared against corresponding GEOS‐CF model concentrations. While GEOS‐CF generally captures the observed vertical structure of O₃, the model tends to underestimate free tropospheric O₃ (approx. 800–300 hPa) over the NH midlatitudes. For these profiles (e.g., De Bilt, Payerne, Madrid, Tateno), the simulated vertical gradient is much less pronounced than observed and the model consistently underestimates O₃ concentrations. This pattern is consistent with previous comparisons of the online GEOS‐Chem module against observations (Hu et al., 2018) and implies two potential model errors: (1) excessive vertical mixing resulting in an overestimation in the lower altitudes and underestimation at higher altitudes, as evident over Payerne, Trinidad Head, and Madrid (Figure 4); and (2) an underestimation of lightning NO_x over the NH midlatitudes, resulting in an underestimation of O₃ production in the upper troposphere. The global lightning NO_x source in GEOS‐CF is 5.9 Tg N yr^‐1, in good agreement with other studies (Schumann and Huntrieser, 2007) and the 6.0 Tg N yr^‐1 reported for the GEOS‐Chem CTM (Murray et al., 2012). However, due to the real‐time nature of the system, the GEOS‐CF lightning parameterization uses the unconstrained cloud top height parameterization (Price and Rind, 1992, 1993, 1994) and does not apply time‐dependent, regional redistribution factors based on Lightning Imaging Sensor (LIS) and the Optical Transient Detector (OTD) satellite observations as is standard in GEOS‐Chem. As described in Murray et al. (2012), this results in an underestimation of simulated lightning flash rates over the NH extratropics, which is likely one of the main reasons for the model underestimation of O₃ in the NH upper troposphere. We also note that the model vertical resolution is approximately 500 m (20 hPa) in the mid‐to upper troposphere, which might be insufficient to resolve the strong vertical gradients across the tropopause boundary and contribute to the model‐observation mismatches in the upper troposphere. Over the Southern Hemisphere (SH), the simulated O₃ profiles are in good agreement with the ozonesonde observations and show an improved O₃ distribution compared to the offline GEOS‐Chem model (Hu et al., 2018).

Figure 4.

Vertical profile of annual average O₃ (2018–2019) at 24 sites as observed by ozonesondes (black) and simulated by GEOS‐CF (magenta). The gray shaded area indicates sonde observation standard deviation.

The GAW observations show a pronounced seasonal cycle for surface O₃ with a maximum of 30–50 ppbv during spring and summer and a low of 10–30 ppbv in winter, a feature that is well reproduced by GEOS‐CF (Figure 5). At remote locations such as the high‐latitude NH GAW sites (Figure 5 top row), the seasonal cycle is more distinct with a peak in early spring. The O₃ maximum develops more slowly at locations that are more heavily influenced by anthropogenic emissions. As already observed in the ozonesondes, the model underestimates O₃ over the Southern Ocean by up to 10 ppbv. This is likely due to excessive O₃ deposition over seawater (Pound et al., 2020), a problem since corrected in newer versions of GEOS‐Chem.

Figure 5.

Monthly average surface O₃ as observed at 48 GAW sites (black) and simulated by GEOS‐CF (magenta). Vertical bars represent the standard deviation of hourly variability. Y‐axis ranges vary by station.

Figure 6 shows monthly average surface O₃ by region, as captured by observations in the OpenAQ database and simulated by GEOS‐CF. The model generally overestimates surface O₃ over Europe, North and South America and Australia by 5–10 ppbv and underestimates it over Africa and Asia by up to 20 ppbv. The model‐observation mismatch is larger than for the remote GAW sites, suggesting that sub‐grid factors such as vertical and horizontal model representation errors and nighttime loss of O₃ by reaction with NO contribute to this difference (Dacic et al., 2020; Travis & Jacob, 2019). Global models tend to underestimate O₃ nighttime depletion due to difficulties to resolve small‐scale horizontal NO₂ gradients, and evaluation of surface O₃ is thus often restricted to daytime or afternoon O₃ (Millet et al., 2015; Hu et al., 2018; Travis & Jacob, 2019). As shown in Figure 7, restricting the analysis to daytime O₃ (12:00–16:00 local time) does indeed improve the comparison and reduce the bias by up to 5 ppbv, in particular over Europe, South America and Australia. The model still shows a systematic positive bias over the US during summer and fall, a known issue in GEOS‐Chem (Hu et al., 2018; Travis et al., 2016; Travis & Jacob, 2019 ). Results for the mean values of the daily maximum 8 h average (MDA8) instead of afternoon O₃ yield the same qualitative result (not shown).

Figure 6.

Monthly average surface O₃ grouped into six regions (see Figure 2) as obtained from the OpenAQ database (black) and simulated by GEOS‐CF (magenta). Vertical bars represent the standard deviation of hourly variability across all sites. Number of sites is given in the inset. Y‐axis ranges vary by region.

Figure 7.

As Figure 6 but using afternoon O₃ only (12:00‐16:00 local time).

Factors that likely contribute to the high NH surface O₃ in GEOS‐CF are uncertainties in the production of O₃ from the oxidation of isoprene (Bates & Jacob, 2019; Travis et al., 2016) and errors in O₃ deposition to wet surfaces (Travis & Jacob, 2019). In addition, most OpenAQ observation sites are located near densely populated areas, and the 25 × 25 km² model simulation cannot fully capture the fine‐scale features characteristic for these environments (Keller et al., 2020).

3.3. Nitrogen Dioxide

As shown in Figure 8, GEOS‐CF captures major features of the global distribution of tropospheric NO₂. The model simulated NO₂ columns (Figure 8a) show distinct hot spots over urban areas (e.g., Eastern China, Europe, and Eastern US), reflecting the dominant contribution of fossil fuel combustion coupled with the short atmospheric lifetime of NO₂ (Duncan et al., 2016; Streets et al., 2013;). Additional hot spots resulting from biomass burning are found over Africa and the boreal region of Canada (Figure 8a). The spatial pattern simulated by GEOS‐CF agrees well with the NO₂ columns observed by OMI. Over Eastern China, the model simulated NO₂ columns are up to 3.0 × 10¹⁵ molecules cm^‐2 (or approx. 40%) higher than the OMI observations (Figure 8b), suggesting a potential overestimation of NO_x emissions or a longer NO_x lifetime in the model (Shah et al., 2020). However, the OMI retrieval algorithm v4.0 tends to underestimate tropospheric NO₂ over polluted areas (Lamsal et al., 2021), which complicates the analysis. As further discussed in the next section, the comparison against surface observations does not support the view of a broad‐based overestimation of surface NO₂ over Asia.

Figure 8.

Comparison of GEOS‐CF against OMI tropospheric column NO₂. (a) the 2018–2019 average tropospheric NO₂ column (1–2 pm local time) as simulated by GEOS‐CF. (b) The difference between model simulated tropospheric column and NASA OMI NO₂ (v4.0) observations. (c) Right panel shows the tropospheric NO₂ time series averaged for six regions, as shown in (a) and (b) (n.b., y‐axis intervals are not the same for each time‐series).

As shown in Figure 8c, the simulated seasonality of tropospheric NO₂ columns is in good agreement with OMI observations. Over areas dominated by anthropogenic activities, such as the US, Europe, and China, the simulated NO₂ columns show a distinct seasonal cycle with the minimum during summer and peak during winter, driven by the seasonal variation in NO_x lifetime against oxidation (Shah et al., 2020). The seasonal cycle observed over China is well captured by GEOS‐CF, while the simulated wintertime peak over the US and Europe is higher than observed by OMI. Chemical loss of NO_x during winter is dominated by N₂O₅ hydrolysis, which is sensitive to NO_x emissions and O₃ concentrations (Jaeglé et al., 2018; Shah et al., 2020). Over Africa, the seasonal cycle is dominated by summer biomass burning, which is well captured by the model.

Figure 9 shows comparisons of monthly mean surface NO₂ at six GAW sites (n.b., all located in Europe) against model simulated concentrations. These comparisons show that GEOS‐CF is in good agreement with the (non‐mountainous) GAW sites, suggesting that it provides a realistic simulation of background surface NO₂ over Europe. Figure 10 further evaluates the simulated distribution of global surface NO₂ in comparison to observations in OpenAQ. The model is in excellent agreement with observations over North America and Africa, and Asia during the summertime, but underestimates concentrations over Europe and South America, as well as Asia during the wintertime. The apparent low bias over Europe and Asia is inconsistent with the tropospheric column comparisons shown in Figure 8 and also the comparison against the GAW observations (Figure 9), which do not show such a systematic underestimation of NO₂ by GEOS‐CF.

Figure 9.

Monthly average surface NO₂ as observed at six GAW sites (black) and simulated by GEOS‐CF (green). Vertical bars represent the standard deviation of hourly variability. Y‐axis ranges vary by station.

Figure 10.

Monthly average surface NO₂ grouped into six regions (see Figure 2) as obtained from the OpenAQ database (black) and simulated by GEOS‐CF (green). Vertical bars represent the standard deviation of hourly variability across all sites. Number of sites is given in the inset. Y‐axis ranges vary by region.

The comparison of surface concentrations of NO₂ is complicated by the fact that most surface observations are based on chemiluminescence using a molybdenum converter, which can result in an overestimation of reported NO₂ concentrations due to interference with other oxidized nitrogen compounds (Grosjean and Harrison, 1985; Steinbacher et al., 2007; Winer et al., 1974). This might explain some of the model underestimation of NO₂ relative to the OpenAQ observations. In addition, since the OpenAQ observations tend to be located in relative proximity to urban areas they often do not represent the regional average NO₂ concentrations captured by GEOS‐CF, which can introduce a systematic bias. While this is an issue for all analyzed species, the problem is particularly pronounced for NO₂ given its large spatial and temporal variability.

3.4. Sulfur Dioxide

GEOS‐CF overestimates surface SO₂ concentrations compared against observations from the GAW network (Figure 11) as well as the OpenAQ database (Figure 12). The GAW sites are all located in Europe, and the model has a mean bias of 0.9 ppbv across the nine stations. Against the low average background concentration of 0.5 ppbv observed at the GAW sites, this indicates an overestimation by the model by a factor of more than two. A similar bias is found at the OpenAQ sites. We attribute the high model bias in SO₂ to a large extent the use of outdated emissions from the OMI‐HTAP emission inventory. Anthropogenic emissions in OMI‐HTAP are based on HTAP v2.2, which is representative for year 2010 (Liu et al., 2018). Since then, anthropogenic emissions of SO₂ have declined significantly in most parts of the world (Feng et al., 2020; Fioletov et al., 2016; Li et al., 2020; Qu et al., 2019). The model overestimation is particularly severe during wintertime, when the atmospheric lifetime of SO₂ is longest (Lee et al., 2011) and the majority of anthropogenic SO₂ is emitted (according to OMI‐HTAP). Another likely explanation is that the model may be under‐accounting for non‐photochemical SO₂ oxidants in winter (Shao et al., 2019).

Figure 11.

Monthly average surface SO₂ as observed at nine GAW sites (black) and simulated by GEOS‐CF (blue). Vertical bars represent the standard deviation of hourly variability. Y‐axis ranges vary by station.

Figure 12.

Monthly average surface SO₂ grouped into six regions (see Figure 2) as obtained from the OpenAQ database (black) and simulated by GEOS‐CF (blue). Vertical bars represent the standard deviation of hourly variability across all sites. Number of sites is given in the inset. Y‐axis ranges vary by region.

3.5. Carbon Monoxide

Our analysis of simulated CO shows that the model has no significant global CO bias compared against satellite and surface observations. Figure 13 shows the global distribution and seasonal cycle of total column CO retrieved from MOPITT and derived from GEOS‐CF. Local MOPITT averaging kernels were applied to the GEOS‐CF CO profiles to obtain the column values. The model simulated CO pattern is in good agreement with MOPITT, with local maxima over major polluted areas (e.g., East China and India) and over biomass burning regions (South America and Central Africa). Similar to NO₂, the simulated CO columns over China are larger than the observations, possibly due to an overestimation of CO emissions over that region. The baseline anthropogenic CO emissions in GEOS‐CF are scaled based on ODIAC emission trends, with strongest increases over China and India (Oda et al., 2018). Our scaling methodology assumes a constant CO/CO₂ emissions ratio, and any decrease in the CO/CO₂ emission ratio, for example, due to a technology shift, would result in an overestimation of CO emissions.

Figure 13.

Comparison of GEOS‐CF against MOPITT V8 total column CO. Top left panel shows the 2018–2019 average total CO column as simulated by GEOS‐CF. Bottom left panel shows the difference between model simulated total column and MOPITT observations. Right panel shows the total CO time series averaged for six regions, as shown on the left.

The seasonal cycle of CO is determined by its shorter chemical lifetime during summer due to increased photochemical activity, which is reflected in lower NH concentrations during July and August where the model underestimates the MOPITT concentrations in the NH middle and high latitudes by 10%–20% (Figure 13c). This underestimation is driven by a stronger than observed decrease of simulated total column CO during summertime, a pattern that is confirmed by comparisons against the GAW surface observations (Figure 14). While the model underestimates summertime surface CO in the NH, the opposite is true for the SH where model simulated concentrations during summer are higher than the observations. A low bias in CO is a long‐standing issue in atmospheric chemistry models, commonly attributed to an underestimation of direct emissions, differences in the CO yield from VOC oxidation, and inconsistencies in the simulated distribution of the hydroxyl radical OH, the main atmospheric oxidant (Flemming et al., 2015; Gaubert et al., 2016; Monks et al., 2015; Shindell et al., 2006; Strode et al., 2015). The air mass‐weighted mean tropospheric OH in GEOS‐CF is 11.9 × 10⁵ molecules cm^˗3, in good agreement with other models as well as estimates derived from methyl chloroform observations (Montzka et al., 2000; Naik et al., 2013; Spivakovsky et al., 2000). The inter‐hemispheric (NH/SH) ratio is 1.33, again in good agreement with other model estimates (Naik et al., 2013) but higher than observation‐derived values of 0.8–1.0 (Krol and Lelieveld 2003; Montzka et al., 2000; Patra et al., 2014; Prinn et al., 2001). This is consistent with an underestimation of summertime CO in the northern high‐latitudes—as shown by the comparisons against MOPITT and the GAW surface observations (Figures 13 and 14)—and an overestimation of SH CO (Figure 14). An overestimation of NH OH in GEOS‐CF is also supported by inversion studies using the GEOS‐Chem CTM, which find that a similar OH correction is needed to match methane observations (Zhang et al., 2020).

Figure 14.

Monthly average surface CO as observed at 54 GAW sites (black) and simulated by GEOS‐CF (red). Vertical bars represent the standard deviation of hourly variability. Y‐axis ranges vary by station.

3.6. Aerosols

The evaluation of model simulated O₃, NO₂ and CO has shown that GEOS‐CF reproduces many of the features of the tropospheric distribution of these trace gases. With respect to aerosols simulated by GEOS‐Chem, our validation shows that GEOS‐CF has a high bias but still captures many of the observed spatial and temporal patterns. Some of the high aerosol bias can be attributed to the high SO₂ in GEOS‐CF (Section 4.4), resulting in excessive formation of sulphate. In addition, a high bias in aerosols is a known issue in GEOS‐Chem v12.0.1 used in GEOS‐CF v1.0, and recent versions of GEOS‐Chem show improved simulation of aerosols including surface PM_2.5 (Carter et al., 2020; Luo et al, 2019, 2020).

Figure 15 shows the global distribution and seasonality of simulated AOD at 550 nm against Aqua MODIS satellite observations. GEOS‐CF captures the seasonal cycle of AOD well but tends to overestimate the aerosol burden, especially over the US, Europe, and China. These regions have seen large declines in SO₂ emissions since 2010 that are not captured by GEOS‐CF and we attribute the high model AOD at least partly to the excessive formation of sulphate aerosols. Further, inadequate treatment of wet scavenging processes is a known issue in GEOS‐Chem v12.0.1, resulting in an overprediction of aerosol nitrate and ammonium, especially over Asia (Luo et al., 2019, 2020).

Figure 15.

Comparison of GEOS‐CF against MODIS Aerosol Optical Depth (AOD). Top left panel shows the 2018–2019 average total AOD as simulated by GEOS‐CF. Bottom left panel shows the difference between model simulated total column and MODIS observations. Right panel shows the total AOD time series averaged for six regions, as shown on the left.

GEOS‐CF tends to underestimate AOD related to strong wildfires events, such as over Canada during the summers of 2018 and 2019, the US in 2018, or Indonesia in 2019 (Figure 15). This points to a potential underestimation of severe wildfire emissions by QFED, possibly due to the thick smoke obscuring fires from satellite detection (Eck et al., 2019; Liu et al., 2020). However, the QFED emissions inventory has been shown to lead to an overall high aerosol bias when used in GEOS‐Chem as the enhancement factors used in QFED (to match MODIS‐observed AOD) are based on the GOCART model (Darmenov and Da Silva, 2015), which can result in an overestimation of biomass burning emissions if the differences between the GEOS‐Chem and GOCART aerosol parameterizations are not accounted for (Carter et al., 2020). This factor likely contributes to the overestimation of AOD over Africa and South America.

Figure 16 compares model simulated AOD at 550 nm against AERONET observations, confirming the systematic model overprediction of AOD. The model bias is highest over Europe and North America, consistent with an overprediction of sulphate aerosols due to excessive SO₂ emissions in GEOS‐CF. To provide an upper bound on the contribution of the SO₂ error to total AOD, we analyzed GEOS‐CF AOD excluding sulphate. As shown by the dashed orange line in Figure 16, this indicates that sulphate indeed constitutes up to 50% of total AOD over Europe, North America, and Asia, while its contribution to total aerosol is smaller in other regions.

Figure 16.

Monthly average Aerosol optical depth (AOD) at 550 nm grouped into six regions (see Figure 2) as observed at AERONET sites (black) and simulated by GEOS‐CF (orange). Vertical bars represent the standard deviation of daily variability. Dashed line indicates GEOS‐CF AOD when excluding sulphate aerosols. Number of sites is given in the inset. Y‐axis ranges vary by region.

Figure 17 shows model‐observation comparisons for surface PM_2.5. As for AOD, GEOS‐CF generally overestimates surface PM_2.5. On a relative basis, the overprediction is most pronounced for Europe and North America, and the largest absolute bias of surface PM_2.5 is found over Asia.

Figure 17.

Monthly average surface PM_2.5 grouped into six regions (see Figure 2) as obtained from the OpenAQ database (black) and simulated by GEOS‐CF (orange). Vertical bars represent the standard deviation of hourly variability. Number of sites is given in the inset. Y‐axis ranges vary by region.

While we conclude that the main reasons for the high aerosol bias in GEOS‐CF is related to three main factors—excessive sulphate aerosols, an underestimation of aerosol wet scavenging, and overestimated biomass burning emissions outside of high‐intensity wildfire episodes—there are other factors that could contribute to the observed model‐observation mismatches. For instance, a general overestimation of wildfire emissions could directly affect modeled organic aerosol as well as the formation of secondary organic aerosol (SOA). Organic aerosol constitutes an important fraction of total aerosol mass—also in GEOS‐Chem (Kim et al., 2015)—but its sources and atmospheric processes are complex and poorly constrained (e.g., Bates et al., 2019; De Gouw and Jimenez, 2009; Marais et al., 2016). Biases in dust emissions could be another contributor to model biases, especially over Africa. The dust emissions calculated by GEOS‐CF total 770 Tg/yr and 630 Tg/yr for 2018 and 2019, respectively, which is at the lower range of current estimates of 426–2726 Tg/yr (Huneeus et al., 2011) and thus unlikely a major contributor to the high aerosol bias.

4. Evaluation of Model Forecasts

4.1. Comparison of Model Forecasts Against Observations

GEOS‐CF v1.0 does not directly assimilate trace gas observations and differences between the 1‐day hindcast and the model forecasts are thus mainly driven by variations between the forecasted and analyzed meteorological state. The meteorology not only impacts the flow of the constituents but also affects deposition and dynamically calculated emissions, including lightning NO_x, biogenic VOCs, sea salt aerosols, and dust (see Table 1). Further, the model forecasts assume persistence in the biomass burning emissions, meaning the fires observed during the hindcast are assumed to continue burning and emitting the same amount for the next five days. The model thus does not capture changes in occurrence or intensity of wildfires, both of which can lead to significant changes in surface air pollution close to and downwind from the fires.

Figure 18 shows model‐observation skill scores for the model hindcast (forecast day −1) and the 5‐days forecasts (forecast days +1 to +5) for O₃, NO₂, and PM_2.5 relative to OpenAQ and GAW surface observations. Skill scores were calculated at each observation site individually using daily mean concentrations. Similar to the results discussed in Section 4.1 (Figure 3), the skill scores were computed for each month separately before aggregating them in the form of boxplots (as shown in Figure 18) in order to remove the impact of seasonality. For all three evaluated species, the model hindcast shows the best agreement with the observations. The median NMB of the model hindcast is 0.1 for O₃, ‐0.31 for NO₂, and 1.56 for PM_2.5, with almost no difference between the hindcast and the model forecasts. Similarly, the forecasted NRMSE are little changed relative to the hindcast; the median NRMSE for the hindcast is 0.61 for O₃, 0.63 for NO₂ and 1.61 for PM_2.5, and the NRMSE only slightly deteriorates with increasing forecast lead time, resulting in NRMSE's for the 5‐days forecast of 0.62 for O₃, 0.65 for NO₂ and 1.65 for PM_2.5.

Figure 18.

GEOS‐CF model skill scores for the hindcast (forecast day −1) and the 5‐days forecasts (forecast day +1 to +5) relative to surface observations (GAW and OpenAQ). Boxplots show the variation in the NMB, NRMSE, and R across all surface sites for daily mean O₃, NO₂, and PM_2.5.

The largest change in skill score between hindcast and forecasts is found for R, which drops from 0.49 for the hindcast to 0.38 for the 5‐days forecast for O₃, 0.39 to 0.26 for NO₂, and 0.52 to 0.36 for PM_2.5. This indicates that errors in the meteorological forecasts (and biomass burning emissions) indeed impact the quality of the surface air quality forecasts. Of the three analyzed skill scores, R is most sensitive to errors in the temporal pattern and we attribute a large fraction in the deterioration in R to transport errors in the forecasts, such as the evolution of frontal systems or the dispersion of smoke plumes. PM_2.5 is particularly sensitive to these factors given its large spatiotemporal gradients, comparatively long atmospheric lifetime, and strong sensitivity to changes in biomass burning emissions.

4.2. Bias‐Corrected Local Forecasts

As discussed above, model‐predicted concentrations of O₃, NO₂ and PM_2.5 can differ from the observations for a number of reasons, including model representation errors, uncertainties in the meteorology, or model biases arising from errors in the model treatment of emissions, deposition, or atmospheric chemistry. One approach to deal with these issues is to quantify and correct these systematic model errors in a post‐processing step. Such bias correction methods can be applied to near real‐time model forecasts and have been found to be an effective tool to significantly improve local model predictions, for example, by using mean subtraction (McKeen et al., 2005; Wilczak et al., 2006), historical analogs (Hamill and Whitaker, 2006), Kalman‐filtering (Delle Monache et al., 2006; Djalalova et al., 2015), or kriging (Honoré et al., 2008). More recently, machine learning (ML) approaches have become popular to relate model output to air quality observations (e.g., Grange et al., 2018; Grange and Carslaw, 2019; Ivatt and Evans, 2020; Petetin et al., 2020). As discussed in Keller et al. (2020), bias‐correction using ML can significantly reduce GEOS‐CF model biases compared to surface observations. This is illustrated in Figure 19, which shows the GEOS‐CF model skill scores for the same surface observations analyzed above (GAW and OpenAQ) but using bias‐corrected model concentrations instead of the original model output. The bias correction methodology is described in detail in Keller et al. (2020). It uses the XGBoost algorithm (Chen and Guestrin, 2016) to correct the original hourly‐average model predictions of O₃, NO₂, and PM_2.5 based on local meteorology and atmospheric composition, as predicted by the GEOS‐CF model. The input features fed into the ML algorithm include 9 meteorological parameters, 51 chemical species, 21 modeled emissions, and calendar information including hour of the day, weekday, and month of the year. The latter allow the ML model to identify systematic observation‐model mismatches related to the diurnal, weekly, and seasonal cycle of the pollutants. Since model biases are often site‐specific—especially the representation error—a separate ML model is trained for each site. Using the hourly model hindcast—observation pairs of year 2018, the ML models are trained to predict the systematic bias between model predicted pollutant concentrations and the actual observation. We then apply the ML algorithm to the 2019 model output and compare the resulting bias‐corrected model concentrations against observations (Figure 19). It is important to emphasize that the ML correction is site‐specific and varies as a function of local meteorological and chemical conditions, as well as time of the year, weekday, and hour of day. As such, it it designed to provide a “localized” model prediction only representative for the given observation site, not the 25 × 25 km² grid cell as a whole.

Figure 19.

Similar to Figure 18, however the boxplots indicate the GEOS‐CF model skill scores for the bias‐corrected hindcast (forecast day −1) and the bias‐corrected 5‐days forecasts (forecast day +1 to +5) relative to surface observations. The median statistics of the non‐corrected model values from Figure 18 shown as diamonds.

Compared to the uncorrected model output (indicated by the diamonds in Figure 19), the bias‐corrected model values agree better with the observations for all species, skill scores, and lead times. For all three species, the NMB of the bias‐corrected values is close to zero for both the hindcast and the 5‐days forecasts. This result is not unexpected given that the ML algorithm is designed to minimize the mean model bias. The NRMSE and R scores are also significantly improved compared to the original data, with median RMSE values of approximately 0.2 and R values of 0.9 for the hindcast. Relative to the hindcast, the NRMSE and R become incrementally worse for the 1‐day to 5‐days lead forecasts relative to the actually observed values. Also, the skill scores of the bias‐corrected forecasts deteriorate more rapidly than is the case for the uncorrected output but still outperform the original forecasts for all five lead days. Since the ML algorithm was only trained on the hindcast data, model errors in the meteorology forecasts will also negatively impact the quality of the bias correction applied to the (forecast) baseline, which explains the steady decline in the skill scores for the bias‐corrected forecasts.

5. Conclusions

The GEOS‐CF system provides global, near real‐time estimates of current conditions (“hindcast”) and 5‐days forecast simulations of atmospheric composition and meteorology at the global resolution of 0.25° based on the GEOS‐Chem chemistry module run within the GEOS GCM. Comparisons against a suite of satellite, ozonesonde and surface observations demonstrate that the model realistically captures the global distribution of O₃, NO₂, and CO. For these three species, the model shows little biases at background locations, with NMB values ranging between −0.1 to +0.1. Like other atmospheric chemistry models, the model overestimates surface O₃ in the NH during summer, especially over the south eastern US (Emmons et al., 2020; Flemming et al., 2015; Travis et al., 2016). While the horizontal resolution of 0.25° is one of the highest for a global atmospheric chemistry forecast model, it is still not fine enough to capture local‐scale features of urban air pollution, which can lead to substantial model‐observation mismatches when compared against urban air quality observations. This is reflected in the poorer model skill against the OpenAQ observations, which are typically located in the proximity of urban areas. The model representation error is an important limitation of global composition models such as GEOS‐CF and needs to be taken into account for air quality applications. One possible solution to this problem is to apply a bias‐correction algorithm to the gridded model output, based on ML and historical observation‐model comparisons. This can lead to significant improvements of the model predictions. This procedure, which can be automated as a post‐processing step, offers the option to provide highly localized forecasts anywhere in the world where surface observations are available. However, it should be emphasized that such localized forecasts are only valid for the given observation site and do not represent a model grid cell average (25 × 25 km²) anymore.

The 5‐days forecasts produced by the GEOS‐CF system have comparable NMB and NRMSE as the 1‐day hindcast (compared to hourly observations), but show a deterioration in temporal correlation with increasing forecast lead time. The current version of GEOS‐CF (v1.0) does not directly assimilate trace gas observations, and the reduction in correlation skill score is thus directly related to errors in meteorological forecasts—which feed back on chemistry, deposition, and dynamically updated emissions—as well as errors in the wildfire emission forecasts which are assumed to persist over the entire 5‐days forecast window.

GEOS‐CF v1.0 systematically overpredicts SO₂ due to the use of outdated emissions data, highlighting the critical need of timely emission information. In it's current form, GEOS‐CF does not account for real‐time changes in anthropogenic emissions, for example, due to reduced human activities in the wake of the COVID‐19 pandemic. GEOS‐CF v1.0 thus represents a business‐as‐usual estimate of the atmosphere (Keller et al., 2020). The inclusion of near real‐time information for dynamic emission adjustment, for example, based on traffic data or satellite observations, is a focus of future model development for GEOS‐CF.

The GEOS‐CF modeling system leverages components developed by the GEOS and the GEOS‐Chem modeling communities and directly benefits from the continuous advancements provided by these groups. The current version of GEOS‐CF (v1.0) incorporates GEOS‐Chem v12.0.1. Several important updates have been added to GEOS‐Chem since, including updates to the chemistry of isoprene (Bates and Jacob, 2019) and halogens (Wang et al., 2019), improved wet scavenging of aerosols (Luo et al., 2019, 2020), and updated O₃ deposition over seawater (Pound et al., 2020). When implemented in future versions of GEOS‐CF, these updates are expected to reduce the high bias observed in PM_2.5 and AOD as well as the high bias in surface O₃ over the SH.

Another model development focus will center around the assimilation of satellite observations of atmospheric constituents, which has been shown to lead to improved representation of atmospheric composition, in particular for longer‐lived species such as O₃ and CO (Benedetti et al., 2009; Inness et al., 2015). In its current form, GEOS‐CF does not directly assimilate tropospheric trace gas observations, and the variability in constituent distribution is thus driven by the anthropogenic emission inventories, real‐time biomass burning information, and the current meteorological state and its impact on dynamic emission sources such as biogenic emissions or sea salt aerosols.

GEOS‐CF offers a new tool for academic researchers, air quality managers, and the public. Applications include flight campaign planning, support of satellite and other remote‐sensing observations, interpretation of field campaign data (Dacic et al., 2020; Johnson et al., 2021), and air quality research (Keller et al., 2020).

Keller, C. A. , Knowland, K. E. , Duncan, B. N. , Liu, J. , Anderson, D. C. , Das, S. , et al. (2021). Description of the NASA GEOS composition forecast modeling system GEOS‐CF v1.0. Journal of Advances in Modeling Earth Systems, 13, e2020MS002413. 10.1029/2020MS002413

Data Availability Statement

All model output is centrally stored at the NASA Center for Climate Simulation (NCCS). Public access to these archives is provided by the GMAO at https://gmao.gsfc.nasa.gov/weather_prediction/GEOS-CF/data_access/in the form of weather maps and through model output access tools, including OPeNDAP and Hypertext Transfer Protocol (HTTP).