UCLALES is a large-eddy simulation which has mainly been used in idealised cloud and fog studies. It is based on the Smagorinsky–Lilly subgrid model and solves the Ogura–Phillips anelastic equations with an Asselin filter. Boundary conditions are doubly periodic in the horizontal and fixed in the vertical. Momentum variables are advected with leapfrog time stepping and scalar variables through forward time stepping. In the standard model, a two-moment warm rain microphysical scheme is used, the vertical is spanned by a stretchable grid, and a sponge layer is applied at the domain top to prevent gravity waves being released into the boundary . The surface scheme explicitly calculates SHF and latent heat flux (LHF) at each time step and is based on a coupled soil moisture and surface temperature scheme by (Eqs. 1, 2 and 3): 1SHF=ρCpTg-Tara2LHF=ρCpγfhesTg-earsurf+ra, where ρ is air density, Cp is specific heat capacity of dry air, Tg and Ta are surface and air temperature, respectively, γ is the psychrometric constant, fh is a dimensionless function related to water volume fraction and takes the value 0.267 in our case. es (Tg) is saturation vapour pressure at surface temperature (Tg) and es is water vapour at 2 m height. rsurf is the surface resistance to bare soil and is related to surface friction velocity (u∗). ra is atmospheric resistance to water vapour and heat and is dependent on atmospheric stability . 3ΔQs=ωChλ212Tg-T‾ Surface parameters, which vary greatly in different environments, can be varied within the model input and largely affect the heat storage term (ΔQs) (Eq. 3). Here, Ch is volumetric heat capacity (J m-3 K-1), λ is thermal conductivity (W m-1 K-1), ω is angular frequency (s-1) and T‾ (K) is the average daily temperature in the 2 cm soil layer. The resulting parameters as well as the overall radiation are utilised in the surface energy balance scheme detailed in (Eq. 4), where Q∗ is net all-wave radiation. 4Q∗=H+LHF+ΔQs

SALSA was developed by and has been coupled with large-eddy simulation models (UCLALES) as well as a climate model (ECHAM) . SALSA bins aerosols according to size, allowing for a variety of aerosol sizes and compositions as well as for aerosols to be either internally or externally mixed (Fig. ; ). When SALSA is used in these simulations, aerosol species included are black carbon, sulfate, organic carbon, nitrate and ammonium, with all aerosols assumed to be internally mixed. In terms of aerosol processes, coagulation and water vapour condensation are switched on, while nucleation, aerosol deposition, emissions and semi-volatile condensation are not considered here for simplicity but may be considered in future work.

UCLALES-SALSA couples the UCLALES with SALSA and is primarily described in the paper by . The version of UCLALES-SALSA here is a fully coupled radiation–dynamical model, whereby the aerosol–radiation interactions in SALSA are fully coupled with the four-stream radiative solver in UCLALES which feeds back on boundary layer turbulence. This is the first time that aerosol–radiation interactions have been dynamically coupled to UCLALES and the work outlined here examines the sensitivity of aerosol loading to these interactions and feedback.

All measurements used in this study were taken at IAP, Chinese Academy of Sciences, as part of the APHH Beijing campaign. Measurements taken include but are not limited to NR-PM1 (non-refractory PM with a diameter <1 µm) composition and aerosol and black carbon size and concentration measurements at the surface, as well as meteorological measurements at 15 levels on a 320 m tower. Sensible and latent heat flux measurements were calculated and a ceilometer was used to infer PBL height. For more information concerning the measurements taken as well as the APHH project and field campaign, the reader is directed to the Introduction to the special issue “In-depth study of air pollution sources and processes within Beijing and its surrounding region (APHH-Beijing)” by .

The domain size for all model simulations spanned 5.4 km in the horizontal with a resolution of 30 m, and the model top was set to 1.8 km in the vertical with a resolution of 10 m. The model uses an adaptive time step with a maximum time step of 1 s. A haze period which took place within the APHH winter campaign period from 24 to 26 November 2016 was used to examine the sensitivity of boundary layer meteorology to varying aerosol concentrations. Meteorological data taken from ECMWF-ERA5 reanalysis and tower meteorological data were used to initialise vertical profiles at 08:00 LST on 24, 25 and 26 November. Simulations were run from 08:00 LST for 14 h (22:00 LST) including 1 h spin-up time. Simulations for all days were considered to be cloudless. Case studies for each day were simulated and compared to each other and are described as follows: case 1 – no aerosols, case 2 – high and low aerosol loading, case 3 – aerosol vertical profiles. For case 2, aerosol vertical profiles were constant in the column, whereas case 3 examined the impact of including a varying aerosol vertical profile. Aerosol size distribution parameters and volume fraction of aerosol components were the same for all simulations, detailed in Tables 1 and 2. The values for aerosol size distribution data used were measured by a scanning mobility particle sizer (SMPS) and aerosol composition measurements were taken with an aerosol mass spectrometer (AMS). All in situ measurements were taken at IAP and values used were averaged between 07:30 and 08:30 LST on 24 November 2016. In all cases, BC can be considered to be the primary absorbing aerosol, with sulfate (SO4-), nitrate (NO3-) and ammonium (NH4+) strongly scattering and organics (OC) predominantly scattering with a small absorbing component. Aerosol growth is considered through the processes of coagulation and water condensation, but semi-volatile condensation is not considered. Both wet and dry deposition are switched off in all simulations.

Simulations in case 1 examine the development of boundary layer dynamics for 24, 25 and 26 November without aerosol–radiation interactions. All 3 d are initialised with different meteorological vertical profiles, taken from ECMWF profiles. On 25 November, there is a strong temperature inversion throughout the whole profile, while on 26 November there is strong vertical wind shear, higher surface humidity and strong stability in the lowest 300 m (Fig. ). Strong vertical wind shear causes mechanical turbulence, while a strong temperature inversion in the morning can suppress boundary layer development through reducing buoyancy. Figure d–f show development of SHF, PBL height and total turbulent kinetic energy (TKE) for the three simulated days with different meteorological conditions initialised in the morning.

SHF is similar in magnitude for all 3 d, while TKE and simulated PBL height are significantly lower for the 25 November simulation. A well-mixed, turbulent boundary layer forms quickly on 24 November; however, on 25 November, a shallow, weakly turbulent boundary layer remains throughout the day, and on 26 November a turbulent boundary layer is much slower to develop (Fig. ). The changing conditions used here are typical for a Beijing haze episode and show that even without the consideration of aerosols, meteorological conditions can largely affect the diurnal development of boundary layer dynamics.

Case 1 shows that simulated boundary layer dynamics are impacted by initial meteorological conditions. In case 2, the sensitivity of boundary layer dynamics to aerosol loading is examined, where aerosol mixing ratios were constant throughout the profile as shown in Fig. . Table 4 shows the impact of including high and low aerosol loading on maximum SHF and maximum PBL height between 12:00 and 16:00 LST.

In all cases, inclusion of aerosols causes cooling in the lower planetary boundary layer and warming above it. This is due to the aerosols absorbing and scattering incoming SW radiation (Fig. ) to reduce the amount of solar radiation reaching the surface. Where there are high concentrations of aerosols through the column, this severely reduces the amount of radiation reaching the surface and consequently causes cooling. Aerosols, specifically black carbon, in the upper layer of the boundary layer will absorb radiation, which causes warming. Including high aerosol concentrations (220 µg m-3) compared to low aerosol concentrations (55 µg m-3) maximised this effect, leading to enhanced temperature inversions, and suppressed PBL development on all 3 d (Table ).

For the case of 25 November, the PBL is already low due to synoptic conditions and aerosols from the previous day causing strong temperature inversions in the morning. Therefore, even though the aerosols cause cooling in the PBL to the same extent on 26 and 25 November, a strong temperature inversion already exists on 25 November and so the PBL is low even without the inclusion of aerosols.

To assess the sensitivity of the model to a varied aerosol vertical profile, case 3 uses the same setup as case 2 but varies the aerosol mass mixing ratio with altitude, as shown in Fig. . This is to assess the impact of high aerosol concentrations aloft in case 2 simulations which may magnify the aerosol–radiation effect, due to higher total loading increasing the total column aerosol optical depth (AOD). In case 3 simulations, total aerosol mass loading throughout the column is ∼22 % less than for case 2 simulations for both high and low aerosol simulations. The aerosol profile was chosen so that at the first time step, the aerosol mass mixing ratio at the surface was the same as those with a constant profile and decreased above the PBL in accordance with the potential temperature profiles, while composition and size remained constant throughout. It should be noted from the varied aerosol vertical profile simulations that total aerosol mass mixing ratio decreases by about 5 % over the course of the day. This is despite dry deposition not being included in these simulations. This is a result of UCLALES-SALSA using the Ogura–Phillips anelastic approximation for filtering out acoustic waves. The approximation assumes that there are only small variations in pressure and density from static reference values over time. Throughout the day, surface fluxes increase air temperature, while subsidence of air at the model top decreases density . The limitations of the anelastic approximation mean that these changes do not fully feed back to change pressure, and fixed boundary conditions mean that volume remains constant. As the model holds to constant volume rather then constant mass, when SALSA aerosol mass tracers are pulled downward, the total air mass increases while the mass of aerosols remains the same; this causes the apparent decrease in aerosol mass mixing ratio (Fig. ). We consider this to be a limitation of using a meteorological model for air quality analysis; however, as the relative reduction is the same for different meteorological conditions, comparisons between different cases can still be performed.

Figure  shows simulation results of potential temperature and aerosol number mixing ratio at 17:00 LST (9 h of simulation) for constant and varied aerosol vertical profiles at high concentrations for 24 and 26 November. When a varied aerosol profile is included, vertical mixing of aerosol occurs, resulting in a difference in the aerosol vertical profile on each day at 17:00 LST due to the difference in meteorology. The difference between the aerosol profiles over time shows the modelled meteorological feedback on aerosol mixing ratios.

Figure  compares the variance in vertical velocity (σw2) for low and high aerosol loading throughout the profile and at high aerosol loading at the surface only for the 24 and 26 November simulations and shows that high aerosol loading, both throughout the column and at the surface, decreases σw2 at 500–750 m in the afternoon of both 24 November and 26 November. It also shows the effect of high aerosol loading throughout the column, which causes an increase in σw2 close to the model top, with the varied aerosol vertical profiles minimising this effect as the aerosols increase vertical velocity, through creating a turbulent layer. This is due to aerosol warming aloft close to the model top causing stratification of the layer. Reduced aerosol concentrations in the entire column indicate that more solar radiation reaches the surface in the varied vertical aerosol profile case, increasing buoyant turbulence and vertical velocity at lower altitudes.

Case 1 identifies the importance of meteorological conditions on boundary layer dynamics throughout the day. Many observations in Beijing found that meteorological conditions are a main driver for both the onset and longevity of haze. Large-scale synoptic conditions such as southerly winds and low pressure often pre-empt pollution episodes which tend to occur every 4–7 d during wintertime in Beijing . These conditions are associated with the beginning of “haze” as the switch in meteorological conditions from strong northwesterly to southerly winds advects pollution from surrounding provinces into Beijing. This change is also associated with a low pressure field within the city, where stagnant air becomes trapped and the dispersion of pollutants is inhibited .

The initial meteorological profiles for the simulations on 24 November are taken prior to the onset of the haze and are associated with clean conditions. This is likely the reason for the quick turbulent boundary layer development along with high TKE and SHF throughout the day. Observations show that aerosol concentrations begin to build up around midday on 24 November and remain constant until the afternoon of 25 November when concentrations build up rapidly, peaking overnight on 25 November and remaining high until the afternoon of 26 November. Therefore, the initial conditions used in the simulation of 25 November will have been slightly affected by aerosol–radiation interactions of the previous evening. Aerosol–radiation interactions reduce the amount of solar radiation reaching the surface which causes cooling; simultaneously, black carbon aerosols will absorb radiation at the top of the PBL. Although absorption by BC occurs throughout the column, several studies have shown that due to the higher incidence of solar radiation and lower density of air, BC causes warming at PBL top to a greater extent than at the surface . Overall, this causes a temperature inversion during periods where pollution is high and causes a shallow PBL to form during the day. This leads to stagnant conditions and can affect the meteorology of the next day, particularly when aerosols are suppressed in a shallow PBL. However, frequently during wintertime in Beijing, changes in pressure can cause warm polluted air to converge with cold clean air to create a layer of cold air under a layer of warm air. These conditions often pre-empt pollution episodes in Beijing and favour the accumulation of pollutants in a shallow boundary layer. A combination of these factors explains the strong temperature inversion in the morning and results in a shallow turbulent boundary layer forming in these simulations, with lower turbulent kinetic energy compared to the 24 November simulation (Fig. ).

Aerosol–radiation interactions cause a reduction in SHF, surface SW radiation and TKE, resulting in a reduction in the daily maximum PBL height for all 3 d examined. In these simulations, the aerosols interact with radiation to cause heating and cooling in different layers which perturbs the temperature profile of the PBL and decreases the sensible heat flux term. The aerosols also take up water to a small extent which decreases latent heat. These effects lead to decreased turbulence in the PBL. High aerosol concentrations enhance this effect due to an increased number of aerosols interacting with radiation. This leads to a reduction in maximum SHF of 44, 33 and 45 W m-2 for high compared to low aerosol loading simulations on 24, 25 and 26 November, respectively. However, results from case 2 show a variation in the magnitude of the aerosol–radiation effect with a larger impact on maximum PBL height for high aerosol simulations on 26 November compared to 24 and 25 November (Table 3). Including high aerosols on 26 November causes >1 ∘C of daytime cooling in the lowest 300 m compared to 0.3 ∘C of cooling on 24 November (Figs.  and ). The larger degree of cooling on 26 November leads to a larger reduction in buoyant turbulence and prevents the full growth of a deeply turbulent boundary layer to a larger extent on 26 November.

High aerosol concentrations are known to stabilise the boundary layer through the reduction of vertical transport of momentum to the surface . This can reduce wind speeds at lower altitudes and thus decrease wind shear and the shear component of TKE. In case 2, high aerosol loadings reduce surface wind speeds, wind shear and surface frictional velocity (u∗) for all 3 d, with a greater reduction on 26 November compared to 24 and 25 November. High aerosol loading also causes a reduction in the variance of vertical velocity (σw2), which can be considered a measure of turbulence . On both 24 and 26 November, simulations with high aerosol loading caused a reduction in the magnitude of σw2 particularly between 500 and 800 m. On 24 November, the decrease at the surface at σw2 is ∼40 %, while on 26 November the reduction is 75 %. In the case of 26 November, this is accompanied by increased values of σw2 in the upper layers close to the model top, which results in two turbulent layers forming separated by a stable layer (Fig. ).

In these simulations, aerosol profiles are constant through the column and high aerosol concentrations aloft. Figure  shows that high aerosol throughout the column causes warming in the upper layers and cooling in the lower layers, which causes strong stability throughout the profile. In reality, aerosols tend to be concentrated closer to the surface and within the boundary layer, although occasionally in Beijing regional transport can lead to higher aerosol concentrations aloft. Therefore, case 3 investigated the effect of limiting pollution to the surface by including aerosol vertical varying profiles.

Case 3 examined the impact of meteorological feedback on aerosol vertical mixing for high and low aerosol loading simulations by including aerosol vertical profiles on 24 and 26 November. Simulations with a varied vertical aerosol profile had the same aerosol concentrations at the surface as the high aerosol simulations in case 2 but reduced concentrations at higher altitudes (Fig. ). This resulted in a small increase in maximum SHF (∼7 W m-2) on both 24 and 26 November. For 26 November, limiting high aerosol loading to the surface results in an afternoon increase in turbulence up to 500 m. Furthermore, the effect of high aerosols throughout the column (case 2) resulted in a highly turbulent layer at model top and a large reduction in surface wind speed on 26 November (Fig. ). As this turbulent layer is significantly reduced with lower aerosols aloft, this effect may be considered to be an artefact of aerosol loading at high altitudes which is not often observed in poor-air-quality events during wintertime in China. However, overall, the contribution of the shear term to turbulence is minimal compared to the buoyancy term, which is greatly reduced by high aerosol loading in both case 2 and case 3. The high aerosol loading in case 2 has a larger effect on boundary layer development than the effect of varying the aerosol vertical profile in case 3. Therefore, we can consider the change in the thermal profile of the atmosphere, due to high concentrations of aerosols increasing aerosol–radiation interactions, to be the prominent cause of the reduction in SHF and PBL height (Table 3).

The large degree of cooling on 26 November compared to 24 November is due to the effects of initial meteorology feeding back on aerosol–radiation interactions. Figure  shows potential temperature and aerosol number mixing ratio vertical profiles for each case (after 9 h of simulation) under high aerosol loading at the surface only (red and green lines) and throughout the profile (blue and turquoise lines). After 9 h of simulation (17:00 LST), surface aerosol concentrations on 26 November are higher than on 24 November. This is due to aerosol–radiation interactions and initial meteorological conditions on 26 November resulting in a shallower PBL (Table 3). This shows the ability of UCLALES-SALSA to simulate the aerosol–radiation–meteorology feedback loop and that the feedback effect can have a significant impact on aerosol surface concentrations, which will consequently feed back further on atmospheric stability.