Development and Evaluation of the Aerosol Forecast Member in 2 NCEP’s Global Ensemble Forecast System (GEFS-Aerosols v1)

. NOAA’s National Weather Service (NWS) is on its way to deploy various operational prediction applications using the Unified Forecast System a community-based coupled, comprehensive Earth modeling system. An aerosol 3 model component developed in a collaboration between the Global Air and Environmental Modeling Center (GSL, CSL, ARL, EMC) was coupled online with the 5 FV3 Global Forecast System (FV3GFS) using the National Unified Operational Prediction Capability (NUOPC)-based NOAA 6 Environmental Modeling System (NEMS) software framework. This aerosol prediction system replaced the NEMS GFS 7 Aerosol Component (NGAC) system in the National Center for Environment Prediction (NCEP) production suite in September 8 2020 as one of the ensemble members of the Global Ensemble Forecast System (GEFS), dubbed GEFS-Aerosols v1. The 9 aerosol component of atmospheric composition in GEFS is based on the Weather Research and Forecasting model (WRF- 10 Chem). GEFS-Aerosols includes bulk modules from the Goddard Chemistry Aerosol Radiation and Transport model 11 (GOCART). Additionally, the biomass burning plume rise module from High-Resolution Rapid Refresh (HRRR)-Smoke 12 based on WRF-Chem was implemented; the GOCART dust scheme was replaced by the FENGSHA dust scheme (developed 13 by ARL); the Blended Global Biomass Burning Emissions Product (GBBEPx version 3) provides biomass burning emission and Fire Radiative Power (FRP) data; and the global anthropogenic emission inventories are derived from the Community 15 Emissions Data System (CEDS). All sub-grid scale transport and deposition is handled inside the atmospheric physics routines, which required consistent implementation of positive definite tracer transport and wet scavenging in the physics 17 parameterizations used by NCEP’s operational Global Forecast System based on FV3 (FV3GFS). This paper describes the details of GEFS-Aerosols model development and evaluation of real-time and retrospective runs using different observations from in situ measurement, satellite and aircraft data. GEFS-Aerosols predictions demonstrate substantial improvements for both composition and wet scavenging in the SAS scheme ; 4) the updated background fields of OH, H 2 O 2 and NO 3 from GMI model; 5) biomass- 1 burning emission calculations based on the GBBEPx V3 emission and FRP provided by NESDIS; and 6 ） global 2 anthropogenic emission inventories derived from CEDS and HTAP. This new model is able to forecast the higher-resolution 3 distribution of primary air pollutants of aerosols: black carbon, organic carbon, sulfate, and dust and sea salt each with five 4 size bins.


Introduction
All sub-grid scale transport and convective deposition related to aerosol is handled inside the atmospheric physics routines of 23 simplified Arakawa-Schubert (SAS) scheme, which required consistent implementation of positive definite tracer transport 24 and wet scavenging in the physics parameterizations GFSv15 and subsequent in the forecast system of GEFSv12.  [Grell et al., 2005;Powers et al., 2017], and the first time to be used in the global model is the Flow-following finite-volume 28 Icosahedral Model (FIM), as FIM-Chem [Zhang et al, 2022], including aerosol modules from GOCART. The metrological nitrate, ammonium and secondary organic aerosol (SOA) in GOCART.

24
The real-time forecast experiments were evaluated using the following ensemble analysis, reanalysis data, satellite and in situ 25 observational data, aircraft measurements, and model predictions. We compare each day model forecast hours with same day 26 reanalysis or analysis data and compute the AOD statistics (e.g. bias, RMSE, correlation etc.) for each grid for each pair of 27 model and reanalysis or analysis data for that model forecast hour. We then calculate that for the entire 4 months of the study 28 period and averaged it over the entire 4 months for each grid points. This method gives an overall estimate of systematic bias 29 of the model in spatial and temporal scale.

22
3) The Aerosol Robotic Network (AERONET), which is a global ground-based network of automated sun-photometer 23 measurements, provides AOT, surface solar flux and other radiometric products [Holben et al., 1998]. It is a well-24 established network of over 700 stations globally and its data are widely used for aerosol-related studies [Zhao et al., 25 2002]. AERONET employs the CIMEL sun-sky spectral radiometer, which measures direct sun radiances at eight 26 spectral channels centered at 340,380,440,500,675,870,940 and 1020 nm. AOT uncertainties in the direct sun 27 measurements are within ±0.01 for longer wavelengths (longer than 440 nm) and ±0.02 for shorter wavelengths [Eck 28 et al., 1999]. Table 1    in ICAP-MME, not use to compute ensemble mean in ICAP-MME for total AOD. All four of the multispecies models 25 incorporate aerosol data assimilation (DA) and satellite-based smoke emissions. ICAP-MME is able to provide real-26 time comparison for synchronous evaluation of operational forecast. The correlation and RMSE between ICAP-MME 27 and AERONET indicating in Table 1 shows that ICAP analysis is quite close to observation, which is good to use it 28 as the global evaluated data, especially when the MERRA-2 data is not available in the real-time or operational 29 forecast.

30
2) The NEMS GFS Aerosol Component Version 2.0 (NGACv2) for global multispecies aerosol forecast developed by 31 NCEP and collaborators was previously used to provide operational global multispecies aerosol forecasts at NCEP    The preprocessor PREP-CHEM-SRC v1.7, a comprehensive tool that prepares emission fields of trace gases and aerosols for 13 use in atmospheric chemistry transport models, was used to generate the anthropogenic emissions, background fields of OH, 14 H2O2, NO3, DMS and dust scheme input of clay and sand at the FV3 grid resolution for GEFS-Aerosols [Freitas et al., 2011].

15
Two global anthropogenic emission inventories were chosen as input to drive the model, both providing monthly emissions.

16
One is from the Community Emissions Data System (CEDS), which provides the emissions of BC, OC and SO2 in 2014 with 17 0.5 degree horizontal resolution [Hoesly et al., 2018]. The CEDS inventory improves upon existing inventories with a more 18 consistent and reproducible methodology applied to all emission species, updated emission factors, and more recent estimates 19 in 2014. The data system relies on existing energy consumption data sets and regional and country-specific inventories to 20 produce trends over recent decades [Hoesly et al., 2018]. The Hemispheric Transport of Air Pollution (HTAP) version 2 21 [Janssens-Maenhout et al., 2015] inventory provides the emissions of BC, OC SO2, PM2.5 and PM10 in 2010.  To validate model performance when using the GBBEPx v3 fire emissions with a plume-rise module based on real-time FRP 23 data, we compare the real-time GEFS-Aerosols AOD with other reanalysis data, satellite observations and the NGACv2 model 24 for the big fire event in August 2019. Smoke from large fires burning in the Amazon rainforest, primarily in Brazil, Bolivia, For both satellites, daily gridded AOD is used to compare against the model forecast at 18z. The GEFS-Aerosols AOD is able 28 to reproduce the enhanced AOD due to several fire events over South America near the border of Bolivia, Paraguay, and Brazil, 29 which were also observed by the VIIRS and MODIS satellite instruments and captured by the MERRA2 analysis. Although     Pacific and southern Atlantic. In contrast, the NGACv2 model does not capture these fire events, and exhibits only a very 2 slight AOD enhancement. NGACv2 AOD is more than 80% smaller than the observations over the fire source region and 3 produces little or no transported smoke over the surrounding areas.

4
Beyond the fires burning in South America, an even greater number of blazes on the African continent are observed by the 5 satellite images at almost the same time in August 2019. Angola experienced almost three times more fires than Brazil in the 6 middle of August 2019. There were around 6,000 fires in Angola, more than 3,000 in Congo and just over 2,000 in Brazil, 7 according to NASA satellite imagery (https://earthobservatory.nasa.gov/images/145421/building-a-long-term-record-of-fire).

24
We evaluated the GEFS-Aerosols model performance with the daily AERONET data globally. The locations of the 60 selected 25 AERONET sites where these comparisons were made are listed in Table 1. It also indicates the correlation and root mean 26 square error (RMSE) of GEFS-Aerosols, ICAP and NGACv2 AOD with respect to that of AERONET observation. The GEFS-

27
Aerosols, NGACv2 and ICAP predictions are sampled at the same locations as the AERONET sites for these comparisons.

28
The left panel in Figure 5 (a and b) shows the correlation coefficients between daily total AOD observed by AERONET and 29 the day 1 forecast of model AOD from GEFS-Aerosols and NGACv2 for the period between 7/5/2019 and 11/30/19. The

Moved (insertion) [14]
Formatted: Font color: Text 1 correlation coefficients range from 0.5 to 0.9 for GEFS-Aerosols at most sites, except for several sites in South America, 1 Africa and eastern Asia near fire source regions, which are slightly lower than those of the ICAP. In contrast, the correlation 2 coefficients of daily total AOD between the NGACv2 and AERONET observations are lower than 0.5 globally, even ranging 3 from 0.1 to 0.3 at most sites. A more quantitative display of correlation coefficients for a selection of 60 AERONET sites for displaying highly correlated total AOD for the AERONET data and GEFS-Aerosols, with the correlation coefficients 7 exceeding 0.7. In contrast, there is only 1 site with a correlation coefficient larger than 0.7 for NGACv2 model vs. AERONET,  Table 1, the ICAP results show the best performance in both the correlation and RMSE.

12
In addition to comparing with the AERONET data, Fig. 6 shows the Day 1 AOD prediction of GEFS-Aerosols and NGACv2

27
[2018] also showed that the predicted dust AOD in NGACv2 over northwestern Africa is much larger than GEFS-Aerosols,

29
OC is a major component emitted from wildfires, and OC AOD is a good indicator of the performance of fire impacts. GEFS-

30
Aerosols OC AOD shows smaller biases compared to the MERRA-2 reanalysis than those of NGACv2 (Figure 7c and d).

31
Positive biases in GEFS-Aerosols OC AOD of less than 0.2 occur mainly over southern Africa, eastern Asia, south Asia and

12
where the underestimates exceed 0.18, and in the eastern US and western Europe, where they exceed 0.1.

13
The summary comparison of the GEFS-Aerosols and NGACv2 Day 1 total AOD prediction biases with respect to MERRA-2 14 reanalysis between 7/5/19 and 11/30/19 is shown in Figure 8. Generally, the GEFS-Aerosols model is able to reproduce the 15 total AOD very well, much better than NGACv2 (see Figure 8a

27
We choose some sites near the major fire source region, which have available observation data for the duration of this study 28 and hold long records based on various previous studies. Figure 9 indicates the total AOD time series of AERONET 29 observations compared against ICAP, NGACv2 and GEFS-Aerosols model predictions at the four AERONET sites near the 30 fire source regions of South America during the period of 7/1/19-11/30/19. At the Alta Floresta site, which is in the middle of   Amazon fire source region, the daily AOD variations of both the ICAP and GEFS-Aerosols day 1 predictions are quite 1 consistent with that of the AERONET data, especially as they are able to reproduce two peaks in AOD enhancements in late 2 August and late September caused by fire plumes (Figure 9a). The correlation (RMSE) is 0.66 (0.23), 0.9 (0.12) and 0.68 (0.31) 3 for GEFS-Aerosols, ICAP and NGACv2. Obviously, NGACv2 results under predict AERONET observations almost 4 throughout the whole period with a significantly larger bias than GEFS-Aerosols or ICAP, and the two August-September 5 peaks in total AOD enhancements are essentially missed in the NGACv2 prediction.

6
The Itajuba site is located southeast of the Alta Floresta site and in the downwind areas of the Amazon fire source region. The

22
The last site of RioBranco is also located in the Amazon fire source region but to the west of the AltaFloresta site. There are 23 some missing data at this site for the AERONET total AOD from middle July to middle September ( Figure 9d). During this 24 period, the GEFS-Aerosols prediction is slightly lower than ICAP by about 5-10%. Both ICAP and GEFS-Aerosols total AOD 25 match the AERONET variations well when the AERONET data are available again from mid-September. Several peaks of 26 total AOD are also captured by GEFS-Aerosols in middle September and early November. The NGACv2 prediction shows 27 enhanced total AOD in middle August, with low biases by more than 2-3 times compared to ICAP and GEFS-Aerosols. For  We also evaluate the total AOD time series of AERONET against ICAP, NGACv2 and GEFS-Aerosols for fire regions of 1 central and southern Africa. The comparisons at seven AERONET sites from July to November are shown in Figure  10. Generally, the GEFS-Aerosols predictions are able to capture the daily total AOD variation measured by AERONET. At 3 the site of Misampfu, the GEFS-Aerosols mode is somewhat better than that of ICAP in predicted the peaks of high AOD. The 4 correlation coefficients at the sites of Maun Tower and Lubango are much high than those of ICAP (see Table 1). While both 5 the GEFS-Aerosols and ICAP overpredicted the total AOD most of time throughout the three months at the station Bamenda Ascension Island located west of the African fire source region, the GEFS-Aerosols and ICAP are able to capture the AOD enhancements in the middle of August, and shows the best performance of the three different models (see Table 1). For other  NGACv2 at 6 of these sites are shown in Figure 11. Overall, the GEFS-Aerosols model is able to closely predict the observed 24 total AOD variation, especially at the sites of Banizoumbu, Tenerife, Saada, Ben Salem, Granada and Sede Boker, with much 25 better performance than those of NGACv2 according to the correlation (RMSE) values in Table 1

28
it does not accurately capture observed temporal variations of total AOD at these sites.

29
We compare the daily AERONET total AOD with the 1-day forecasts of total AOD from GEFS-Aerosols and NGACv2 at the 30 AERONET sites of Cape Verde, Tamanrassett and Tenerife located in the dust source region over North Africa in Figure 12. The slope of the linear regression of AERONET total AOD vs. GEFS-Aerosols is quite different from that of NGACv2 at the 1 site of Tamanrassett, which is located in southern Algeria and in the middle of the Saharan dust source region. The GEFS-2 Aerosols linear regression slope is much closer to 1 than that of NGACv2, and the R 2 in the NGACv2 model is lower by a

23
The European region has the largest differences between GEFS-Aerosols and MERRA-2 reanalysis total AOD among the 9        10 tend to be more than a factor of 5 lower that the observations. 11 Table 3 gives median bias and correlation statistics for column sums of all GEFS-Aerosols model cases as well as the NGAC

20
Correlations of dust are also much lower than for other species, and there is a very obviously difference between GEFS-21 Aerosols and NGAC model forecast statistics as discussed further below. We note that sea-salt columns are not calculated or compared to ATom-1 observations, due to the large amount of observations below the detection limit, especially above 2 km 23 altitude.

26
The chemical component of atmospheric composition in GEFS-Aerosols is based on WRF-Chem, which is a community

19
biomass burning plumes in Southern Hemisphere, South America, Northwestern America and Eastern Europe, polluted air 20 over Eastern and Southern Asia, and high-altitude sea-salt bands. We also sampled the forecast total AOD of GEFS-Aerosols 21 and NGACv2 in the same location as 60 AERONET sites, which are spread globally and represent different aerosol regimes, and compared their variations for the 7/5/19-11/30/19. Much higher correlation coefficients against AERONET data are 23 indicated for the GEFS-Aerosols than those for NGACv2 globally, and are quite comparable to those of the ICAP-MME.

24
During the biomass burning events, GEFS-Aerosols captured major fires over southern Africa, Siberia, Central Amazon and 25 Central South America much better than NGACv2. Part of the improvement may be due to the vertical transport by the plume- areas, which may be due to the lack of removal process or uncertainties of fire emission in central and southern Africa. In 1 contrast, the NGACv2 results show under prediction in total AOD forecast at most of the AERONET sites in this region.
2 Overall, the model predicts total AOD variation by GEFS-Aerosols indicates much better performance than that of NGACv2 3 over western North Africa. Although GEFS-Aerosols shows reductions in dust emissions over the Saharan dust source, the 4 correlations with observations from downwind AERONET sites in western Africa are improved over those for NGACv2. The 5 largest biases and discrepancies of GEFS-Aerosols and NGACv2 are both indicated in the sites in Tajikistan, which may be 6 associated with a missing dust source near this site for both models.

7
We also evaluated predicted aerosols concentrations with different resolution against the ATom-1 aircraft measurements from OC, BC and sulfate, and the location of fire plumes was captured well overall. Sulfate over the Pacific, southern and tropical

11
Atlantic is significantly underpredicted, suggesting an underestimation in the oceanic sulfur sources, such as DMS. A clear 12 trend in increased overprediction with altitude for BC suggests that further refinements in characterizing precipitation 13 scavenging of aerosol in GEFS-Aerosols is needed, since this is the only loss process for BC other than surface deposition.     , 9, 1905-1919, https://doi.org/10.5194/gmd-9-1905-2016, 39 2016.