Numerical simulations of the impact of ultrafine cloud condensation nuclei (CCN) on deep convection are analyzed to investigate the idea proposed by Fan et al. that addition of ultrafine CCN to an otherwise pristine environment leads to convective invigoration. The piggybacking methodology is applied, allowing rigorous separation of the impact of aerosols from different flow realizations that typically occur when even a small element of the model physics or modeling setup is changed. The setup follows the case of daytime convective development over land based on observations during the Large-Scale Biosphere–Atmosphere (LBA) experiment in Amazonia. Overall, the simulated impacts of ultrafine CCN are similar to the previous study by the authors on the impact of pollution on deep convection. There is no convective invigoration above the freezing level, but there is a small invigoration (increase in vertical velocities) below due to the supersaturation and buoyancy differences in conditions with additional ultrafine CCN compared to unperturbed pristine conditions. As in the previous study, the most significant impact is on the upper-tropospheric convective anvils that feature higher cloud fractions in conditions with ultrafine CCN. The increase comes from purely microphysical considerations as the increased cloud droplet concentrations from ultrafine CCN lead to increased ice crystal concentrations and, consequently, smaller sizes and fall velocities, and longer residence times. Mesoscale organization due to low-level shear has a small effect on the simulated ultrafine CCN impacts. Finally, an alternative explanation of increased lightning above oceanic shipping lines seen in satellite observations and argued to result from convective invigoration is provided.
The impact of atmospheric aerosols serving as cloud condensation nuclei (CCN) and ice-forming nuclei on clouds and cloud systems continues to be discussed in the literature; see recent reviews by Tao et al. (2012), Boucher et al. (2013), Fan et al. (2016), and Mülmenstädt and Feingold (2018). The impact concerns many aspects of cloud dynamics and microphysics as well as interactions between clouds and their large-scale environment. Observations, theory, and numerical modeling have all been used to advance the understanding of those aspects. This paper concerns a specific facet of cloud–aerosol interactions, the so-called convective invigoration in polluted environments. The argument for the invigoration (e.g., Rosenfeld et al. 2008, and references therein) conjectures that inefficient warm-rain processes below the freezing level in a polluted environment result in increased amounts of condensed water transported through the melting level that subsequently freeze aloft, release the latent heat of freezing, and invigorate convective updrafts (see a more detailed discussion in the next section). More recently, Fan et al. (2018) applied observations from the Amazon field project and model simulations to suggest that adding small (ultrafine) CCN to an otherwise clean aerosol environment provides additional invigoration. The modeling study reported here addresses specifically the small CCN impact on convective dynamics.
As far as observations are concerned, correlation between aerosols and deep convection does not imply causality: aerosols and convection may simply covary due to meteorological factors, see examples in Nishant and Sherwood (2017) and Varble (2018). Koren et al. (2010b) recognize the role of meteorology and attempt to separate the impact of aerosols from meteorological factors in satellite observations of deep convection over the tropical Atlantic. However, observations of the cloud fraction and cloud-top height changes do not necessarily imply changes of the convection strength. This is because these cloud properties may change solely in response to the modification of the cloud microphysics (e.g., Morrison and Grabowski 2011). Moreover, Grabowski (2018) argues that the accuracy of atmospheric observations is insufficient to separate impacts of aerosols from meteorological effects that independently affect moist convection. In other words, it is impossible to find two situations that are exactly the same with convection developing differently only because of the impact of aerosols. This is possibly a problem in the observational part (and maybe modeling as well) of the Fan et al. (2018) study because the observations were collected in an area known for the presence of river breezes and spatially variable mean rain accumulations (e.g., Santos et al. 2014, and references therein).
Modeling can eliminate meteorological ambiguity because the same environmental conditions (i.e., the initial sounding and the forcing) can be used in model simulations with the only difference coming from modified aerosols. However, the simulations need to be long enough to include realistic interactions among an ensemble of clouds and their environment, and to span an appropriate range of meteorological conditions. Such long simulations [multiyear as in Seifert et al. (2012) and multiweek as in Fan et al. (2013)] show strong microphysical impact on condensate amounts (e.g., increased upper-tropospheric cloud fractions) and a small impact on the surface precipitation. Short simulations applying idealized cloud initialization [e.g., such as in Khain and Pokrovsky (2004), Khain et al. (2005), and Teller and Levin (2006)], and applying simplified frameworks [e.g., in two-dimensional geometry as in Fan et al. (2009)] may not be sufficient. Short simulations of convective clouds also face additional complication of separating the impact of aerosols from different flow realizations. This is because convection is highly nonlinear, and changing a single element of the model physics or including small perturbations to the initial conditions leads to different flow realizations. As a result, clouds develop in different locations and may follow different life cycles. This makes it difficult to evaluate underlying causes and mechanisms that drive differences between the two simulations.
One way to eliminate the impact of different flow realizations is to apply a technique referred to as piggybacking; see a recent review in Grabowski (2019). The key idea is to apply two sets of thermodynamic variables (the temperature, water vapor, and all aerosol, cloud, and precipitation variables) in a single cloud field simulation. One set is coupled to the dynamics and drives the simulated flow (the driver), and the other set piggybacks the flow; that is, thermodynamic variables are carried by the flow but they do not affect it (the piggybacker). Piggybacking allows extracting the impact of a physical process (i.e., the aerosols in the case considered here) from the effects of different flow realizations.
Grabowski and Morrison (2016; GM16 hereafter) applied piggybacking to study differences in the cloud field characteristics in simulations assuming either pristine or polluted CCN. The results showed a relatively small impact on the cloud dynamics (i.e., a small increase of the strongest updraft contribution to the updraft statistics, Fig. 11 therein) and surface precipitation (10% to 15% increase in the polluted case, Fig. 6 therein). The cloud dynamics was affected below the freezing level because the smaller supersaturations in the polluted case led to larger cloud buoyancies (Fig. 8 therein). There was no invigoration above the freezing level. The most significant impact was on the upper-tropospheric cloudiness. The polluted cases featured significantly larger convective anvil cloud fractions because smaller ice crystals have smaller fall velocities and thus extended anvil lifetime (Figs. 1 and 2 therein). The key mechanism involves higher ice crystal concentrations resulting from higher cloud droplet concentrations within convective drafts, see Fig. 14 therein. This agrees with model simulations discussed in Morrison and Grabowski (2011) and in Chen et al. (2017), and is consistent with the impact documented in observations (e.g., Koren et al. 2010a). However, Chen et al. (2017) focus their interpretation of simulation results on the invigoration, whereas the interpretation of Koren et al. (2010a) involves a combination of the postulated invigoration and the environmental wind impact on the anvil spread, not the effect of smaller ice particles.
GM16 considered two sets of ensemble simulations to contrast impacts of pristine and polluted CCN populations. The first set included CCN distributions centered at relatively large CCN (50 nm) and total concentrations of either 100 mg−1 in the pristine case or 1000 mg−1 for the polluted case.1 The second set included an additional CCN distribution centered at 10 nm and total concentration of either 500 or 5000 mg−1 for pristine and polluted simulations, respectively. The surface precipitation difference between pristine and polluted in the two sets changed only slightly, but the anvil cloud fraction difference was significantly reduced. The differences between the two sets in GM16 suggest that small CCN can have a noticeable impact on the deep convective cloud field. Simulations reported in this paper aim at providing detailed quantification of the impact of small CCN.
The paper is organized as follows. The next section reviews the two postulated invigoration mechanisms: above the freezing level as suggested by Rosenfeld et al. (2008) and below the freezing level as suggested by Fan et al. (2018). Section 3 presents the model and modeling setup that follow GM16. Section 4 discusses simulation results. Results of additional simulations that consider the impact of the prescribed large-scale flow are presented in section 5. Section 6 discusses recent lightning observations (Thornton et al. 2017) and model simulations (Blossey et al. 2018) that seem to support the invigoration hypothesis. Using additional analysis of the simulation in Blossey et al. we suggest a different interpretation of the lightning observations. A brief summary in section 7 concludes the paper.
2. Review of postulated invigoration mechanisms
a. Invigoration above the freezing level
The purpose of this section is to show that the postulated convective invigoration of Rosenfeld et al. (2008) hinges on the off-loading of the condensate transported across the melting level in polluted environments. This because the latent heating due to freezing only balances the weight of the liquid water carried across the melting level. To show that, we consider combined effects of the condensed water weight and the latent heat of freezing on the convective updraft buoyancy. The buoyancy is defined by the density potential temperature Θd (its deviation from the environment to be exact) that is approximated as
where Θ is the potential temperature, qυ and qc are the water vapor and condensate mixing ratios, and ε = Rυ/Rd − 1 ≈ 0.6 (Rυ and Rd are gas constants for the water vapor and dry air, respectively). If δqc is the liquid water mixing ratio change between pristine and polluted environments, its impact on the density potential temperature for the case of latent heating versus condensate loading can be approximated as
where the first term represents the density potential temperature change due to latent heating and the second term represents the condensate loading change. The change in Θ from freezing of δqc is approximately given by δΘ ~ Lf/cpδqc (Lf = 3 × 105 J kg is the latent heat of freezing and cp = 103 J kg−1 K−1), that is, about 3 × 102 × δqc. The condensate loading terms is also about 3 × 102 × δqc. This shows that the impact on buoyancy of additional liquid water crossing the 0°C level and freezing approximately balances the weight of carrying the condensate.
The buoyancy above the freezing level can increase when the additional frozen condensate is converted to precipitation and falls out. A similar off-loading mechanism was shown to increase the depth of warm ice-free shallow cumuli (see Fig. 1 in Grabowski et al. 2015). It may potentially increase buoyancy above the freezing level in the polluted case as well. However, the off-loading mechanism was shown in GM16 to insignificantly affect buoyancy above the freezing level. This will be shown again in this paper.
b. Invigoration below the freezing level
As shown in GM16, the difference in the supersaturation between pristine and polluted conditions results in a noticeable difference in the buoyancy below the freezing level. This has been theoretically shown in Grabowski and Jarecka (2015) for shallow convection and extended to deep convection in Grabowski and Morrison (2017). Thus, the buoyancy difference below the freezing level can provide a small increase of convective updraft strength. However, the interpretation provided in Fan et al. (2018) (and in some earlier publications; e.g., Igel et al. 2015; Chen et al. 2017) is incorrect. In reference to the presence ultrafine aerosol particles (UAP), the caption to Fig. 1 in Fan et al. (2018) says, “With added UAP, an additional number of cloud droplets are nucleated above cloud base, which lowers supersaturation drastically by enhanced condensation, releasing additional latent heat at low and middle levels, thus intensifying convection.” As shown below, it is false to assume that increasing droplet concentration leads to the increased condensation.
The condensation rate Cd can be expressed as Cd ~ N dm/dt, where N is the droplet concentration and dm/dt is the growth rate of a single droplet mean mass m. Applying the condensational growth equation allows the condensation rate to be approximated as Cd ~ NrS, where r is the droplet radius and S is the supersaturation. The equation describing evolution of the supersaturation inside an updraft can be derived from the temperature and water vapor equations and the result is (Squires 1952)
where A is the temperature-dependent coefficient [e.g., Eq. (3) in Politovich and Cooper 1988], w is the updraft strength and τ is the phase relaxation time. The phase relaxation time depends on the droplet concentration and radius as τ ~ 1/(Nr). It is on the order of 1 s for typical droplet concentrations and radii [see, for instance, Table 1 in Politovich and Cooper (1988) or Table 1 in Grabowski and Wang (2013)]. The key point is that supersaturation will rapidly, within a few seconds, approach the quasi-equilibrium Seq as given by dS/dt = 0 or Seq = Awτ. Figure 15 in Grabowski and Morrison (2017) shows that the Seq provides a good approximation for the actual supersaturation inside convective updrafts below the freezing level in simulations of deep convection. It follows that as long as the S = Seq, the condensation rate depends on the vertical velocity alone, that is, Cd ~ Seq/τ = Aw, and it is independent of the droplet concentration and radius. In other words, arguing that by increasing droplet concentration one increases condensation rate is simply incorrect as long as the vertical velocity does not change. The correct argument is that polluted updrafts are stronger because they feature lower supersaturations and higher buoyancy (Grabowski and Jarecka 2015; Grabowski and Morrison 2017). Stronger updrafts do lead to larger condensation rates and thus greater latent heating. This will be illustrated by the simulations discussed in this paper.
3. The model and model simulations
Almost all details of the simulations presented in this paper follow those in GM16. The model is the serial (nonparallelized) version of the Eulerian–Lagrangian (EULAG) model (http://www.mmm.ucar.edu/eulag/), referred to as babyEULAG. The microphysics parameterization is the double-moment warm-rain and ice microphysics scheme of Morrison and Grabowski (2007, 2008a,b). The scheme includes in-cloud prediction of the supersaturation field, Kohler-theory-based representation of CCN activation, heterogeneous and homogeneous ice initiation, and representation of the ice field growth by the diffusion of water vapor and by accretion of the supercooled liquid water; see the detailed discussion in section 2a of GM16. The modeling setup, also the same as in GM16, considers daytime convective development over a warm-season continent. It features early morning formation of a convective boundary layer that is followed by midmorning development of shallow convection, subsequent transition from shallow to deep convection, and only convective anvils that are present in the final couple hours. The 12-h simulations (i.e., from 0730 to 1930 local time) apply evolving surface fluxes as in Grabowski et al. (2006). The horizontal domain of 50 km × 50 km is covered with a uniform 400-m grid. In the vertical, the domain extends up to 24 km, applying 81 levels with a stretched grid. The vertical grid length is around 100 m near the surface, with about a dozen levels below 1.5 km. The vertical grid length increases to about 300 and 400 m at 5 and 15 km, respectively. The model time step is 4 s.
We contrast convection developing within a pristine environment (PRIS hereinafter) with convective development when an additional mode of small CCN is added (ADCN hereafter). PRIS assumes total CCN concentration of 100 mg−1 with a single mode distribution centered at 100-nm mean radius; this is an increase from 50 nm in GM16. As in GM16, the geometric standard deviation of the distribution and the soluble fraction are taken as 2.0 and 0.7, respectively. In ADCN, the PRIS CCN are supplemented with a second mode having 500-mg−1 total concentration and 20-nm mean radius (it was 10 nm in GM16’s simulations presented in section 4 therein). Ice initiation is exactly the same in PRIS and ADCN and includes deposition/condensation–freezing nucleation (Meyers et al. 1992), cloud droplet and raindrop heterogeneous freezing (Bigg 1953), contact freezing as in Morrison and Pinto (2005), and homogeneous freezing of all liquid water at −40°C; see section 2a in GM16 for more details.
A key aspect of the modeling setup, as in GM16, is the piggybacking methodology that enables the impact of aerosols to be separated from different flow realizations. With piggybacking, the simulations are referred to as D-PRIS/P-ADCN and D-ADCN/P-PRIS with D and P designating the driver and piggybacker thermodynamic sets. Small seven-member ensembles are conducted for D-PRIS/P-ADCN and for D-ADCN/P-PRIS. Additional sensitivity simulations to be introduced later will be also discussed.
As in GM16, snapshots of model fields are saved every 6 min of the simulation time. Surface precipitation is saved every 3 min as the average over all time steps from the preceding 3-min period. These datasets are subsequently used in the analysis of model results.
Overall, simulation results are consistent with those discussed in GM16. Below we discuss specific results.
a. Cloud macrophysical and microphysical properties
Figure 1 shows evolution of the mean cloud fraction profiles in a format similar to Fig. 1 and part of Fig. 14 in GM16. The figure documents daytime convective evolution applying the Grabowski et al. (2006) setup. As discussed in GM16, no significant differences exist until the second half of the day when convection begins to dissipate and only upper-tropospheric anvils are present. As discussed in GM16, slower sedimentation of smaller ice particles means they stay longer aloft. Differences at hour 12 are large. Figure 2, in the same format as Fig. 14 in GM16, shows statistics of cloud droplet concentrations in the lower troposphere (in the layer between 2 and 4 km) versus ice crystal concentrations in the upper troposphere (in an approximately 3-km-deep layer below the cloud top). The data come from all time levels between hours 6 and 8 (i.e., when deep convection is at its maximum strength) and for cloudy columns with a cloud base below 2 km, cloud top above 11 km, and column-maximum vertical velocity larger than 5 m s−1. Figure 2 clearly shows that adding small CCN leads to an increase of both droplet and ice concentrations.
Figure 3 compares point-by-point various quantities inside convective updrafts at 3-km height and hour 6 of the simulations. The horizontal axis for all three panels is the driver supersaturation for PRIS (ADCN) in the left (right) panels.2 The vertical axes are the piggybacker supersaturation (bottom panels), the driver vertical velocity (middle panels), and the concentration of activated CCN (the latter provides an approximate measure of the droplet concentration; see section 2c of Morrison and Grabowski 2008a). Bottom panels show that supersaturations for ADCN are smaller regardless whether it drives (bottom-right panel) or piggybacks (bottom-left panel). The concentration of activated CCN increases with the driver supersaturation. This is because driver supersaturations are larger for larger vertical velocities as shown by the middle panel. One has to keep in mind, however, that smaller CCN concentrations require higher supersaturations to activate and this explains the data in the upper panels, that is, higher concentrations of activated CCN for higher supersaturations. However, the link between the local supersaturation and the local concentration of activated CCN is indirect as activation typically happens below the 3-km level (e.g., near the cloud base). For PRIS, all available CCN are activated at 3 km for updrafts larger than a couple meters per second. In contrast, only a fraction of available CCN is activated for ADCN unless the vertical velocities exceed several meters per second. Vertical velocities seem on average larger when ADCN drives, an aspect to be quantified later in the paper. Although results shown in Fig. 3 come from only two simulations, they are representative of all ensemble members.
To further document supersaturations inside the simulated clouds, Fig. 4 compares statistics of the supersaturation from the PRIS and ADCN drivers for updrafts stronger than 1 m s−1 and with the total condensate larger than 1 g kg−1 for all time levels during hour 6 of the simulations and for all seven ensemble members. The figure shows that PRIS simulations feature a larger supersaturation range and larger mean values. Smaller mean supersaturations for ADCN agree with higher droplet concentrations as shown in Fig. 2. However, one needs to keep in mind that activation of additional small CCN requires higher supersaturations; grid volumes where this happens are excluded from the figure because of the total condensate threshold. It follows that there exists a balance between higher supersaturations required to activate additional CCN (excluded from Fig. 4) and smaller supersaturations where CCN has already activated, leading to increased droplet concentrations and thus more efficient condensational growth (i.e., smaller supersaturations). The 10th–90th-percentile range increases with height below the freezing level, then varies little within the mixed-phase zone (i.e., between 0° and about −40°C), and increases again near the homogeneous freezing level around −40°C. The increase below the freezing level is due the increase of the updraft strength with height (not shown), whereas the increase in the upper troposphere is likely due to the rapid production of ice near the homogeneous freezing level that drives down supersaturation from subsequent ice growth. The mean supersaturations are around 1% in the mixed-phase zone where water and ice coexist. The mean supersaturation difference between PRIS and ADCN increases with height below the freezing level and decreases aloft.
b. Convective dynamics
Differences in the supersaturation affect the density potential temperature and thus the buoyancy. As shown in Grabowski and Jarecka (2015) and Grabowski and Morrison (2017), smaller supersaturations result in larger parcel buoyancies. In the lower troposphere, the difference in the density potential temperature can be as large as 0.1 K for a 1% supersaturation difference; see Fig. 2 in Grabowski and Morrison (2017). In the upper troposphere, the difference is typically much smaller than 0.1 K (Fig. 2 in Grabowski and Morrison 2017). Figure 5 compares point-by-point buoyancies between the driver and piggybacker for PRIS and ADCN at the height of 3 km (temperature around 283 K) and at 7 km (temperature around 261 K).3 The driver and piggybacker buoyancies are similar at 7 km (top panels in Fig. 5, in agreement with results discussed in GM16), but there is a clear difference at 3-km height (bottom panels in Fig. 5), with ADCN featuring buoyancies larger by up to a few hundredths of 1 m s−2 regardless of whether ADCN drives (bottom-right panel) or piggybacks (bottom-left panel). Larger buoyancies for ADCN at 3 km come from lower supersaturations, whereas the small supersaturation differences at 7 km have a small impact on the buoyancy. In a nutshell, we argue that the differences shown in Fig. 5 follow directly from the supersaturation differences documented in Fig. 4.
Figure 6 compares updraft statistics at 3- and 7-km height between PRIS and ADCN ensembles. Only points with vertical velocity larger than 1 m s−1 and total condensate larger than 1 g kg−1 are included in the statistics. There is a clear shift toward stronger updrafts at 3 km for ADCN as the mean and median velocities are larger by about 0.5 m s−1, and the 90th-percentile velocities differ by around 1 m s−1. The velocities are stronger at 7 km in agreement with the increase of the buoyancy integral with height (i.e., the cumulative CAPE; see Fig. 2 in Grabowski and Prein 2019), but there is virtually no difference between PRIS and ADCN except for hour 11 when deep convection is already dissipating. The similarity of PRIS and ADCN distributions at 7 km arguably reflects the impact of entrainment as the adiabatic maximum vertical velocity (i.e., the square root of twice the CAPE) is in excess of 60 m s−1 for approximately 2000 J kg−1 CAPE around hour 6 (see Fig. 3 in Grabowski and Prein 2019). Including the perturbation pressure forcing (Morrison 2016a,b) reduces the maximum vertical velocity significantly, but the maximum velocities shown in the upper panel of Fig. 6 (i.e., below ~10 m s−1) are still much smaller. Hence, the impact of entrainment is the likely reason for similar vertical velocity distributions at 7 km.
The ~0.01 m s−2 buoyancy difference at 3-km height, as shown in the lower panels in Fig. 5, can lead to noticeable differences in the vertical velocity. If B is the buoyancy and w is the vertical velocity, then the increase of the squared vertical velocity with height in steady-state conditions is given by dw2/dz = 2B, neglecting perturbation pressure forcing and momentum entrainment. For buoyancies around 0.05 and 0.06 m s−2 in PRIS and ADCN, respectively, from Fig. 5 (regardless of which drives and which piggybacks), integrating this expression over a height difference Δz of 1 km gives a difference in w of ~1 m s−1. This is somewhat larger than the actual increase of mean and median w of ~0.5 m s−1 between PRIS and ADCN seen in Fig. 6. However, this estimate neglects perturbation pressure forcing that can further reduce the w difference by a factor of 2 or more (Morrison 2016a,b). Thus, the w difference of about 0.5 m s−1 at 3-km height is roughly in line with the buoyancy differences between PRIS and ADCN.
c. Surface precipitation
Evolution of the domain-averaged surface rain accumulations is shown in Fig. 7. The top two panels show evolution of surface rain accumulations using solid lines for drivers and dashed lines for piggybackers. Driver–piggybacker (D–P) differences are small as seen in the third panel from the top that shows evolution of the D–P differences for all ensemble members. The spread in D–P is small and even a single pair of the piggybacking simulations would be sufficient for assessing the impact on the mean surface rain accumulation. The D-ADCN minus P-PRIS difference is significantly larger than D-PRIS minus P-ADCN. In GM16, the D–P difference also depended on which one drives (see Fig. 6 therein, the bottom-left panel), but the contrast was not as large as in the current simulations. The bottom panel shows evolution of the ensemble mean difference between drivers, with thick vertical lines representing standard deviation within the ensemble. On average, ADCN has slightly more rainfall, about 0.15 mm or about 5% of the mean accumulation. The standard deviation among ensemble members is still large, about two-thirds of the mean difference.
A possible explanation for the contrast between the D–P difference depending on which one drives the simulations comes from spatial distribution patterns of the rain accumulation. These are shown in Fig. 8. The left panels show accumulations from the drivers that feature elongated structures arguably resulting from advection of precipitating system by the mean wind. The right panels show patterns of the D–P difference. The patterns change between top- and bottom-right panels: in the top panel, negative values are typically located to the southeast of the positive values, and the pattern reverses in the bottom-right panel (i.e., negative values are typically to the northwest of the positive values). A possible reason may come from the shear in the prescribed mean flow that is shown in Fig. 9. The figure shows that the mean flow is from the northwest in the lower troposphere (leading to the pattern shown in the left panels of Fig. 8), and then it changes to southerly and easterly flow in the upper troposphere. Since driving with ADCN gives typically stronger updrafts below 3 km as shown above, the vertical versus horizontal advection of cloud and precipitation elements is likely different compared to the case when PRIS drives. This arguably leads to the pattern evident in the right panels of Fig. 8.
Finally, probability distributions of rain rates in the drivers and piggybackers from the two simulation sets for the period with deep convection (hours 5 and 6) were also examined. The distributions turned out to be similar between PRIS and ADCN sets, with slightly smaller values for precipitation rates around 20 mm h−1 and larger values for precipitation rates over 60 mm h−1 for ADCN (not shown). This is in contrast to the results shown in Chen et al. (2017, Fig. 2 therein) that documents a significant difference between pristine and polluted conditions.
5. Sensitivity simulations
There are two factors that motivate the sensitivity simulations discussed in this section. First, as mentioned above, the environmental wind shear could be responsible for the asymmetry between D–P difference in the mean accumulation as show in Fig. 7. Moreover, shear allows more precipitation to fall outside clouds and evaporate before reaching the ground (e.g., Ferrier et al. 1996). Thus, the impact of additional CCN for the case without shear may be different than discussed in the previous section. This aspect has not been considered in simulations discussed in GM16.
Following the above arguments, two additional three-ensemble member sets of piggybacking simulations were run. The first set excludes the impact of shear above 1 km by assuming the mean winds aloft are the same as at 1 km. Hence, these simulations include the low-level shear and feature uniform northwesterly flow above 1 km. In general, results of these simulations are similar to those shown so far. Evolution of the cloud fraction profiles is similar to that shown in Fig. 1, although with smaller differences between PRIS and ADCN in the final 3 h of the simulations. Surface mean accumulations, including D–P differences, are also similar to those shown in Fig. 7. Other statistics analyzed for the original simulations are similar as well. Because of these similarities, the results from these sensitivity simulations are not shown.
The second three-ensemble member set of additional simulations assumes no mean wind across the entire atmosphere as in Wu et al. (2009) and Böing et al. (2012). These simulations are referred to as PRIS-NW and ADCN-NW (NW for no wind). Most results, such as the supersaturation statistics (Fig. 4 for the original simulations), buoyancy at 3- and 7-km height (Fig. 5) and updraft statistics (Fig. 6) are similar to those discussed above and thus are not shown. Results that highlight differences are shown in Figs. 10, 11 and 12. Figure 10 shows evolution of the cloud fraction profiles. A clear difference from Fig. 1 is a smaller spread among ensemble members and reduced D–P difference of upper-tropospheric cloud fractions in the second half of the simulations. The reduced D–P difference in Fig. 10 arguably comes from the absence of the upper-tropospheric large-scale shear that affects spread of the anvils and the way ice particle sedimentation impacts the anvil life cycle. The former is consistent with the observational study in Koren et al. (2010a). One can also notice an earlier transition from shallow to deep convection in the no-wind simulations as seen by comparing profiles for hour 4 in both simulations. The large-scale wind has a small impact on the surface precipitation statistics as seen in Fig. 11 that should be compared with Fig. 7 showing results of the original simulations. Without the large-scale wind (and shear), differences between the drivers and the standard deviation between the drivers are reduced (the former in agreement with the cloud fraction profiles and the latter despite the smaller ensemble). The D–P differences are larger than in the original set, but the asymmetry in D–P remains.
The most interesting is the impact of the mean wind on surface rain accumulation patterns shown in Fig. 12. For the total accumulation (left panels), the elongated structures in PRIS and ADCN evident in Fig. 8 are replaced by more numerous, smaller, and more rounded structures in Fig. 12 [see also Figs. 7 and 8 in Böing et al. (2012)]. This arguably comes from the impact of mean wind that leads to the advection of precipitating systems across the computational domain. Advection of systems allows them to access boundary layer air that has been affected by the surface fluxes (i.e., air with the increased equivalent potential temperature), thus extending their lifetime. The spatial pattern of the D–P difference (right panels in Fig. 12) shows small-scale features, mostly positive (red color) versus negative (blue color), in the top versus bottom panels.
Comparison of the surface accumulation patterns in Figs. 8 and 12 suggests that systems in PRIS and ADCN simulations not only live longer, but they also appear to be larger. Figure 13 shows maps of the total condensate and precipitation path (i.e., vertical integral of the sum of all liquid and ice mixing ratios) for three different times during the strongest convection phase for single members of PRIS and PRIS-NW ensembles. Colors represent different times: red for hour 6.5, green for 12 min later, and blue for 24 min later. The plots, representative of simulations from the other two ensemble members, clearly show that systems simulated in the no-wind case are smaller and seem to represent single convective cells progressing through the classical convective cell life cycle (Byers and Braham 1948). In contrast, convective systems in the PRIS simulations are larger and they seem to represent multicellular storms (see the extreme example near the lower-right corner in the right panel). Because convective systems in the simulations with no shear above 1 km briefly mentioned above are similar to those in PRIS (e.g., the total surface rain accumulations resemble those shown in Fig. 8; not shown), one can argue that it is the low-level shear that leads to the formation of larger, multicellular storms. This is likely because of the interactions of convective cold pools with the environmental low-level shear that affect new cell formation (e.g., Liu and Moncrieff 1996, and references therein). However, the shear is not strong enough nor deep enough to allow formation of a squall line as in the classical RKW theory (Rotunno et al. 1988).
The key conclusion from the sensitivity simulations is that the impact of additional CCN is similar regardless of the mesoscale convective organization.
6. Recent lightning observations and convective invigoration
In contrast to GM16 and results presented herein, the recent study of Thornton et al. (2017) seems to support the hypothesis of convective invigoration in polluted environments. They analyzed 12 years of satellite lightning observations and document up to a twofold increase of lightning density over major oceanic shipping lanes when compared to the shipping lane environment. Thornton et al. argue that CCN particles resulting from ship exhaust are added to the otherwise clean maritime environment, invigorate convection, and modify ice processes above the shipping lanes, leading to enhanced lightning. To support the Thornton et al. hypothesis, Blossey et al. (2018) applied a cloud resolving model in an attempt to simulate the impact of shipping lane CCN enhancement on deep convection and to better understand the associated processes. Their simulations applied a convective-radiative quasi-equilibrium setup with a prescribed uniform sea surface temperature of 29°C and used a radiation transfer model to represent radiative cooling of the atmosphere. A doubly periodic horizontal domain of about 400 km × 400 km was used with horizontal grid length of 1 km. A narrow strip of continuous emission of additional CNN mimicked the shipping lane. A double-moment bulk microphysics scheme with saturation adjustment was used.
The simulations of Blossey et al. (2018) indeed show convective “invigoration” over the shipping lane (we put invigoration in quotes to stress a different meaning as we discuss below). There are larger spatially averaged values of liquid water (below the freezing level) and graupel (above the freezing level) mixing ratios and larger upper-tropospheric reflectivities over the lane when compared to the lane environment; see Fig. 3 in Blossey et al. (2018). However, because the microphysics scheme applies saturation adjustment, the only significant impact on the dynamics discussed in GM16 and in the current study is eliminated. So where does the “invigoration” come from? Fig. 14 shows results from additional analysis of the Blossey et al. simulations and suggests a possible explanation. The figure compares profiles of the area-averaged vertical velocity and the mean in-core convective updraft velocity for areas outside and inside the shipping lane (the core is defined as a positively buoyant updraft with velocity larger than 1 m s−1). Although the mean ascent is stronger over the shipping lane, the convective updraft strength is practically unchanged. The larger area-averaged vertical velocity and the same updraft strength imply that there are simply more convective drafts without changing their strength. Hence, the “invigoration” in the Blossey et al. (2018) simulations represents more convection, not stronger convection. More convection likely comes from a weak mesoscale circulation that develops between the shipping lane and its environment, similarly to simulations discussed in Morrison and Grabowski (2013). The circulation is likely initiated by the difference in radiative cooling between the lane and its environment. In the quasi-steady state, the circulation is maintained by the differences in the latent heating resulting from more convection above the shipping lane as well as by the differences in the radiative heating (P. Blossey 2020, personal communication). The same argument can be used for the observations reported in Thornton et al. (2017) with a caveat that latent heating below the freezing level due to supersaturation differences as discussed herein and in GM16 could also play role in driving a mean circulation between the shipping lane and its environment.
7. Summary and conclusions
The simulations discussed in this paper clearly show a small impact of additional small CCN on convective dynamics and a significant microphysical effect on upper-tropospheric anvils. The impact on the dynamics is limited to below the freezing level and comes from buoyancy differences associated with differences in the in-cloud supersaturations. One can argue that this in principle agrees with the argument put forward in Fan et al. (2018). However, we feel there are significant differences. First, the explanation suggested in Fan et al. is incorrect as shown in section 2. The invigoration does not come from the enhanced condensation resulting directly from higher droplet concentration. As long as the quasi-equilibrium supersaturation is a valid approximation (and there are good reasons to believe so as discussed in section 2), the condensation rate depends only on the updraft speed. The correct argument is that reducing the quasi-equilibrium supersaturation due to higher droplet concentration leads to the increased cloud buoyancy, stronger updraft, and thus more condensation. Second, observed and simulated updraft speed increases presented in Fan et al. (up to a factor of 3; Figs. 2 and 3 therein) are much larger (unrealistic in our view) than simulated here. As mentioned in the introduction, we believe key factors important to understand the difference between our results and results in Fan et al. (2018) concern the variable environment near Manaus (see Santos et al. 2014) that makes extraction of the CCN impact difficult and an insufficient accuracy of observations that preclude elimination of meteorological impacts on convection (e.g., mesoscale circulations or river breezes) in addition to differences in CCN. Finally, Fan et al. (2018) also argue about the invigoration above the freezing level (i.e., as in Rosenfeld et al. 2008) that is inconsistent with the results here and in GM16.
The microphysical impact on the anvils (larger for the cases with environmental wind shear) comes from higher droplet concentrations leading to higher ice concentrations aloft. With approximately the same mass of condensed water transported into the upper troposphere, higher ice concentrations imply smaller ice particles that sediment with slower fall velocities and extend the anvil life cycle. The presence (in the original simulations) or absence (in the no-wind simulations) of the mesoscale organization has a small impact on the dynamics below the freezing level and a more significant impact on the upper-tropospheric anvils. The latter agrees with satellite observations reported in Koren et al. (2010a), although their explanation involves convection invigoration that is incorrect in our view. Finally, we suggest in section 6 that recent observations of enhanced lightning over oceanic shipping lines argued to result from convection invigoration (Thornton et al. 2017) may simply represent the impact of a persistent mesoscale circulation that develops between the line and its immediate environment. As a result, there is simply more convection over the shipping line, without significant invigoration of individual convective drafts.
In summary, the invigoration of deep convection in polluted environments, either resulting from increased total CCN concentrations as suggested in Rosenfeld et al. (2008) or from addition of small CCN to the pristine environment as in Fan et al. (2018), is not supported by theoretical arguments and cloud simulations presented here.
The authors acknowledge partial supported from the U.S. DOE ASR Grants DE-SC0016476, DE-SC0020104, and DE-SC0020118. Peter Blossey provided graphs used in Fig. 13 together with comments on the early draft. NCAR is sponsored by the National Science Foundation.
Supersaturation refers to the supersaturation with respect to liquid water saturation throughout the paper.