The Role of Aerosol‐Radiation Interaction in the Meteorology Prediction at the Weather Scale in the Numerical Weather Prediction Model

The current capabilities of the numerical weather prediction (NWP) model to simulate aerosol‐radiation interaction (ARI) impacts on weather prediction are still rarely considered compared with climate models. Here, the NWP model GRAPES_CUACE is used to evaluate the role of the ARI in meteorology prediction at the weather scale in China. The results show that the online calculation of ARI in the model can extensively improve the meteorology prediction accuracy involving temperature, wind, and pressure at most vertical levels in relatively clean, light, and heavy pollution stages. This improvement significantly occurs in the meteorology factors below 950 hPa prediction such as temperature at 2 m, 1,000, and 950 hPa, and mean sea level pressure, particularly in heavy pollution areas and stages. Unlike temperature, the improvement of ARI in the predicted wind at the height of the boundary layer is more significant than near‐surface. However, this improvement declines when the low‐cloud exists.

an unprecedented increase of aerosol optical depth (AOD) and PM 2.5 mass concentration have occurred in Jing-Jin-Ji (Cai et al., 2017;Che et al., 2015;Y. Sun et al., 2014;Z.-b. Sun et al., 2016). Such a regional study of the ARI effect in Jing-Jin-Ji is nothing more than representative. Some studies have been carried out to investigate the impact of ARI on air quality and weather (Huang et al., 2018;H. Wang et al., 2018;Z. Wang et al., 2020;Zhong et al., 2017). Scattering properties of composite aerosol is dominant during HPEs in Beijing, which caused the significant decrease of the surface downward shortwave radiation (SDSR) and surface temperature . A two-way feedback mechanism between aerosol and meteorology due to the ARI is found through the analysis of HPEs in Beijing Zhong et al., 2017). By using the NWP model (e.g., the Global-Regional Assimilation and Prediction System coupled with the Chinese Unified Atmospheric Chemistry Environment [GRAPES_CUACE]), some studies find that the ARI can reduce PM 2.5 negative errors in the model (H. Wang et al., 2018;H. Wang, Shi, et al., 2015). Moreover, the ARI effect over northern China helped to reduce the bias of simulated temperature at 2 m and horizontal wind speed from Weather Research and Forecasting model by ∼73.9% and ∼7.8%, respectively (Yang et al., 2020). All these indicate that the impact of the ARI on the weather is evident in Jing-Jin-Ji. Still, most of the research focuses on the air quality forecast and the prediction of a single meteorological factor rather than predicting from a weather scale perspective. Therefore, it is necessary to study the impact of the ARI on meteorology prediction at the weather scale in the current NWP model. This article uses the GRAPES_CUACE model to study the impact of the ARI on 24 hr weather prediction in Jing-Jin-Ji. The result will help to understand how much the ARI can improve the meteorology prediction of the current NWP model at the weather scale associated with aerosol pollution in China and provide a meaningful reference for a more accurate meteorology forecast in the future.

Data
The hourly PM 2.5 mass concentration data are provided by the Ministry of Ecological Environment of the People's Republic of China. So far, more than 1,400 PM 2.5 monitoring sites have been established in various cities in China (the red dots in Figure 1a). The hourly surface radiant exposure datasets are collected by three radiation stations of China Meteorological Administration (CMA; the black rectangles in Figure 1b). The hourly ground meteorological observation data are provided by more than 250 automatic meteorological observatories of CMA in the Jing-Jin-Ji (the purple rectangles in Figure 1b), including wind, temperature, relative humidity, pressure, and low-cloud amount. The meteorological vertical observation data twice a day (00:00 UTC and 12:00 UTC) are obtained from four sounding stations of CMA (the yellow signs in Figure 1b). The date range of meteorological data and PM 2.5 data used in this study is from 1 December 2016 to 9 January 2017.
Daily cloud optical thickness (COT) data (from 11 to 14 December 2016) and AOD data (from 1 December 2016 to 9 January 2017) come from the Suomi National Polar-orbiting Partnership Visible Infrared Imaging Radiometer Suite (VIIRS) NASA Level-3 (L3) product. The horizontal resolution of COT and AOD data is 1° × 1° on a lat-lon global map. The National Centers for Environmental Prediction (NCEP) global final (FNL) reanalysis data are prepared operationally every 6 hr of a day with a horizontal resolution of 0.25° × 0.25°. These data from 1 December 2016 to 9 January 2017 are used as initial field and boundary conditions of the model.
The model domain includes the study area (Jing-Jin-Ji; Figure 1a), which has a horizontal resolution of 0.15° × 0.15° and 31 vertical layers. The emission data entered in the model are the Multi-resolution Emission Inventory for China of Tsinghua University in 2016 (M. Li et al., Zhang, 2014;F. Liu et al., 2016). Selected parameterization schemes and detailed ARI parameterization in the model are presented in the Supporting Information (Table S1 and Text S1 in Supporting Information S1). The model simulation period is from 1 December 2016 to 9 January 2017 and the forecast time is 24 hr. There are 72 hr spin-up to ensure the model can run stably and reduce model errors from the initial and boundary conditions.
In this study, we conduct two experiments (Table S2 in Supporting Information S1) to determine the impact of the ARI on simulations in the model by controlling whether to activate the ARI.

Model Validation
To analyze the role of the ARI in weather scale prediction, it is necessary to evaluate the simulated aerosol changes to ensure that the GRAPES_CUACE model can reproduce the aerosol characteristics in Jing-Jin-Ji. Previous studies have shown that the extinction of aerosol in winter in China is mainly due to the scattering and absorption of short-wave solar radiation by PM 2.5 (Cheng et al., 2017) and there is a high degree of consistency between PM 2.5 and AOD (Chudnovsky et al., 2014;Guo et al., 2017;Song et al., 2018). Figure S1 in Supporting Information S1 shows the comparison of time series of PM 2.5 mass between observations and simulations (NO-ARI experiment) from 1 December 2016 to 9 January 2017. This model can simulate the time changes of PM 2.5 mass including the accumulation and dissipation of PM 2.5 mass within a few hours and the calculated correlation coefficients (R) are greater than 0.5 (Figures S1a-S1f in Supporting Information S1). Besides, the error of simulated AOD is acceptable ( Figure S2 in Supporting Information S1). Both of them indicate that the results predicted by this model can reproduce the aerosol characteristics in Jing-Jin-Ji, which can be used to analyze the impact of the ARI on the meteorology prediction at the weather scale.

The Impact of the ARI on SDSR
The fundamental reason for the impact of the ARI on weather is solar short-wave radiation (Rosário et al., 2013;Ruiz-Arias et al., 2016). Figure 2 shows the observed and simulated SDSR from 1 December 2016 to 9 January 2017. The simulated SDSR from the NO-ARI experiment differs greatly from the observation during the whole period. The mean bias (i.e., bias = Xsim − Xobs where Xsim and Xobs represent the simulations and observations) is 39.4 MJ/m 2 in Jing-Jin-Ji and 43.1 MJ/m 2 in Beijing. When the ARI is considered in the model, the simulated SDSR is in agreement with the observation (Figures 2a and 2b). At the same time, compared with simulations from the NO-ARI experiment, the mean bias can be reduced by 80% in Jing-Jin-Ji (73% in Beijing). However, when the low-cloud period (i.e., the period when the daily low-cloud amount is greater than 20%) occurs, the simulated SDSR from the ARI experiment exceeds the observed SDSR by as much as a factor of one. The ARI does not play its due role in improvement (Figures 2a and 2b). We give related reasons in Section 3.4 for the lower improvement of the ARI in model prediction during the low-cloud period.

The Impact of the ARI on Predicted Meteorological Factors
The improvement of simulated SDSR caused by the ARI will inevitably lead to changes in the forecast of meteorological factors at the weather scale in the model. Figure 3 shows the hourly mean absolute error skill score (MAESS) of predicted meteorological factors (temperature at 2 m, u and v components of wind at 10 m, mean sea level pressure, and temperature and u and v components of wind at 1,000, 950, 900, 850, 800, 700, 500 hPa) from the ARI experiment based on simulations from NO-ARI experiment. The ARI positively affects the improvement of most predicted meteorological factors in Jing-Jin-Ji, Beijing, and Xingtai. The MAESS of predicted meteorological factors from the ARI experiment below 800 hPa are more remarkable, especially temperature at 2 m, 1,000, and 950 hPa, and mean sea level pressure, which indicates that the ARI has a more prominent impact on the lower atmosphere. To compare the impact of the ARI on the forecast in different aerosol pollution stages, clean stages (i.e., PM 2.5 mass ≤ 75 μg m −3 ), light pollution stages (i.e., 75 μg m −3 < PM 2.5 mass ≤ 150 μg m −3 ), and heavy pollution stages (i.e., PM 2.5 mass > 150 μg m −3 ) are distinguished according to China's national environmental quality standards. We find that the MAESS of predicted meteorological factors in the heavy pollution stages is the largest, followed by the light pollution stages and the clean stages. This implies that the ARI can improve the forecast more significantly in the heavy aerosol pollution period. It should be noted that the ARI effectively improves the accuracy of simulation in relatively clean stages, especially in Jing-Jin-Ji (Figure 3a). Even in the future, when aerosol pollution continues to decline, the role of the ARI cannot be ignored in the NWP model. Besides, the improvement of ARI in the predicted wind in the vertical direction is different from the temperature. The hourly MAESS of u and v components of wind at the height of the boundary layer (e.g., 950, 900, and 850 hPa) is greater than that at the height of the near-surface layer (e.g., 10 m and 1,000 hPa), which reflects a more obvious improvement of ARI. The improvement in the upper-level wind is probably related to Ekman pumping and enhanced secondary circulations (J. Li et al., 2021).

The Impact of the ARI on the Meteorology Prediction in Different Regions
Four meteorological factors (temperature at 2 m, 1,000, and 950 hPa, and mean sea level pressure) most obviously improved by the ARI in Beijing and Xingtai are selected to explain the impact of the ARI on meteorology prediction in different regions. During the whole period, the hourly MAESS of four predicted meteorological factors from the ARI experiment in Xingtai is greater than that in Beijing (Figures 3b and 3c). Furthermore, we choose the rising process of PM 2.5 mass from 15 to 20 December 2016 with similar PM 2.5 mass in early stages ( Figure S3 in Supporting Information S1) and find that the PM 2.5 mass is more easily to accumulate in Xingtai caused by the geographical location and topography (Figure 1b). At the same time, compared with the NO-ARI experiment, the simulations from the ARI experiment reduce the mean bias of four meteorological factors by 26%, 15%, 21%, and 19% in Beijing and 35%, 20%, 30%, and 21% in Xingtai, respectively (Figure 4). ARI plays a more pronounced role in the meteorology prediction of the current model in Xingtai. Of course, the bias of the model system and local emissions cannot be ignored (F. Liu et al., 2016;H. Zhang et al., 2016) and further studies are needed to determine the details.

The Performance of Improvement of ARI During the Low-Cloud Period
From Section 3.1, we have found that although SDSR from the ARI experiment has been closer to observed SDSR than that from the NO-ARI experiment, the improvement degree of the ARI on simulated SDSR decreases during the low-cloud period. It can be seen from Figure 5 that the ARI reduces the mean bias of simulated SDSR by 56% in Jing-Jin-Ji and 47% in Beijing during the low-cloud period, both of which are lower than reductions of the mean bias during the non-low-cloud period. Besides, compared with the simulations during the whole period, the hourly MAESS of each predicted meteorological factor during the low-cloud period has been significantly reduced (Figure 5c). It indicates that the ARI still plays a crucial role in improving the current model prediction at the weather scale during the low-cloud period, but the improvement degree is reduced.
According to previous studies, the more complex aerosol and weather processes during the low-cloud period significantly increase the uncertainty of the model prediction and limit the impact of the ARI (Brient & Bony, 2013;W. Li et al., 2011). Figure S4 in Supporting Information S1 shows the daily COT of VIIRS and simulations from 11 to 14 December 2016, which can help to understand the possible reasons for the declining improvement of the ARI during the low-cloud period. The model simulates the spatial distribution and temporal changes of COT. However, the mean bias of simulated COT during the low-cloud period in Jing-Jin-Ji is much greater than that during the period without low-cloud and the ARI does not improve. This situation will undoubtedly be an important factor affecting the accuracy of current model prediction (Mulcahy et al., 2014). More detailed reasons need to be found. At the same time, combined with the decreased PM 2.5 mass concentration and the emergence of low-cloud ( Figure S5 in Supporting Information S1), we speculate that mean bias of simulated COT during the low-cloud period may also be affected by aerosol-cloud interaction (ACI; L. Liu et al., 2019;Poore et al., 1995), but we are reluctant to present it, as we do not have sufficient evidence for the mechanism of low-cloud formation during this period. In the future, we will add the ACI into the model to find out the reasons for the mean bias of COT during the low-cloud period.

Conclusions
ARI is widely concerned and recognized in the climate model for the past many years, but there are few studies on understanding the importance of the ARI for the NWP model. In this study, we conduct NO-ARI and ARI experiments by using the GRAPES_CUACE model to analyze the role of the ARI in the meteorology prediction of the current NWP model in China. The simulated PM 2.5 mass and AOD from the NO-ARI experiment are consistent with the observations and can reproduce the aerosol pollution level in Jing-Jin-Ji. The results of the sensitivity experiment show that the ARI reduces the mean bias of SDSR during the whole study period by 80% and 73% in Jing-Jin-Ji and Beijing, respectively. Moreover, the ARI can more effectively improve the 24 hr forecast results of the vertical structure of temperature and wind in heavy pollution stages compared with that in light pollution stages and clean stages. Predicted meteorological factors below 800 hPa are more obviously improved by the ARI, especially temperature at 2 m, 1,000, and 950 hPa, and mean sea level pressure. It is worth noting that the improvement of ARI in u and v components of wind at the height of the boundary layer (e.g., 950, 900, and 850 hPa) is more significant than that at the near-surface (e.g., 10 m and 1,000 hPa). At the same time, the improvement of the ARI on meteorology prediction in Xingtai is more significant than that in Beijing, which is affected by topography and geographical location. In addition to the Jing-Jin-Ji of China, the role of the ARI in the weather forecast in other areas with high aerosol loading over land (e.g., northern India and southern Africa) cannot be ignored (Huang & Ding, 2021). However, this improvement has declined when the low-cloud exists caused by more significant uncertainty of simulated cloud characteristics from the model. Further studies should consider more works about the simulation of cloud characteristics, the impact of the ACI on the ARI effect, etc.