How does the Environment Modulate Aerosol Impacts on Tropical Sea Breeze Convective Systems?

. This study investigates how the enhanced loading of microphysically and radiatively active aerosol particles impacts tropical sea breeze convection and whether these aerosol impacts are modulated by the multitudinous environments that support these cloud systems. To achieve these goals, we have performed two large numerical model ensembles, each comprised 10 of 130 idealised simulations that represent different initial conditions typical of tropical sea breeze environments. The two ensembles are identical with the exception of the fact that one ensemble is initialised with relatively low aerosol loading or pristine conditions, while the other is initialised with higher aerosol loading or polluted conditions. Six atmospheric and four surface parameters are simultaneously perturbed for the 130 initial conditions. Analysis of the ten-dimensional parameter simulations was facilitated by the use of a statistical emulator and multivariate sensitivity techniques. 15 Comparisons of the clean and polluted ensembles demonstrate that aerosol direct effects reduce the incoming shortwave radiation reaching the surface, as well as the outgoing longwave radiation, within the polluted ensemble. This results in weaker surface fluxes, a reduced ocean-land thermal gradient, and a weaker sea breeze circulation. Consequently, irrespective of the different initial environmental conditions, increasing aerosol concentration decreases the three ingredients necessary for moist convection: moisture, instability, and lift. As reduced surface fluxes and instability inhibit the convective 20 boundary layer development, updraft velocities of the daytime cumulus convection developing ahead of the sea breeze front are robustly reduced in the polluted environments. Furthermore, the variance-based sensitivity analysis reveals that the soil saturation fraction is the most important environmental factor contributing to the updraft velocity variance of this daytime cumulus mode, but that it becomes a less important contributor with enhanced aerosol loading. It is also demonstrated that enhanced aerosol loading results in a weakening of the convection initiated along the sea 25 front. This suppression is particularly robust when the sea breeze-initiated convection is shallower, and hence restricted to warm rain processes. However, when the sea breeze-initiated convection is deep and includes mixed-phase processes, both the sign and magnitude of the convective updraft responses to increased aerosol loading are modulated by the environment. The less favourable convective environment arising from aerosol direct effects also restricts the development of sea breeze-initiated deep convection. While precipitation is ubiquitously suppressed with enhanced aerosol loading, the magnitude of this 30 suppression remains a function of the initial environment. Altogether, our results highlight the importance of evaluating aerosol impacts on convection systems under the wide range of environments supporting such convective development. comparison between rOn-500 and rOn-2000 demonstrated that aerosol direct effects resulted in less shortwave radiation reaching the surface. The associated longwave emission from the land surface was also reduced, thereby producing 515 a cooler land surface, a smaller land-sea thermal contrast and a weaker sea breeze circulation. The land-surface sensible and latent heat fluxes, the surface-based mixed layer depth, and the sea breeze inland extent were all subsequently reduced in the polluted scenario as a result. The three ingredients of moist convection, i.e., moisture, instability, and lift (Doswell, 1987), were all therefore reduced in the presence of aerosol direct effects. The reduction in these three convective ingredients as a result of enhanced aerosol loading was found in all of the 130 ensemble simulation pairs, and therefore suggest a robust result 520 that is independent of environmental conditions. The magnitude of the reduction does vary as a function of the initial environmental. These results extend those of Grant and van den Heever (2014), in which a similar sensitivity of tropical sea breeze convection to aerosol loading was demonstrated, albeit for only one set of initial conditions.

interactions may have important feedbacks to the environment and the resulting sea breeze convective system. Nevertheless, studies of aerosol impacts on cloud properties, in particular deep convection, have quite often neglected aerosol-radiation interactions in the interest of focusing specifically on aerosol indirect effects (e.g., Storer and van den Heever, 2013;Miltenberger et al., 2018;Marinescu et al., 2021).
Depending on their size and composition, various aerosol particles can serve as cloud condensation nuclei (CCN). 70 Under the assumption of the same liquid water content, when the number concentration of CCN is increased, a greater number of smaller cloud droplets are formed, thereby increasing the cloud albedo (Twomey, 1974) and the cloud lifetime (Albrecht, 1989), in what are often referred to as the first and second aerosol indirect effects, respectively. Less efficient collision and coalescence processes between the population of more numerous smaller cloud droplets suppresses the warm rain process, resulting in the longer cloud lifetimes described in the second indirect effect. These aerosol effects were initially postulated 75 for shallow convective clouds, but have also been observed to occur in deep convective cloud systems (Tao et al., 2012). It has been further hypothesised that with the suppression of the warm rain process, more numerous cloud droplets are lifted above the freezing level and frozen, thereby releasing additional latent heating and potentially strengthening updrafts. This process, which takes places in the mixed-and cold-phase regions of clouds, is often referred to as "cold-phase invigoration" and has been reported in both observational and modelling studies (e.g., Andreae et al., 2004;Khain et al., 2005;Koren et al., 80 2005;van den Heever et al., 2006;Rosenfeld et al., 2008). Under increased aerosol loading, the condensational growth rate of the population of more numerous smaller cloud droplets within the warm phase regions of clouds has also been observed to increase due to the increased exposed total droplet surface area, thereby releasing more latent heating and enhancing convective updrafts closer to cloud base. These warm-phase aerosol-cloud dynamical feedbacks have collectively been termed "condensational invigoration" (Kogan and Martin, 1994;Seiki and Nakajima, 2014;Saleeby et al., 2015;Sheffield et al., 2015) 85 or "warm-phase invigoration" (Fan et al., 2018). It has recently been suggested that warm-phase invigoration may be more significant and robust than its cold-phase counterpart (Marinescu et al. 2021;Igel and van den Heever 2021).
While convective invigoration theories have proposed stronger updrafts and/or heavier precipitation for convective clouds developing within enhanced aerosol loading conditions, a number of studies have questioned the robustness of these theories for various reasons, one of which is the modulation of these effects by the cloud environment. For example, wind 90 shear (Fan et al., 2009;Lebo and Morrison, 2014;Marinescu et al., 2017), CAPE (Lee et al., 2008;Storer et al., 2010;Storer et al. 2014), boundary layer instability (Marinescu et al., 2021) and moisture (Khain et al., 2005(Khain et al., , 2008Tao et al., 2007;Grant and van den Heever, 2015), have all been found to be modulating factors of aerosol impacts on cloud systems. Such environmental modulations have not been explored in the context of sea breeze convective systems, in spite of the fact that they are one of the most ubiquitous forms of convection in the tropics (Hadi et al., 2002;Perez and Silva Dias., 2017), and as 95 such, forms the focus of this paper. The primary goals of this study are twofold. The first goal is to investigate how radiatively and microphysically active aerosols change the convection that develops within tropical sea breeze regimes, specifically over the land. The second goal is to determine whether and how these convective responses to enhanced aerosol loading are modulated by the large number of environmental factors supporting sea breeze convective systems.
To achieve our stated goals, we have examined aerosol-cloud interactions within a fully interactive aerosol-100 radiation-cloud-land surface framework and under a wide range of different tropical sea breeze environments. This is accomplished through the use of a large ensemble of idealised numerical simulations in which we investigate the impact of varying aerosol number concentrations on the characteristics of sea breeze convective systems developing under a wide range of different initial environmental conditions previously identified in the literature as being important to such systems. More specifically, we extend Park et al. (2020) in which we assessed the relative importance of ten different thermodynamic, wind, 105 and surface properties under a relatively pristine aerosol scenario, by running the same large model ensemble under more polluted conditions. We use the term "polluted" here simply to mean increased aerosol loading and are not necessarily referring specifically to anthropogenic aerosols. Comparing these two large model ensembles allows us to assess the role of radiatively and microphysically active aerosols on sea breeze convection and how these aerosol effects may change as a function of the wide range of tropical sea breeze environments represented here. 110 Detailed experiment design, model configuration, and analysis methodologies are demonstrated in Section 2. Section 3 describes the development of sea breeze convective systems, including convection developing ahead of and along the sea breeze front. The overall differences between pristine and polluted ensembles in terms of radiation and associated convective environments are then shown in Section 4. In Section 5, changes in convective responses represented by cloud top heights, updraft velocities, and surface precipitation with enhanced aerosol loading are discussed. The environmental modulation of 115 the various aerosol impacts on the sea breeze convective system is simultaneously examined throughout Sections 4 and 5 and summarised in section 6.

RAMS model configuration
Two large ensembles of idealised numerical simulations of tropical sea breezes are conducted using the Regional 120 Atmospheric Modelling (RAMS) version 6.2.08 (Cotton et al., 2003;Saleeby and van den Heever, 2013). Idealised simulations are useful in that they are sufficiently complex to capture the storm systems of interest but are also sufficiently simple to allow for the isolation and evaluation of the critical physical processes at play, without the addition of unnecessary confounding factors often present in more complex case-study simulations. As our focus is on tropical sea breeze convection, the idealised simulations are initialised using conditions that are representative of equatorial coastal rainforest regions, more specifically, 125 the Cameroon rainforest region (Grant and van den Heever, 2014).
The RAMS model configuration used is here is identical to that used in Park et al. (2020). The boundary conditions, initialisation, and physical schemes are summarised in Table 1. The horizontal grid spacing is 1 km, and a 100 m vertical grid spacing near the surface is vertically stretched to 1 km near the model top. With this spatial resolution, the structure of the sea breeze circulation and associated convective clouds are well represented (Park et al., 2020). Each simulation is run from 0000 130 Local Time (LT) for 24 h using a 3 s time step. Output files are saved every ten minutes. RAMS is coupled to the Land-https://doi.org/10.5194/acp-2021-693 Preprint. Discussion started: 9 September 2021 c Author(s) 2021. CC BY 4.0 License.
Ecosystem-Atmosphere Feedback version 3 (LEAF-3), a fully interactive soil-vegetation-atmosphere parameterization (Lee, 1992;Walko et al., 2000). To simulate an idealised sea breeze circulation, two different surfaces, one land and one ocean, are separated by a straight coastline located at the centre of the domain. In our idealised setup, the western half of the domain represents the land region and is specified to be a rainforest with evergreen broadleaf trees and sandy clay loam soil type, 135 following that of Grant and van den Heever (2014). The eastern half of the domain is over ocean, and the sea surface temperature and horizontal gradient are kept fixed throughout the simulation. These relative locations of land and ocean are chosen arbitrarily and could have just as easily been the other way around.
All of the simulations are initialised with horizontally homogeneous thermodynamic and wind profiles. As described in Park et al. (2020), we make use of 130 different initial environmental conditions, where ten different lower-tropospheric 140 thermodynamic, wind, and surface properties (Table 2) are simultaneously perturbed across a range of values representative of tropical equatorial regions, and which were sourced from the sea breeze literature (Igel et al., 2018). These ten parameters are perturbed using maximin Latin Hypercube sampling (Morris and Mitchell, 1995), a space-filling algorithm, to ensure optimal coverage of the ten-dimensional parameter space with a minimum number of parameter combinations, and hence determines the minimum number of simulations that need to be conducted. As shown in Park et al (2020), this amounts to 130 145 simulations in each of the ensembles conducted here. It should be emphasised that these 130 different initial conditions represent more than 59,049 (=3 10 ) different initial conditions had only three values for each of the ten parameters been selected.
For the lower-tropospheric thermodynamic properties, five parameters defining the structure of the boundary layer and inversion layer are considered: (1) the boundary layer potential temperature; (2) the boundary layer relative humidity; (3) the boundary layer height; (4) the inversion layer depth; and (5) the inversion layer strength. The upper tropospheric 150 thermodynamic profiles above the inversion layer are identical for all 130 initial conditions and utilize the initial sounding of Grant and van den Heever (2014). The initial zonal wind speed without vertical shear is also considered and represents the sixth environmental factor examined. Finally, sensitivity to the variations in the surface characteristics are also examined including: (7) the soil saturation fraction; (8,9) the temperature difference between land/ocean surface and the atmospheric temperature at the lowest model level; and (10) the horizontal gradient of sea surface temperature (SST). The ranges of the ten 155 parameters tested are based on previous studies, as described in Igel et al. (2018) and Park et al. (2020), and are also shown in Table 2.
For the sake of simplicity, the aerosol type utilised in this study is restricted to single-mode, submicron, ammonium sulphate. We chose ammonium sulphate due to its ubiquity in the atmosphere and its ability to serve as CCN as a result of its high solubility. While the aerosol field is initialised horizontally homogeneously using appropriate profiles of aerosol (as 160 discussed below), aerosol particles are allowed to be redistributed via advection, convection, and nucleation following initialization. Aerosol sources and sinks, including the return of aerosols to the environment following evaporation, are all also incorporated (Saleeby and van den Heever, 2013). https://doi.org/10.5194/acp-2021-693 Preprint. Discussion started: 9 September 2021 c Author(s) 2021. CC BY 4.0 License.

Experiment setup
Two different ensembles of idealised simulations with different aerosol concentrations are performed for this analysis. 165 The nomenclature and description of each of the ensembles are summarised in Table 3. Here, "r" refers to aerosol-radiation interactions, and "On" means aerosol particles are allowed to interact with radiation. It should be noted that radiation is always fully interactive with both aerosols and hydrometeor species in both of the ensembles analysed here. We have conducted additional ensembles in which we have turned the aerosol-radiation interactions off to investigate the impacts on the deep convective mode, the results of which will be published in a subsequent paper. The rOn-500 ensemble is the suite of simulations 170 examined in Park et al. (2020) in which the model is initialised with surface aerosol concentrations of 500 mg -1 (blue line) and which decrease exponentially with height, as shown in Figure 1. To examine the impacts of increasing the aerosol loading on sea breeze convection, this ensemble of simulations is repeated, but with aerosol concentrations of 2000 mg -1 (red line in Figure 1) at the surface, and again using an exponentially decreasing profile. This model ensemble is referred to as "rOn-2000". The surface aerosol concentrations are chosen based on observations made in equatorial Africa (Andreae et al., 1992;175 Kacarab et al., 2020). The exponentially decreasing initial profiles are used in the interests of simplicity, and they soon evolve as a result of the processes described above. We shall use the terms "pristine" and "rOn-500" interchangeably, and "polluted" and "rOn-2000" interchangeably, throughout this manuscript. It is important to note that while 130 different initial conditions are implemented in each ensemble, that the polluted and pristine ensembles utilise the same 130 different initial conditions, thereby facilitating direct comparisons between the corresponding pristine and polluted ensemble pairs. Also, as the 130 initial 180 environmental conditions represent the ten-dimensional input parameter space, these pristine and polluted ensembles allow us to evaluate the ways in which direct and indirect aerosol impacts on sea breeze convection are modulated by the large number of different environments that support the development of such sea breeze regimes.

Analysis methodology
In a manner similar to Park et al. (2020), we apply an advanced statistical algorithm developed by Lee et al. (2011) 185 andJohnson et al. (2015) that includes statistical emulation (O'Hagan, 2006) and variance-based sensitivity analysis (Saltelli et al., 1999), over the ten-dimensional input parameter space. Due to its computational efficiency, this advanced statistical algorithm has been successfully utilised in several modelling studies that quantify the sensitivity of numerical model responses to a range of input parameters (Feingold et al., 2016;Igel et al., 2018;Wellmann et al., 2018Wellmann et al., , 2020Glassmeier et al., 2019;Marshall et al., 2019;Park et al., 2020). This algorithm incorporates the Gaussian process emulation (O'Hagan, 2006;190 Rasmussen and Williams, 2006) to build a statistical surrogate representation of complex cloud-resolving model responses of the parameters of interest. Over the ten-dimensional parameter space, the emulator estimates the cloud-resolving model responses of interest at untried input parameter combinations, thereby densely sampling the output of interest and thus allowing us to understand the relationship between the perturbed input parameter and output responses, without having to conduct the simulations representative of each set of perturbed parameters. We can then quantify the relative importance of the perturbed 195 https://doi.org/10.5194/acp-2021-693 Preprint. Discussion started: 9 September 2021 c Author(s) 2021. CC BY 4.0 License.
input parameters on the output of interest via variance-based sensitivity analyses. Further details of how this approach is used to determine the predominant environmental factors impacting tropical sea breeze convective systems are included in Park et al. (2020). We now begin our analysis by examining the morphology of the convection that develops within the pristine and polluted ensemble of simulations.

Basic description of the sea breeze simulations 200
In this section, an overview of the convective morphology and development within our tropical sea breeze simulations is provided. A sea breeze circulation develops after sunrise (0600 LT), and convergence along the leading edge of the sea breeze front becomes evident at the coastline, where the highest land-sea thermal contrast is established (not shown).
Throughout the daytime hours, and even shortly after sunset (1800 LT), the sea breeze front continues to propagate further inland, which is in keeping with classical theory of such baroclinic circulations. The sea breeze front is observed to develop in 205 all 260 simulations comprising the rOn-500 and rOn-2000 ensembles (130 simulations in each ensemble), the location of which is detected at every output timestep using an identification algorithm developed by Igel et al. (2018).
Two types of convection are evident over land between 1200-1800 LT (Figure 2), as is often the case with tropical sea breeze systems. Ahead of the leading edge of the sea breeze, daytime heating and mixing induce cumulus convection, whereas along the leading edge of the sea breeze front, where low-level air parcels may be lifted to the level of free convection 210 through low-level convergence, sea breeze-initiated deep convection may occur. As shown in Park et al. (2020), the strength of the sea breeze circulation and the associated convergence along the leading edge vary strongly as a function of the initial environmental conditions. Such variations result in a range of shallow through deep convective clouds being produced in association with the sea breeze front. The left column of Figure 2 displays one example from the rOn-2000 ensemble, where both the daytime cumulus convection ahead of the sea breeze and the convection developing along the sea breeze front are 215 shallow (cloud top heights < 4 km AGL), and remain so throughout the simulation. While both types of convection remain shallow in this scenario, the sea breeze-initiated convection is characterised by slightly deeper clouds and stronger updrafts than the daytime cumulus convection forming ahead of this line. In the right column of Figure 2, the development of deep convection (cloud top heights > 7 km AGL) along the sea breeze front in another ensemble member is shown. In this case, the convection initiated by the sea breeze is significantly deeper than the previous case, and this sea breeze-initiated "deep mode" 220 is accompanied by significantly heavier precipitation and stronger updrafts than those associated with the sea breeze-initiated shallow modes. The vast majority of the ensemble members fall into "shallow mode" scenario, whereas only a handful of the ensemble members display "deep mode" convection. Throughout the rest of the manuscript we will refer to the convection developing out ahead of the sea breeze front as "daytime cumulus convection", and to the convection developing along the sea breeze front as the "sea breeze-initiated convection". The terms "shallow mode" and "deep mode" will be used to distinguish 225 the different types of sea breeze-initiated convection. We now examine the impacts of aerosols on the developing convective environments (Section 4) and the resulting convection both ahead of, and along, the sea breeze front (Section 5).

Aerosol impacts on the convective environment
In order to examine the impacts of aerosols on the convective environment, we first examine the effects of varying aerosol concentrations on the surface radiation budget. As ammonium sulphate aerosols are allowed to scatter and absorb 230 radiation as a function of the wavelength, median radius, and relative humidity within RAMS (Saleeby and van den Heever, 2013), less incoming solar radiation is expected to reach the surface. These aerosol direct effects are addressed by analysing the differences in surface downwelling shortwave and surface upwelling longwave radiation between rOn-500 and rOn-2000 for clear-sky columns. To identify the clear-sky columns, we have assumed that the total condensate mixing ratio is smaller than 0.01 g kg -1 at all vertical levels within the column. As the aerosol number concentrations are increased between the pristine 235 and polluted ensembles, the clear-sky downwelling shortwave radiation decreases throughout the atmosphere over land and ocean, with the maximum difference occurring at the surface for all 130 corresponding pairs of simulations of the rOn-500 and rOn-2000 ensembles (Figures 3a, b). The daytime-averaged (0600-1800 LT) surface shortwave radiation difference between rOn-500 and rOn-2000 is 81.2 W m -2 for the ensemble average, and has a standard deviation of 6.5 W m -2 .
The surface upwelling longwave radiation (Figures 3c and 3d) reflects the aerosol-induced changes in incoming solar 240 radiation. Due to the much lower heat capacity of the land surface compared to the ocean surface, the longwave emission significantly decreases over the interactive land surface as it rapidly responds to the reduction in shortwave radiation in the polluted ensemble. With less surface upwelling longwave radiation, the air above the land becomes cooler (Figure 3e) for the conditions of enhanced aerosol loading compared to the cleaner conditions. Over the ocean regions, the SST is fixed throughout the duration of each ensemble member simulation (although it can be different between ensemble members based on the 245 environmental parameter being perturbed). As such, the potential temperature of the lowest level air does not change significantly over the ocean since the longwave radiation emitted from the surface remains almost the same ( Figure 3f).
However, even if the SST had been allowed to be interactive, given the higher heat capacity of the ocean, the upwelling longwave radiation from the ocean surface would not respond as quickly to the aerosol-induced reduction in shortwave radiation as that over the land surface. As a result of these aerosol interactions with the radiation, the ensemble-mean surface 250 temperature difference between the ocean and the land is less in the polluted case when compared with the pristine case ( Figures 3e and 3f), which has important implications for the strength of the thermally driven sea breeze circulation. These are discussed in more detail below.
The aerosol-induced reduction in incoming solar radiation reaching the surface also impacts the surface fluxes. Figure   4 displays the temporal evolution of the surface sensible and latent heat fluxes of the environment ahead of the sea breeze 255 front, averaged over all 130 simulations in each ensemble. Both the ensemble-mean sensible and latent heat fluxes are reduced in rOn-2000 compared with rOn-500 as a result of the enhanced aerosol loading. The maximum aerosol-induced reduction is 45% and 60%, and the mean is 19% and 15%, for the sensible and latent heat fluxes, respectively. This reduction in incoming solar radiation and associated surface flux responses due to enhanced aerosol loading is in close agreement with the findings from previous studies (Yu et al., 2002;Koren et al., 2004;Feingold et al., 2005;Jiang and Feingold, 2006;Zhang et al., 2008;260 https://doi.org/10.5194/acp-2021-693 Preprint. Discussion started: 9 September 2021 c Author(s) 2021. CC BY 4.0 License. Grant and van den Heever, 2014). These reductions in sensible and latent heating will negatively impact the convective boundary layer by limiting the heating and moistening of this layer. The surface-based mixed layer depth, defined here as the level above the surface at which the vertical gradient of the potential temperature first exceeds 2 K km -1 , decreases in rOn-2000 due to this reduction in surface sensible heat flux and associated turbulent mixing. The percentage difference of the mean surface based-mixed layer depth ahead of the sea breeze front between rOn-500 and rOn-2000 is shown in Figure 5a, where it 265 is evident that the mixed layer in each member of the polluted ensemble is shallower than the corresponding mixed layer in the pristine ensemble, with differences ranging from 1.6% to 23%.
The results of this section thus far indicate that two of the three convective ingredients (Doswell, 1987), lowertropospheric moisture (as represented by the latent heat fluxes) and instability (as represented by the mixed layer depth), become less favourable for convective development in all of the rOn-2000 members compared with rOn-500 members, as a 270 result of the aerosol-induced reduction of incoming solar radiation. The third convective ingredient is vertical lift. In these simulations, lift is provided both by the sea breeze circulation itself and the convergence along the sea breeze front. Classical sea breeze theory dictates that, to first order, the faster the sea breeze moves, the further inland the sea breeze travels during the day and the stronger the convergence along the sea breeze front, and hence the greater the vertical lift along the front. Here we examine the maximum inland extent of the sea breeze front as an indicator of sea breeze intensity and hence vertical lift 275 ( Figure 5b). The maximum inland extent of the sea breeze front is identified as the last inland location of the sea breeze front detected by the sea breeze front algorithm (Igel et al., 2018). The percentage difference of the maximum sea breeze inland extent between the polluted and pristine ensembles varies from −2% to −53% across the ensembles, and it is evident from Figure 5b that the sea breeze extent is significantly less in rOn-2000 than rOn-500 for each and every one of the ensemble pairs. It is therefore evident from these results that the third primary element of convection, vertical lift, is also reduced with 280 increasing aerosol loading. As all three convective ingredients (moisture, instability, and lift) are reduced in association with aerosol direct effects on the sea breeze circulation and the convective boundary layer, we could therefore expect the presence of increased aerosol concentrations to reduce the intensity of the sea breeze convective system in all of the different environments represented by the ensemble members. However, this assumes that aerosol indirect effects do not significantly contribute to convective storm intensity, thereby offsetting some of these direct effects. We investigate these effects in more 285 detail in the next section.
Finally, a reduction in the three convective ingredients with enhanced aerosol loading was also observed in Grant and van den Heever (2014). However, in that study, they only utilised one set of initial environmental conditions. Therefore, one might wonder whether such findings are applicable to the multiple other environments known to support sea breeze systems, or whether various sea breeze environments produce a different response to aerosol loading. Since rOn-500 and rOn-2000 each 290 have 130 different initial conditions, and as these initial conditions are representative of the wide range of environmental parameter values previously identified as being responsible for generating sea breeze systems, these results suggest that the aerosol-induced weakening of sea breeze-initiated convection through direct effects is indeed a robust result, and one that occurs irrespective of the sea breeze environment.

Impacts of enhanced aerosol loading on continental convection 295
In the previous section, we demonstrated that the three convective ingredients of moisture, instability, and lift, are reduced in the polluted ensemble, when compared with the pristine ensemble, and that this is due to the direct impacts of enhanced aerosol loading on incoming solar radiation. Now we seek evidence of the impacts of increased aerosol number concentrations on the intensity of the convection within tropical sea breeze convective system as a result of both direct and indirect aerosol effects. More specifically, we investigate the impacts of enhanced aerosol loading on cloud top heights, 300 convective updraft velocities, and precipitation, all of which are common indicators of convective intensity (e.g., Nesbitt and Zipser, 2003).

Cloud top heights
First, we examine the impacts of aerosol loading on cloud top height by analysing the following: (1) the frequency distribution of the low (cloud top height < 4 km AGL), middle (cloud top heights between 4-7 km AGL), and deep (cloud top 305 heights > 7 km AGL) clouds and (2) the maximum cloud top height during the afternoon hours (1200-1800 LT) anywhere over land. This includes clouds both ahead of and along the sea breeze front. To identify cloud top heights, we start at the bottom of each model column and locate the lowest and highest level where the total condensate exceeds 0.1 g kg -1 . This column is then marked as a cloudy column when all the points between these two levels contain total condensate mixing ratios greater than 0.1 g kg -1 . We regard the highest level of the cloudy column as the convective cloud top height. This check on 310 cloud contiguity means that we do not inadvertently take into account those situations in which there may be multiple clouds at different levels within the column.
As shown in Figure 6a, the vast majority of ensemble members in the pristine ensemble are dominated by low clouds, with cloud tops less than 4 km AGL. In other words, only shallow convective clouds develop both ahead of and along the sea breeze front in most of the ensemble members, and while all of the environmental conditions tested here support the 315 development of sea breeze convection, most do not support the development of the sea breeze-initiated deep convective mode.
This low cloud dominance is also observed for the polluted ensemble ( Figure 6a). Specifically, among the 130 simulations in each ensemble, there are 104 and 113 simulations with low clouds only in rOn-500 and rOn-2000, respectively.
The suppression of sea breeze convective intensity in rOn-2000, when compared with rOn-500, is evident in the maximum cloud top heights (Figure 6c), as well as in the percentage difference in the maximum cloud top height (Figure 6d). 320 It is clear from these figures that the maximum cloud top heights are always less in rOn-2000 than in rOn-500, except for three sea breeze-initiated deep mode cases (Tests 27, 79, and 104). In other words, the maximum cloud top height of the sea breezeinitiated shallow mode and the daytime cumulus convection always decreases with more aerosol loading, regardless of different initial conditions. For the sea breeze-initiated deep mode cases, however, the sign of the maximum cloud top height response to enhanced aerosol loading varies with the initial environmental conditions. Furthermore, Figure 6c demonstrates 325 that while there are 12 cases (Tests 27,29,37,41,59,65,75,79,100,104,106, and 120) with a deep mode in rOn-500, only https://doi.org/10.5194/acp-2021-693 Preprint. Discussion started: 9 September 2021 c Author(s) 2021. CC BY 4.0 License. 7 of these 12 cases (Tests 27,59,65,75,79,104, and 120) have a deep mode when the aerosol loading is enhanced in rOn-2000. The reasons for this are discussed below.

Convective updraft velocities
In order to better understand the dynamical response of convection to enhanced aerosol loading, updrafts developing 330 over land between 1200 and 1800 LT with velocities greater than 1 m s -1 are analysed. The maximum updraft velocity ( Figure   7a) represents the intensity of the most vigorous continental convection. In all of the simulations of the pristine and polluted ensembles, the strongest updrafts are always found in association with the convergence along the sea breeze front, irrespective of whether the convection produced by this process is shallow or deep. Aerosol-induced suppression of the maximum updraft velocities is evident in most of the member pairs (Figure 7a), however, there are 10 ensemble pairs (Tests 18,28,36,67,76,335 79,83,84,111,and 124) in which the maximum updraft velocities are stronger in rOn-2000. As such, enhanced aerosol loading may therefore produce weaker or stronger maximum updraft velocities in the sea breeze-initiated convection. As this aerosolinduced response varies as a function of the initial environmental conditions, it is environmentally modulated. The mean In the following subsections, we now seek to determine whether the key environmental parameters identified in Park et al. (2020) as the predominant parameters driving the convective updrafts in rOn-500 are impacted by aerosol loading, and if so, why that is the case. We first examine the sea breeze-initiated convection, followed by the daytime cumulus convection developing ahead of the sea breeze front. We examine these two types of convection separately as they are driven by different 350 processes, with convergence and convective instability along the sea breeze front being critical to the sea breeze-initiated convection, and daytime surface heating, boundary layer mixing, and thermal buoyancy driving the convection developing ahead of the sea breeze.

Sea breeze-initiated convective updrafts
Due to the steep vertical velocity gradients between sea breeze-initiated shallow and deep convective modes in rOn-355 500, Park et al. (2020) found that constructing a statistically robust emulator for the maximum updraft velocities was not feasible. The same is true here for the rOn-2000 simulation, and thus we have had to rely on other analysis methods to determine why deep convection is absent in 5 out of 12 cases in the polluted scenario. As noted in Park et al. (2020), the 12 cases with https://doi.org/10.5194/acp-2021-693 Preprint. Discussion started: 9 September 2021 c Author(s) 2021. CC BY 4.0 License. sea breeze-initiated deep modes in rOn-5000 all have initial boundary layer potential temperatures greater than 297 K as a common parameter. Given the range of initial boundary layer potential temperatures tested in this study (Table 2), this critical 360 threshold for the deep mode falls at the upper end of this range and is the reason for why the vast majority of ensemble members have shallow modes. Figure 8 is a scatter plot relating various convective characteristics to the initial boundary layer potential temperature, for all 130 ensemble members. It is clear from this figure that the sea breeze-initiated deep convective mode in rOn-500 occur in all of the ensemble members in which the initial boundary layer potential temperature is 297 K or greater (Figures 8c and 8e), and in which the mixed layer CAPE is greatest (Figure 8a). A similar dependence on the initial boundary 365 layer potential temperature threshold is shown for rOn-2000 (Figures 8d and 8f). However, in rOn-2000 the threshold above which the deep convective mode occurs is 299 K, 2 K greater than that in rOn-500. As demonstrated above, the surface temperatures are reduced in the presence of enhanced aerosol loading in the rOn-2000 ensemble, thereby reducing the convective potential in these simulations. Therefore, greater initial boundary layer potential temperatures are necessary to offset the aerosol-induced surface cooling in the polluted ensemble, thereby providing sufficient environmental CAPE (Fig.  370 8b) to support the production of deep convection along the sea breeze front.

Daytime cumulus convection updrafts
Unlike the sea breeze-initiated convective updrafts, constructing a robust emulator was possible for the updraft velocities of the daytime cumulus convection forming ahead of sea breeze front. We can therefore draw on the emulator results in our analysis of these updrafts. The bar graphs in Figures 9a (rOn-500) and 9b (rOn-2000) indicate how much of the variance 375 in the median updraft velocity is explained by the individual perturbations to the ten environmental parameters tested. Each stacked bar graph's height refers to the summation of first-order contributions of the ten parameters to the median updraft velocity. Any blank space remaining above the bar indicates the contributions made by higher-order nonlinear interactions involving multiple parameters. It is evident from the left most bar graph (labelled "Overall") that the same two parameters, the soil saturation fraction (dark grey) and the inversion layer strength (pink), are the predominant contributors to the median 380 updraft velocities in rOn-500 and rOn-2000, although their percentage contributions differ. In order to understand the impacts of aerosols on these relative contributions, we first need to understand the processes driving these two predominant contributions. We now turn to the impacts of enhanced aerosol loading on the roles of the soil saturation fraction and the inversion layer strength. When comparing Figure 9a with 9b, the relative percentage contribution of soil saturation fraction to updraft 415 variability decreases from 78% to 68% with enhanced aerosol loading, whereas that of the inversion layer strength increases from 4% to 8%. The soil saturation fraction plays an important role in the daytime cumulus convection ahead of the sea breeze front through its control of the magnitude of the surface sensible and latent heat fluxes, and its partitioning between them. As discussed in Section 4, the aerosol-induced reduction in surface downwelling shortwave radiation in the polluted ensemble reduces the incoming shortwave radiation, the surface longwave emission, the associated surface sensible and latent heat completely absent in WET soil regimes regardless of aerosol loading. Comparing the emulator-predicted median updrafts between the pristine and polluted conditions (Figure 9c), the trends in the relationship between the soil saturation fraction and the updraft velocities are very similar, but median updrafts are stronger in rOn-500 due to the stronger sensible heat fluxes.
In rOn-500, inversion layer strength is the second-most important parameter for the mixed-layer depth and the median updraft velocity. When the initial inversion layer is weaker, the lower troposphere is less stable, which promotes daytime 430 turbulent mixing and leads to stronger vertical motions. While the inversion layer is still the second-most important parameter for the median updraft in rOn-2000, its relative importance is increased from 4% to 8%. As the contributions made by the rest of the factors to the velocity variance remains much the same between the pristine and polluted conditions (Figures 9a and b), this increase primarily reflects the reduced contribution by the soil saturation fraction and the concomitant increase in the contribution of the inversion layer strength to boundary layer development, and hence, to boundary layer updrafts. 435 To summarize, the daytime cumulus updraft velocities become weaker with enhanced aerosol loading. This is consistent with the reduction in surface fluxes and mixed layer depth under polluted conditions described in section 4. The ten-dimensional variance-based sensitivity analysis described here demonstrates that the relative importance of the soil saturation fraction and inversion layer strength in contributing to the updraft velocity variance varies as a function of aerosol loading. The degree of aerosol-induced weakening of the updrafts is thus modulated by the aerosol-environmental feedbacks. 440

Surface accumulated precipitation
The changes in surface accumulated precipitation with increased aerosol loading are now analysed. Figures 10a and   10b display land-averaged accumulated surface precipitation at 1800 LT in rOn-500 and rOn-2000, respectively. The impact of enhanced aerosol loading on precipitation is represented by the percentage differences between rOn-2000 and rOn-500 ( Figure 10c). The percentage differences are only determined for those simulations that produce at least 0.1 mm of land-445 averaged accumulated precipitation in rOn-500 (36 out of 130 simulations). Only 19 simulations produce more than 0.1 mm of area-averaged precipitation in both rOn-500 and rOn-2000, and no simulation in rOn-2000 produces more than 0.1 mm of precipitation when its corresponding ensemble member in rOn-500 does not produce precipitation. Figure 10c distinctly shows that the accumulated precipitation is reduced in all 36 of the rOn-2000 precipitating ensemble members when compared with their corresponding counterparts in rOn-500, the percentage differences of which vary from −16.3% (Test75) to −96.7% 450 (Test110).
We now examine the differences in the bulk microphysical processes contributing to the differences in the surface precipitation in the 36 precipitating ensemble pairs (Figure 11). The following five processes contribute to the production of precipitation at the surface: 1) cloud-to-rain: cloud water transferred to rain through collection; 455 2) vapor-to-rain or rain-to-vapor: vapor diffusion onto rain (gain term) or evaporation from rain (loss term); 3) melting of ice: ice mass transferred to rain via thermodynamic melting as the ice species fall below the freezing level; https://doi.org/10.5194/acp-2021-693 Preprint. Discussion started: 9 September 2021 c Author(s) 2021. CC BY 4.0 License. 4) ice-to-rain: collisional ice melting due to collection of warmer rain; and 5) rain-to-ice: rainwater collected by ice species through riming. 460 These process rates are averaged over the land domain between 1200 and 1800 LT. Since the cloud frequency over land is heavily weighted by the daytime cumulus convection mode, the averaged mixed-phase process contributions are relatively small. As most of the differences in the processes contributing to differences in surface rainfall occur below 10 km AGL, we restrict the axes in Figure 11 to between the surface and 10 km AGL to enhance figure clarity. It should also be noted that the freezing level varies from simulation to simulation due to the different initial temperature and moisture profiles. It is evident 465 that in all simulations, regardless of the different initial conditions, the average warm rain production rates (i.e., cloud-to-rain) are greater in rOn-500 than rOn-2000. Similarly, average rain evaporation rates (i.e., rain-to-vapor) are also greater in magnitude in rOn-500 than rOn-2000, although they are less than the cloud-to-rain production term. In all of the ensemble pairs, except for Test68 and Test120, where the ice-to-rain process is the most dominant gain term, the warm rain production rates are greater than cold rain production rates throughout the whole profile. Figure 11 therefore demonstrates that, in the 470 mean, precipitation suppression in the polluted environment primarily occurs as a result of the suppression of warm rain formation. Furthermore, as this response is evident across 34 out of 36 precipitating members, it appears that these warm rain trends are independent of the environmental conditions, although the magnitudes in the response certainly do vary with environment. No common characteristics are, however, found in the changes in cold rain production with enhanced aerosol loading. As such, our results suggest that the environment modulates cold rain response to enhanced aerosol loading. 475 Examining the most noticeable differences in the initial conditions of the two extrema rainfall difference cases (Test75 and Test110) provides some insights into the environmental controls on these precipitation responses to aerosol loading ( Figure   12). Test75, which demonstrates the least reduction in precipitation between the pristine and polluted cases, has a very moist, warm and deep lower troposphere that develops in association with a soil saturation fraction of 0.89, a boundary layer relative humidity of 89.8% and a boundary layer potential temperature of 299 K. On the other hand, Test110, which shows the greatest 480 difference in precipitation between the pristine and polluted cases, has only moderately moist and warm initial conditions with a soil saturation fraction of 0.44, a boundary layer relative humidity of 84.2%, and a boundary layer potential temperature of 288 K. The warmer, moister conditions in Test75 will (a) allow for more effective aerosol nucleation and the subsequent growth of the more numerous, smaller cloud droplets that form in the polluted case compared to drier, polluted conditions, and (b) more effectively limit the impacts of evaporation on raindrop populations than in the drier, polluted cases. As such, is 485 appears that the warmer, moister conditions assist in muting the impacts of aerosols on the warm rain production, resulting in rain gain and loss terms that are of similar magnitude in the polluted and clean conditions, and hence relatively small aerosolinduced reductions in net rain production. Drier conditions, on the other hand, enhance the aerosol-induced differences in the rain source and sink terms through less efficient droplet growth and enhanced raindrop evaporation. Similar trends can be found for other simulations with warm, moist boundary layers, such as Test84, and drier, colder boundary layers, such as 490 Test17 ( Figure 12). Now, when the initial boundary layer temperature and relative humidity are similar between ensemble members, then lower soil saturation appears to be associated with greater aerosol-induced precipitation suppression. For (soil saturation = 0.83) compared with that in Test40 (soil saturation fraction = 0.64) appears to limit the aerosol-induced 495 precipitation suppression. The same could be applied to Test43 and Test59 (Test43: 298.1 K and 76.4%;Test59: 299.8 K and 76.2%) where the soil saturation fraction is higher in Test43 (soil saturation fraction = 0.68) than Test59 (soil saturation fraction = 0.57). However, the same cannot be said for all 36 precipitating pairs, and as such, the extent to which aerosols suppress precipitation therefore appears to be a complex function of different environmental variables.
In summary, under enhanced aerosol concentrations, the reduction of surface accumulated precipitation is observed 500 irrespective of the initial environmental conditions. While the trends of aerosol-induced decreases in precipitation is robust across the sea breeze environments investigated here, the magnitude of precipitation reduction varies significantly across the ensemble members, and appears to be a relatively complex function of the environment.

Summary and discussion
The primary goals of this study have been 1) to investigate how microphysically and radiatively active aerosol 505 particles influence the wide range of convective environments supporting sea breeze convective regimes and the convective cloud characteristics developing under these regimes, and 2) to determine whether the convective environment modulates these aerosol impacts on the convection. In order to achieve our goals, we conducted two large numerical model ensembles, where the only difference between the ensembles was the aerosol loading. Each of these two ensembles contained 130 members initialised with 130 different initial conditions, representing the simultaneous perturbation of ten thermodynamic, wind, and 510 surface properties. The selection of these 130 initial conditions was based on a statistical space-filling method (Morris and Mitchell, 1995) and the clean (rOn-500) and polluted (rOn-2000) ensembles were then compared using the statistical framework developed by Johnson et al. (2015).
The comparison between rOn-500 and rOn-2000 demonstrated that aerosol direct effects resulted in less shortwave radiation reaching the surface. The associated longwave emission from the land surface was also reduced, thereby producing 515 a cooler land surface, a smaller land-sea thermal contrast and a weaker sea breeze circulation. The land-surface sensible and latent heat fluxes, the surface-based mixed layer depth, and the sea breeze inland extent were all subsequently reduced in the polluted scenario as a result. The three ingredients of moist convection, i.e., moisture, instability, and lift (Doswell, 1987), were all therefore reduced in the presence of aerosol direct effects. The reduction in these three convective ingredients as a result of enhanced aerosol loading was found in all of the 130 ensemble simulation pairs, and therefore suggest a robust result 520 that is independent of environmental conditions. The magnitude of the reduction does vary as a function of the initial environmental. These results extend those of Grant and van den Heever (2014), in which a similar sensitivity of tropical sea breeze convection to aerosol loading was demonstrated, albeit for only one set of initial conditions. https://doi.org/10.5194/acp-2021-693 Preprint. Discussion started: 9 September 2021 c Author(s) 2021. CC BY 4.0 License.
The resulting aerosol-induced changes in convection over land in the afternoon (1200-1800 LT) were then examined for convective clouds developing along the sea breeze front. In the vast majority of ensemble members in rOn-500 and rOn-525 2000 the sea breeze-initiated convection remained stronger and deeper than daytime cumulus convection forming ahead of the sea breeze front. The shallow mode of the sea breeze-initiated convection showed an overall reduction in maximum cloud top height and maximum updraft velocity with enhanced aerosol loading, despite different initial environmental conditions, suggesting that the aerosol-induced suppression of shallow convective velocities is robust for the wide range of environments tested here. However, the sign and magnitude of changes to the sea breeze-initiated deep convective velocities in response to 530 enhanced aerosol loading varied as a function of initial environmental conditions, with some member ensembles showing an increase in updraft velocity, while others showed a decrease or no change to the strength of the updraft. This demonstrates that aerosol impacts on the deep convective updrafts are environmentally modulated. Of the ten environmental parameters perturbed in the initial conditions, the initial boundary layer potential temperature was shown to have important implications for the deep convective mode. The sea breeze-initiated deep mode was only observed in those rOn-2000 ensemble simulations 535 in which the initial boundary layer potential temperatures were greater than 299 K, compared with a corresponding threshold of 297 K in rOn-500. With enhanced aerosol loading inducing cooler and more stable near-surface environments in rOn-2000, the warmer initial boundary layer potential temperatures were found to be necessary to facilitate deep convection through higher CAPE.
The aerosol-induced reduction in surface fluxes and mixed-layer depth also led to the suppression of daytime cumulus 540 convection ahead of sea breeze front, irrespective of different initial conditions, and thus appears to be a robust response.
Furthermore, a ten-dimensional variance-based sensitivity analysis, combined with statistical emulation, revealed that the relative importance of soil saturation fraction on the surface fluxes and mixed-layer depth, and hence on the daytime cumulus updrafts, was reduced in the presence of enhanced aerosol loading. A nonlinear sensitivity of boundary layer updraft velocities to soil saturation fraction was also found, with the greatest convective updraft response observed in the MID soil saturation 545 fraction regime, followed by a moderate response in the DRY soil saturation fraction regime and no response in the WET soil regime. These sensitivities were found to exist in both the clean and polluted environments. The sensitivity analysis results therefore emphasise the importance of considering atmosphere-radiation-land feedback processes in the modelling studies of aerosol-cloud interactions.
Finally, changes in the surface-accumulated precipitation to increased aerosol loading were analysed. Again, a 550 consistent aerosol-induced reduction was observed across the entire ensemble of model simulations. Spatiotemporally averaged rain production and evaporation rates exhibited robust aerosol-induced behaviours. While the trends in precipitation reduction were found to be consistent irrespective of the environment, the magnitude of precipitation suppression was found to be modulated by environments.
Many of the past studies assessing aerosol impacts on deep convection have tended to neglect the role of aerosol 555 direct forcing in order to focus on aerosol indirect effects. However, our results indicate the importance of considering both aerosol direct and indirect effects when determining the impacts of aerosols on convective systems, and suggest that future https://doi.org/10.5194/acp-2021-693 Preprint. Discussion started: 9 September 2021 c Author(s) 2021. CC BY 4.0 License. studies should consider both aerosol direct and indirect effects to fully understand aerosol impacts on deep convection. A subset comprised of the members of the clean and polluted ensembles that produced deep convection has been further analysed to assess the relative roles of aerosol direct and indirect effects specifically on deep convection, as well as to assess the 560 robustness of warm and cold phase invigoration under varying environments. These results are described in Park and van den Heever (2021).
In this study, we only tested a single aerosol species, i.e., ammonium sulphate, which is a strong scatterer, but a poor absorber of radiation. It is worth noting that absorbing aerosols, such as mineral dust or smoke, are also substantial contributors to aerosol emissions in equatorial Africa (Adams et al., 2012;Chakraborty et al., 2015). The potential importance of the 565 interactions of absorbing aerosol with radiation and microphysics has been reported in previous studies. For instance, Saide et al. (2015) found that smoke absorption tends to either burn off clouds or enhance the capping inversion, depending on the smoke's location. Such radiation absorption in the presence of other aerosol species may therefore produce different effects on sea breeze convection to those reported here for sulphate. This study's perturbed parameter ensemble approach could be extended to such studies considering absorbing aerosol species. 570 Finally, we conclude that the results of this study underscore the importance of considering both aerosol direct and  Zonally open-radiative (Klemp and Wilhelmson, 1978), meridionally periodic Initialisation 130 horizontally homogeneous thermodynamic and wind profiles where the ten parameters are simultaneously perturbed within the indicated value ranges (Igel et al., 2018;Park et al., 2020) Random potential temperature perturbations within the lowest 500 m of the domain with a maximum perturbation of 0.1 K at the surface are used to initiate convection