This study is presented as an indication of the uncertainty that can arise from the choice of microphysics scheme alone, and thus we restrict our comparison to two double-moment bulk microphysics schemes rather than diversifying into a comparison of bin schemes against bulk schemes. The literature surveyed in Sect.  indicates the wide range of differences that may be expected when comparing bulk against bin schemes. A significant body of work has shown that two-moment bulk microphysics schemes generally represent cloud and precipitation characteristics more realistically than single-moment schemes most recently, and thus our study is restricted to the comparison of two five-class, double-moment schemes commonly used in WRF and shown by to perform well against satellite observations of cloud in North America: that described by  and (hereafter Morrison, or abbreviated to MORR), and that described by (hereafter Thompson, or THOM). Both schemes are two-moment in rain and ice (prognostic mass and number), while the Morrison scheme is also two-moment in snow and graupel. Both are single-moment in cloud water: mass is the only prognostic liquid cloud variable, and CDNC is prescribed at a given value. Following the method used in many previous studies including that of , we prescribe CDNC values (in this study, at 100, 250 and 2500 cm-3) as a proxy for CCN varying under conditions ranging from clean to highly polluted. The list of microphysics configurations tested and the abbreviations used to describe them are summarised in Table .

The real-data Congo simulations use a model domain covering a 2100×2100km region over the Congo basin (Fig. ) chosen due to the high frequency of isolated deep convective systems occurring in the region and also due to the presence of strong sources of biomass burning aerosol. The model initial and boundary conditions were generated from ERA-Interim reanalysis starting at 00:00 UTC on 1 August 2007. The simulation start date was chosen to coincide with the onset of the seasonal peak in precipitation and the simulation was integrated for 10 days (with a time step of 12 s) in order to identify the nature of the convection and its response to CDNC perturbations over timescales greater than that of the life cycle of any individual convective system. We use a horizontal grid length of 4 km and 30 vertical levels with the standard WRF stretched vertical grid. This gives a vertical grid spacing of about 100 m in the lower levels with grid spacing increasing towards the upper levels. Although 30 vertical levels may seem relatively coarse, it has been shown in a previous study to be sufficient to reproduce observed cloud morphology and resolve the vertical structure of aerosol and precipitation and their interactions in this region . Longwave and shortwave radiation in the simulations are parameterised by the RRTM and Goddard schemes, respectively. Other physics parameterisations (other than the microphysics schemes previously discussed) are the MM5 Monin–Obukhov similarity surface layer scheme available in WRF which uses stability functions and surface fluxes from, the NOAH land surface model and the YSU boundary layer scheme , also shown by to perform well.

The idealised supercell set-up follows the standard 3-D idealised supercell case available as part of the WRF modelling system. Boundary conditions are open on all lateral boundaries, and the model top and surface are free-slip. For consistency with the Congo basin simulations, we use a horizontal grid length of 4 km. The model domain is 1600×1600km in the horizontal and, for consistency with the Congo simulations, also uses 30 vertical levels with a model lid at 20 km. A Rayleigh damper with a damping coefficient of 0.003 s-1 is applied in the top 5 km of the model to prevent spurious wave reflection off the model top. Following the set-up commonly used in idealised supercell studies e.g., surface energy fluxes, surface drag, Coriolis acceleration and radiative transfer are neglected for simplicity, and the subgrid-scale horizontal and vertical mixing is calculated with a prognostic turbulent kinetic energy scheme . The model is initialised as in the idealised quarter-circle supercell test case available in WRF using the analytic sounding of and the quarter-circle supercell hodograph of with the shear extended to a height of 7 km. Convection is triggered using a thermal perturbation in the centre of the domain with a maximum perturbation potential temperature of 3 K centred at a height of 1.5 km and with horizontal and vertical radii of 10 and 1.5 km, respectively. All simulations are integrated for 2 h with a time step of 12 s (the same time step used in the Congo simulations).

The warm-rain shallow cumulus set-up deviates from the other simulations in that it follows the LES intercomparison guidelines for the Precipitating Shallow Cumulus Case 1 of the Rain in Shallow Cumulus Over the Ocean RICO; project and uses the RICO WRF LES package provided by . The model domain is 12.8×12.8×4km with a horizontal grid spacing of 100 m and uses 100 vertical levels, implying a vertical grid spacing of about 40 m. The lateral boundary conditions are doubly periodic. As in the idealised supercell simulations, surface energy fluxes, surface drag, Coriolis acceleration and radiative transfer are neglected for simplicity, and the subgrid-scale horizontal and vertical mixing is calculated with a prognostic TKE scheme. The surface conditions, wind and thermodynamic profiles, large-scale forcings and large-scale radiation, geostrophic wind, initial perturbations and translation velocity are prescribed following the RICO case guidelines . For consistency, we prescribe cloud droplet number concentrations at 100, 250 and 2500 cm-3, following the other simulations in our study, instead of the 70 cm-3 suggested for the standard RICO case. However, we also perform an extra simulation at 50 cm-3. The simulations are integrated for 24 h with a time step of 1 s.

Maps of simulated outgoing longwave radiation (OLR) and surface precipitation at 07:00 UTC on 7 August 2007 (7 days into the simulation) indicate that the cloud morphological and precipitation differences for different microphysics schemes are much greater than the cloud and precipitation response within each scheme to different CDNC values (Fig. ). In the CONGO-MORR simulations, low OLR values (indicating cold, high cloud) are distributed across the domain. Precipitation at this time occurs only in cloud north of 3∘ S, but there is a large band of non-precipitating cold cloud across the south of the domain. There is little discernable response of the morphology of the OLR and precipitation in the CONGO-MORR simulations to different CDNC values (Fig. a–c). In comparison, cold cloud in the CONGO-THOM simulations occurs mostly north of 3∘ S (Fig. d–f). Less cloud forms in CONGO-THOM compared to CONGO-MORR, and the cloud generally has greater OLR values than that in CONGO-MORR. Some non-precipitating cloud occurs south of 3∘ S in the CONGO-THOM simulations, but the band is significantly weaker and warmer than in CONGO-MORR. The differences at this snapshot are representative of differences that persist throughout the simulation. Frequency distributions of OLR over the entire 10-day simulation period show that CONGO-MORR has a much higher frequency of occurrence of colder, higher cloud (values of about 120 Wm2) than CONGO-THOM (which increases in frequency slightly with increased CDNC), while CONGO-THOM has a much higher frequency of occurrence of warmer cloud (values of about 270 Wm2) than CONGO-MORR (Fig. a). When compared to observations of OLR from the Geostationary Earth Radiation Budget GERB; over the same region and period, CONGO-THOM represents warm cloud more consistently with GERB than CONGO-MORR, despite overpredicting colder cloud somewhat, while CONGO-MORR overpredicts higher cloud and underpredicts warm cloud compared to the observations (Fig. a). However, despite a poorer prediction of cloud radiative properties, CONGO-MORR predicts surface precipitation better than CONGO-THOM when compared to observations from the Tropical Rainfall Measuring Mission TRMM; merged product. Both schemes significantly overpredict surface precipitation compared to observations from the TRMM 3B42 product (although the spatial patterns of precipitation are reasonably similar); however, total accumulated surface precipitation over the 10-day simulation period is much greater in CONGO-THOM than CONGO-MORR (Fig. ). Further differences are seen when the distributions of precipitation rates are compared, with CONGO-THOM overpredicting and CONGO-MORR underpredicting the occurrence of low precipitation rates compared to TRMM, CONGO-MORR overpredicting and CONGO-THOM underpredicting moderate rates, and CONGO-THOM overpredicting the frequency of occurrence of very high precipitation rates (Fig. b). That CONGO-MORR overpredicts the frequency of moderate rain rates and CONGO-THOM overpredicts the frequency of very high rain rates likely explains why both schemes overpredict total accumulated surface rain compared to the observations. Additionally, the overprediction of the frequency of very high precipitation rates by CONGO-THOM is likely the reason that the total accumulated surface precipitation is much greater in this scheme than in CONGO-MORR (Fig. a and b).

Further to the significant difference between the two schemes in their reproduction of cold cloud and precipitation rates, the updraught dynamics respond very differently to aerosol loading. Joint histograms of cloud top height in the convective updraughts and the radius of the updraughts show that the most significant dynamical difference between the simulations comes from the choice of microphysics scheme: the Morrison scheme has a tendency towards higher frequencies of wider updraught radii with higher cloud tops than the Thompson scheme (Fig. S1 in the Supplement). Under increased values of CDNC, convection in the CONGO-MORR simulation shifts towards wider cores and higher core tops for midsized cores (radius 11 to 22 km), whilst there is a reduction in the frequency of smaller cores of all core top heights (Fig. c). Conversely, convection under polluted conditions in the CONGO-THOM simulation shows a reduced frequency of occurrence of the highest updraught cloud tops for all updraught radii under polluted conditions with an increased frequency of occurrence of small updraught radii with lower cloud tops (Fig. d). Therefore, a consistent aerosol response is observed in CONGO-THOM, resulting in smaller and lower convective updraughts (i.e. weakened convection under polluted conditions). Interestingly, both of these effects contradict the findings of , who found an ice-phase response to aerosol in which cloud top heights and anvil ice mixing ratios increase under polluted conditions due to increased freezing of larger numbers of cloud droplets and subsequent higher ice particle concentrations with smaller sizes and reduced fall speeds. However, we note that we consider different values of CDNC/CCN to and that responses may be nonmonotonic . We also consider a different case of convection (indeed, our 10-day Congo simulation covers many convective lifecycles). We note that the response of the convective updraughts to aerosol loading in these two bulk schemes cannot be attributed to saturation adjustment alone the suggested effects of which on updraught invigoration are detailed in because both schemes use this method.

Not only does the simulated cloud and precipitation morphology differ significantly between microphysics schemes irrespective of the CDNC values used in the comparison, zonal-mean vertical sections of the mass mixing ratios of the different hydrometeor classes show significant differences in the hydrometeor classes (due to microphysics) between CONGO-MORR and CONGO-THOM (Fig. ). The most significant difference between the two microphysics schemes is that south of 3∘ S, CONGO-MORR produces a large amount of high ice cloud between 300 and 150 hPa (Fig. a–c). Analysis of these vertical sections at hourly intervals throughout the simulation in conjunction with hourly maps of OLR, as in Fig. , show that this upper-level ice is transported from the convective anvils in the north of the domain to the non-convective region in the south of the domain (not shown). In comparison, CONGO-THOM produces significantly less ice with almost no ice visible at this contour value (Fig. d–f). However, all three CONGO-THOM simulations form a large amount of non-precipitating low-level (950 to 850 hPa) liquid cloud south of 3∘ S. The bands of cloud seen south of 3∘ S in Fig.  are therefore high ice cloud in the CONGO-MORR simulations and low liquid cloud in the CONGO-THOM simulations, illustrating not only a cloud morphological difference between the microphysics schemes but also a significant difference in the simulated hydrometeor classes and in the vertical distribution of hydrometeors. Even in the convective precipitating region in the north of the domain, the simulated hydrometeor classes differ significantly between the microphysics configurations with the CONGO-MORR simulations generating more ice and less liquid cloud (Fig. a–c) and the CONGO-THOM simulations producing less ice and more liquid cloud (Fig. d–f). Rain is confined to the convective region in the north in CONGO-THOM, while in CONGO-MORR it is also present at low levels in the non-convective southern region of the domain which is dominated by liquid cloud in CONGO-THOM. We explain the mechanisms behind these differences later, but here we highlight that it is clear from Fig.  that the differences in the simulated hydrometeors between microphysics schemes are much greater than the differences due to different levels of CDNC.

Because the partitioning of water into liquid and ice phases in the full-physics model configuration appears to depend strongly on the microphysics scheme, vertical sections of reflectivity occurrences derived from model hydrometeor fields passed through the QuickBeam radar simulator are compared against equivalent reflectivity occurrences from the CloudSat 2B-GEOPROF product (Fig. ). The histograms are derived from the reflectivity fields thresholded to include all values greater than -20 dBZ. The largest reflectivity values produced by the model occur in the convective region in the north of the domain where the largest reflectivity values are detected by the satellite radar (Fig. ), which is also in agreement with the TRMM precipitation observations (Fig. ). However, both CONGO-MORR and CONGO-THOM have a large positive bias in reflectivity compared to the observations (Fig. ), which is indicative of limitations in the ability of both bulk microphysics schemes to represent the observed vertical cloud structure in this geographic region over this time period. In general, CONGO-MORR has a much larger positive bias in reflectivity than CONGO-THOM (Fig. ). The CloudSat observations show a small frequency of occurrence of reflectivities detected at altitudes of 10 to 15 km in the south of the domain, which is well represented by CONGO-THOM and indicates the overproduction of ice in CONGO-MORR (Fig. ).

Differences in the simulated hydrometeor classes between the schemes persist throughout the simulation and are illustrated by mean profiles of hydrometeor mass mixing ratios (Fig. ). There is significantly more ice-phase condensate in the CONGO-M250 configuration (Fig. a), whereas the CONGO-T250 profile is dominated by a large amount of liquid cloud mass between the near surface and 750 hPa (Fig. b). The differences in the total cloud water mass between the schemes are very large: at 950 hPa (the altitude with the greatest liquid cloud mass in CONGO-T250; Fig. ), cloud water mass contents are about 140 times greater in CONGO-T250. The liquid cloud mass is always greater in CONGO-T250 than CONGO-M250 (Fig. c) by several orders of magnitude at some levels, but despite this the liquid phase does not appear to drive differences in precipitation between the microphysics schemes: CONGO-M250 has about 4 times more rain mass in the mid-levels and 2 times more rain mass near the surface than CONGO-T250 (Fig. a and b). In the ice phase, CONGO-M250 has only slightly more snow mass than CONGO-T250 but up to 10 times more graupel mass (Fig. a and b), and while ice is a significant hydrometeor at upper levels in CONGO-M250, CONGO-T250 has almost no cloud ice at all (Fig. a and b). We note that the magnitude of the difference due to the choice of scheme is the same when a bin scheme is used (Figs. S2 and S3 in the Supplement).

Mean profiles over all condensed points (i.e. representing the mean values of each hydrometeor type but not accounting for changes in absolute quantities across the model domain) show that CONGO-T250 has consistently more cloud water through the depth of the mean cloud compared to CONGO-M250 (Fig. a and b), while CONGO-M250 produces more rain (Fig. a). That rain production in CONGO-M250 occurs mostly through the depths of the atmosphere where cloud water persists suggests that a significant proportion of the rain may be produced though autoconversion in CONGO-M250, although note that these mean cloud profiles are calculated over the entire domain and therefore incorporate both the deep convective region in the north and the warm-cloud region in the south, as seen in Fig. . Further, the two schemes show differences in the frozen hydrometeors with the mean cloud in CONGO-M250 containing more graupel and less snow than CONGO-T250 (Fig. c). This may be a result of the use of distinct and different definitions of ice-phase hydrometeor categories in the two schemes, which have been shown to cause deficiencies in simulations of observed squall lines .

Not only does the partitioning of ice amongst the hydrometeor classes differ between schemes, the response of the hydrometeors to CDNC perturbations also differs between schemes (Fig. ). First note that the scale of the hydrometeor response to CDNC perturbations in the CONGO-MORR simulations is an order of magnitude smaller than the scale of the response in the CONGO-THOM simulations. Over the entire domain, liquid cloud mass appears insensitive to CDNC perturbations in the CONGO-MORR configuration (Fig. a), although a reduction in mean-cloud liquid cloud mass under polluted conditions (Fig. c) indicates that there must be very few liquid cloud points in the CONGO-MORR simulation compared to other hydrometeor types, notably ice (Fig. a). Very weak decreases in domain-mean near-surface rain mass may be evident under polluted conditions in CONGO-MORR, but this difference is on the order of 10-8 kgkg-1 (Fig. a) A reduction in rain mass under polluted conditions is more evident in the mean rain profile (Fig. c), again indicating how few rainy points exist compared to other hydrometeor types in CONGO-MORR when considering the entire domain (Fig. a). Nearly all of the hydrometeor response in CONGO-MORR occurs in the ice-phase processes: graupel mass decreases significantly under polluted conditions (Fig. a and c), while ice mass increases at upper levels in both a domain mean and ice mean sense (Fig. a and c). In contrast, the hydrometeor response to CDNC perturbations in the CONGO-THOM configuration is an order of magnitude greater than in CONGO-MORR and the dominant hydrometeor response to CDNC perturbations in CONGO-THOM occurs in the liquid phase. Not only does the CONGO-THOM configuration generate significantly more liquid cloud than the CONGO-MORR configuration (Fig. c), but the liquid cloud mass also increases under polluted conditions by an order of magnitude more than any other hydrometeor response (Fig. b and d). Rain mass is relatively insensitive to increased CDNC in CONGO-THOM (Fig. b and d). The significant difference between the response of the two schemes to perturbations in CDNC, with CONGO-MORR producing less liquid cloud and rain under polluted conditions while CONGO-THOM produces more cloud water, indicates significant differences in the cloud processes represented by the two schemes in this meteorological regime.

The results from the real-data Congo basin simulations indicate that the development of the simulated hydrometeor classes and the response of the hydrometeors to CDNC perturbations depend strongly on the choice of microphysics scheme. Although some previous studies have focused on the response of real-data case studies to both microphysics scheme and CDNC response e.g., there is a much larger body of literature that investigates the response of idealised supercell simulations to CDNC (or CCN) perturbations e.g.. We therefore place our study in the wider context of the existing literature by investigating the response of a single isolated idealised supercell under both the MORR and THOM microphysics configurations to the same CDNC perturbations used in our Congo simulations, simultaneously allowing us to explore the case dependence of the deep convective response to aerosol effects.

Figure  shows mean hydrometeor profiles from the idealised supercell model configurations under “moderately polluted” prescribed CDNC values of 250 cm-3. As in the Congo basin case, it is clear that the simulated hydrometeor classes differ significantly between schemes. In contrast to the Congo basin configuration, both the SUPER-MORR and SUPER-THOM configurations show similar behaviour in the liquid phase, producing similar profiles of liquid cloud mass and rain mass in both a domain mean and hydrometeor-class mean sense (Fig. a, d and b, e), and instead the most significant differences occur in the ice phase. Graupel dominates as the frozen precipitating hydrometeor in the SUPER-M250 configuration, amounting to about 4 times the snow and ice masses at their peak amounts (Fig. a and d). In contrast, snow is the dominant frozen precipitating hydrometeor in the SUPER-T250 configuration, amounting to about 1.5 times the graupel mass at peak amounts and virtually no ice present (Fig. b and e). Although there is very little difference between the SUPER-MORR and SUPER-THOM configurations in the liquid phase (except for the SUPER-MORR configuration producing about 2×10-7 kgkg-1 less domain-mean rain mass at the surface than SUPER-THOM; Fig. c), the SUPER-MORR configuration forms significantly more ice, more graupel and less snow than SUPER-THOM (highlighting that the partitioning of ice-phase hydrometeors into categories is very different, by design, in different microphysics schemes). Greater total quantities of frozen hydrometeors are present between 600 and about 150 hPa in SUPER-MORR compared to SUPER-THOM (Fig. c and f). This is a significant difference from the Congo real-data configuration in which the dominant contribution to the difference between the CONGO-MORR and CONGO-THOM configurations came from the liquid cloud (Fig. c).

There is a more significant aerosol impact on hydrometeor mass in the supercell case than in the Congo case for both microphysics schemes, with mean responses over each hydrometeor type an order of magnitude greater in the supercell case (Fig.  compared to Fig. ). Although many past studies have shown that aerosol impacts depend on cloud dynamics and thermodynamics e.g., we note that not only do the individual schemes respond differently to CDNC in different cases of convection (as expected), but the way the schemes differ from each other in their response to CDNC is also significantly different in the supercell case compared to the Congo case. The SUPER-MORR and SUPER-THOM cases differ qualitatively from the CONGO-MORR and CONGO-THOM cases, respectively, both in the altitudes at which the response occurs and the sign of the response of some of the hydrometeors. In the SUPER-MORR configuration, cloud water mass increases under polluted conditions, and rain mass is suppressed at mid-levels (between 600 and 450 hPa) but shows negligible response at the surface (Fig. a and c). In the ice phase, cloud ice increases under polluted conditions in SUPER-MORR, while graupel and snow decrease (Fig. a and c). Similarly, the hydrometeor response of the SUPER-THOM case to CDNC perturbations also differs in sign and in altitude to CONGO-THOM. In SUPER-THOM, cloud water mass increases and rain mass decreases under polluted conditions (Fig. b and d), but unlike SUPER-MORR the decrease in rain is evident at the surface. Graupel mass decreases under polluted conditions in SUPER-THOM, similarly to SUPER-MORR, but occurs over a much larger range of heights (Fig. b and d); this is unlike CONGO-THOM, which shows very little response to polluted conditions (Fig. c). Interestingly, this is in contrast to , who found an increase in graupel mass with increased CDNC in the Thompson scheme. However, their study was of 2-D idealised squall line simulations and considered CDNC values of 100, 500 and 100 drops per cm-3. The dominant domain-mean response to increased CDNC perturbations in SUPER-THOM is an increase in snow mass between 550 and 150 hPa (Fig. b), which likely comes from lofting of an increased mass of cloud water (Fig. d). This is in contrast both to SUPER-MORR in which the dominant hydrometeor response occurred in the ice class (Fig. a), despite an almost equal increase in lofted cloud water (Fig. c), and to CONGO-THOM in which the dominant hydrometeor response occurred in the liquid cloud (Fig. b). That both schemes show an increased lofting of cloud water under polluted conditions (Fig. c and d), but SUPER-MORR responds by generating more cloud ice (Fig. a and c) while SUPER-THOM shows an increase in snow (Fig. b and d), suggests differences in the processes that convert cloud ice to snow. This is explored later in Sect. . We emphasise that our main result shows that the variability due to microphysics scheme dominates any aerosol impacts on microphysics. Results using the WRF-SBM in the idealised supercell case show that aerosol impacts in the bin scheme are of equal magnitude to those in the bulk schemes (Fig. S4 in the Supplement).

To further investigate the importance of the difference in microphysics representations and the difference in their response to CDNC perturbations, Fig.  includes the domain-mean total latent heating (sum of the latent heating from individual microphysical processes) contributions for each of the idealised supercell configurations. It can be seen that the choice of microphysics scheme can result in thermodynamic differences in the supercell system equal in magnitude to those arising from CDNC perturbations: between 500 and 250 hPa, the latent heating rate in the SUPER-M2500 configuration is almost identical to that in the SUPER-T250 configuration (solid red and dashed blue lines, Fig. a). Thus, the magnitude and sign of the difference in the latent heating rate between SUPER-M250 and SUPER-T250 (blue solid and dashed lines, Fig. a) is the same as that between SUPER-M2500 and SUPER-M250 (red and blue solid lines), and likewise the magnitude and sign of the difference in the latent heating rate between SUPER-M2500 and SUPER-T2500 (red solid and dashed lines) is the same as that between SUPER-T2500 and SUPER-T250 (red and blue dashed lines). In general, the SUPER-THOM configuration has a much stronger thermodynamic response to CDNC perturbations than the SUPER-MORR configuration, with latent heating rates consistently stronger throughout the atmosphere (Fig. b). Overall, there is little evidence of convective invigoration (defined here as increases in upper tropospheric heating, updraught strengths, cloud top height and surface precipitation) under increased CDNC values in either bulk microphysics scheme. Although both schemes show increased latent heating in the upper troposphere and decreased heating at mid-levels under polluted conditions (Fig. b), it has already been shown that there is no evidence of increased surface precipitation (Fig. ), and the upper tropospheric peak in latent heating can be seen to correspond to an increase in ice (SUPER-M250) or snow (SUPER-T250) at these levels (Fig. ). There is no systematic or consistent evidence of increased mean updraught velocity in the convective cores following the method of under polluted conditions (not shown) or in increased cloud top heights of the convective cores (Fig. c and d). This may not be surprising, as it has been suggested that bulk microphysics schemes are unable by design to produce convective updraught invigoration effects due to limitations in their representation of nucleation, sedimentation and the way in which saturation adjustment limits diffusional growth detailed in. Indeed, found no latent heating effect of increased CCN in a bulk scheme used to simulate idealised deep convection, whereas with a bin scheme increased latent heating aloft was demonstrated. However, found that saturation adjustment methods used in bulk schemes could explain differences in the response of cold pool evolution and convective dynamics between bin and bulk schemes to aerosol loading but could not explain large differences in the response of surface precipitation. Further, some simulations using bulk schemes have identified invigoration-like effects under aerosol loading. For example, found evidence of convective invigoration under increased aerosol loading in a bulk scheme under weak shear conditions (and suppressed convection under strong shear), similar to the findings of who found the same response in a bin scheme. also found higher overshooting tops and larger sizes of cumulonimbus in a weak shear environment with increased aerosol loading. Thus although our results agree with the body of the literature which does not identify a convective updraught invigoration effect when bulk microphysics schemes are used, this is not necessarily attributable to the saturation adjustment method alone and may also only hold for the particular convective environment (idealised supercell in strong shear) we consider.

The results presented in Sect.  and indicate that not only is the way in which the schemes differ from each other not systematic between cases of convection, but the difference between the response of the two schemes to CDNC across types of convection is also not systematic. The largest difference between the microphysics schemes in the real-data Congo basin simulations occurs in the liquid-phase hydrometeor development and response to CDNC. Making the assumption that the liquid phase is the first to respond to CDNC perturbations and the perturbation subsequently propagates to the ice phase, we consider a case of precipitating shallow cumulus convection to investigate the liquid-phase differences between the schemes. Note that the “baseline” hydrometeor profiles in Fig.  show data from the configurations using a prescribed CDNC value of 100 cm-3 (rather than the baseline value of 250 cm-3 used in the Congo basin and idealised supercell deep convection cases in Figs.  and ), as this is more appropriate for a pristine marine environment. Even when we restrict our simulations to the liquid phase, differences in the simulated hydrometeor classes are evident. The dominant domain-mean difference between the two schemes in the RICO case is clearly in the rain profile, with RICO-T100 producing significantly more rain than RICO-M100. Very little rain is present in the RICO-M100 configuration (Fig. a and d), whilst the RICO-T100 configuration produces a peak domain-mean rain mass of about 10-6 kgkg-1 (Fig. b). The liquid cloud profile is similar in both schemes, with RICO-M100 forming more cloud mass than RICO-T100 between 805 and 775 hPa in both the domain mean and hydrometeor-class mean sense (Fig. c and f).

The response of the hydrometeors to CDNC perturbations also differs between schemes in the warm-rain RICO case (Fig. ). In the RICO-MORR configuration, domain-mean rain and cloud mass both decrease under polluted conditions, although the rain response is very weak (on the order of 10-8 kgkg-1) and the dominant response is a reduction in liquid cloud mass (Fig. a). In contrast, a reduction in rain mass is the dominant hydrometeor response under polluted conditions in the RICO-THOM configuration, and the decrease is nearly 2 orders of magnitude greater than that in RICO-MORR (Fig. b). The liquid cloud response to polluted conditions in RICO-THOM is weaker than the rain response but still stronger than the cloud response in RICO-MORR. Cloud mass decreases under polluted conditions between 935 and 825 hPa but increases at higher levels (Fig. b). Note that once again the response of the simulated hydrometeors to CDNC perturbations differs between cases: under polluted conditions, RICO-MORR exhibits a decrease in cloud and rain mass, while CONGO-MORR exhibits a decrease in rain mass with little response in the liquid cloud (Fig. a), and SUPER-MORR shows almost no liquid-phase response at all (Fig. a). Likewise, RICO-THOM exhibits a decrease in rain and an increase in cloud mass under polluted conditions, while CONGO-THOM exhibits similar behaviour (Fig. b), but SUPER-THOM shows a decrease in rain mass with little response in the liquid cloud (Fig. b). When mean profiles of each hydrometeor class are considered, the two schemes actually show similar responses to CDNC (increased upper-level cloud mass and suppressed rain; Fig. c and d). This indicates that the main response to CDNC in this case is not through the individual microphysical processes but through the absolute amounts of cloud and rain that are generated.

To illustrate the difference in the strength of response of the schemes to CDNC, total accumulated surface rain is shown for each RICO configuration in Fig. a along with an extra configuration using a “very pristine” CDNC value of 50 cm-3 and a series of sensitivity tests that will be discussed later. The 50 cm-3 CDNC configuration has been added because even at a prescribed CDNC value of 100 cm-3 very little rain production occurs in the RICO-MORR configuration. Warm rain formation differs strongly between schemes: very low CDNC values are required for the RICO-MORR configuration to produce any rain, whereas RICO-THOM produces significantly more rain at all CDNC values (Fig. a). Even under very pristine conditions, the RICO-M50 configuration produces an order of magnitude less rain than RICO-T50 (Fig. a). The different schemes also respond differently to CDNC perturbations. Rain production in RICO-MORR (which produces much less rain than RICO-THOM) shuts down very quickly as CDNC is increased: rain amounts are on the order of 102 mm at a CDNC value of 50 cm-3, 101 mm at a CDNC value of 100 cm-3 and 10-1 mm at a CDNC value of 250 cm-3; rain production ceases completely at a CDNC value of 2500 cm-3 (Fig. a). In contrast, rain production persists for much larger CDNC values in RICO-THOM: rain amounts are on the order of 103 mm at CDNC values of 50 cm-3, 102 mm at CDNC values of 100 cm-3 and 101 mm at CDNC values of 250 cm-3. While rain amounts are very low at CDNC values of 2500 cm-3 (on the order of 10-5 mm), rain production has not shut down completely (Fig. a).

showed that the rain rates predicted by different autoconversion formulae in bulk schemes can vary by orders of magnitude. This sensitivity of results is also well highlighted in ; note in particular their reference to . Autoconversion is parameterised differently in the two microphysics schemes used in the current paper. The Thompson scheme follows an adaptation of , while the Morrison scheme follows the method of .

and justify their choice of an adapted version of the autoconversion parameterisation through favourable comparison to results from the bin scheme of . Furthermore, implementation of in the Thompson scheme begins the collision–coalescence production of warm rain at almost exactly 14 µm. It is known that raindrop onset begins when the mean volume radius exceeds a critical value of 13 to 14 µm . This is one of the principle reasons the autoconversion scheme was chosen by rather than .

While the autoconversion scheme was initially developed and applied for LES of stratocumulus, other than varying the prescribed values of cloud droplet number concentrations we run the microphysics schemes in their baseline configurations. Thus, although we do not advocate the use of for non-stratocumulus cases, the Morrison scheme is frequently used for simulations of deep convection. Similarly, as the autoconversion scheme was developed for LES-scale studies, the authors recognise the potential importance of subgrid cloud variability at the scales used in the present study. However, we note that we are running the model and microphysics schemes in the typical set-up for a convection-permitting model (that is, neglecting subgrid cloud variability), as one of the main aims of this study is to highlight uncertainty in commonly used model configurations which are exactly based on these schemes.

The autoconversion rates as a function of cloud water content for each of the model configurations are shown in Fig. . Also shown is the cloud water content (up to the mean plus 2 standard deviations) of each configuration . It is immediately clear that the threshold cloud liquid content for autoconversion in the Morrison scheme (solid lines) is significantly lower than that in the Thompson scheme (dashed lines); i.e. rain production can occur at much lower cloud liquid water contents in Morrison. It is also clear from the mean, (mean +1 SD) and (mean +2 SD) cloud water content limits that rain production through autoconversion ought to be possible in all model configurations. However, despite the higher cloud water content threshold for autoconversion in the Thompson scheme, autoconversion rates are much greater once the threshold is reached, and liquid cloud is converted to rain much faster in Thompson than in Morrison. From Fig. , it appears that the threshold for autoconversion is unlikely to be reached very often in any of the T2500 cases. In the deep convective cases, rain can be generated through ice- and mixed-phase processes, but in the RICO warm-rain case this cannot occur. This explains why, compared to more pristine conditions, cloud mass increases in RICO-T2500 while rain mass decreases (Fig. b).

Because Fig.  indicates that the autoconversion threshold may be at least in part responsible for this response in the RICO-THOM case, we replace the autoconversion parameterisation in the Morrison scheme with that from the Thompson scheme and vice versa. We use the notation M100T to denote the Morrison microphysics scheme (at a CDNC value of 100 cm-3) with Thompson autoconversion that of and T100M to denote the Thompson scheme with Morrison autoconversion that of. Differences in the domain-mean hydrometeor mixing ratio profiles for each of the autoconversion swapped configurations in the RICO case are shown in Fig. . It is immediately clear that, in the warm-rain configuration, simply swapping the autoconversion treatment makes the hydrometeor developments of the microphysics schemes much more like each other. The differences between the RICO-M100 configuration with the Morrison and Thompson autoconversion parameterisations (Fig. a) is quantitatively and qualitatively very similar to the difference between the RICO-M100 and RICO-T100 configurations (Fig. c). Likewise, the difference between the RICO-T100 configuration with the Morrison and Thompson autoconversion parameterisations (Fig. b) and finally the difference between the RICO-T100 configuration with the Morrison autoconversion parameterisation and the RICO-M100 configuration with the Thompson autoconversion parameterisation (Fig. c) are also very similar to the difference between the RICO-M100 and RICO-T100 configurations (Fig. c).

Similarly, swapping the autoconversion parameterisations between the microphysics schemes in the RICO cases makes the surface rain production of the microphysics schemes much more similar. The accumulated surface rainfall in the RICO-M100T configuration looks much more similar to the surface rainfall in the RICO-T100 configuration than it does to the RICO-M100 configuration (Fig. a). Rain amounts are on the order of 102 mm in RICO-M100T and RICO-T100, whereas in RICO-M100 it is 2 orders of magnitude smaller (Fig. a). Likewise, the accumulated surface rainfall in the RICO-T100M configuration is on the order of 101 mm compared to 102 mm in the standard RICO-T100 case (Fig. a). To further test the importance of autoconversion in the liquid phase simulations, we first turn off autoconversion completely in the 100 cm-3 CDNC simulations, and then allow autoconversion to occur but prevent the accretion of cloud water by rain. By design, in the absence of ice processes no precipitation occurs without autoconversion of cloud water to rain (Fig. a, M100noAUTO and T100noAUTO). However, in the RICO 100 cm-3 CDNC liquid-phase configuration, the Thompson scheme can produce surface rain from autoconversion alone (albeit 2 orders of magnitude less than when rain can also accrete cloud water (Fig. a, T100noACCR and T100), showing that autoconversion acts almost like a “trigger” in this scheme after which accretion takes over the rain production process. (Indeed, in nearly all schemes, rain formation from accretion, once triggered, is orders of magnitude larger than from autoconversion.) In contrast, zero surface precipitation is produced in RICO M100noACCR (Fig. a), showing that in this (liquid-phase only) configuration the Morrison scheme requires both the autoconversion of cloud droplets to rain and the accretion of rain by cloud droplets in order to produce surface precipitation.

Despite the significant effect of autoconversion in the liquid-phase simulations, changing the autoconversion parameterisation in the idealised supercell case has very little effect on the hydrometeor development (results not shown). This is unsurprising, as ice- and mixed-phase processes will dominate this shear-driven deep convective environment. However, the Congo basin configurations show large differences between microphysics schemes in the partitioning of water into liquid and ice phases (CONGO-THOM produces much more liquid cloud; CONGO-MORR produces much more ice). In the CONGO-THOM configurations, the liquid-phase response to increased CDNC is also very similar to the RICO-THOM response (increased liquid cloud mass and decreased rain mass; Fig. b). When the Thompson autoconversion treatment is implemented in the Morrison scheme, rain production in the southern half of the domain ceases in CONGO-M250T, and the liquid phase is instead represented by low-level cloud with structure similar to the CONGO-T250 configuration (Fig. a compared to Fig. e). To test if radiative effects associated with large amounts of anvil ice drive or contribute to the differences in low cloud, we also set the ice extinction coefficient to zero in both the longwave and shortwave radiation schemes in CONGO-M250. However, this has no effect on the low-cloud characteristics (Fig. e compared to Fig. b), and we therefore conclude that autoconversion of cloud water to rain is the factor dominating the absence of low-level cloud in the south of the domain in the CONGO-MORR simulations. In contrast, autoconversion is a less significant process in the CONGO-T250 configuration. Implementing the Morrison autoconversion treatment in the Thompson scheme has very little effect on the hydrometeor structure in the CONGO-T250M configuration compared to the CONGO-M250 configuration (Fig. b compared to Fig. b). As a final test, the autoconversion process is turned off in both of the microphysics schemes. This confirms that autoconversion dominates the lack of low cloud in CONGO-M250: the resulting liquid-phase hydrometeor structure (Fig. c) is similar to both CONGO-T250 (Fig. b) and CONGO-M250T (Fig. a). This also confirms that autoconversion is much less significant in the CONGO-THOM configurations: the bulk hydrometeor structure when autoconversion is turned off in CONGO-T250 (Fig. d) is very similar to both CONGO-T250 (Fig. b) and CONGO-T250M (Fig. b).

We have shown that the autoconversion process is responsible for the removal of the large cloud mass at low levels in the model configuration with the Morrison microphysics scheme. We also see that this low-level liquid-phase cloud mass forms when we run the same simulation using the WRF bin microphysics implementation the SBM part of the Hebrew University Cloud Model;, although to a lesser extent than in the Thompson simulations, and the warm cloud produced by the WRF-SBM produces rain (Figs. S2 and S3 in the Supplement). We therefore suggest that it is not the Thompson scheme per se which is responsible for producing the low-level cloud mass, but rather the larger-scale meteorological conditions in which these simulations are performed.

A further significant difference between the two schemes in the Congo simulations is the generation of large amounts of upper-level ice in CONGO-M250, which is not present in CONGO-T250 (Fig. b and e). In the Thompson scheme, the fraction of ice mass with a diameter greater than 125 µm is instantaneously transferred into the snow category . The same threshold size for cloud ice autoconversion to snow is used in the Morrison scheme, but the process is parameterised differently . Because the Morrison scheme appears to produce large amounts of anvil cloudiness for the Congo case, which is not seen in the observations (Fig. ), we perform further sensitivity tests in which we reduce the threshold size for cloud ice autoconversion in the Morrison scheme to 50 % of its original value (Fig. f) and 10 % of its original value (Fig. g). We then finally replace the autoconversion of cloud ice to snow in the Morrison scheme with the parameterisation used in the Thompson scheme (Fig. h). In all tests, the upper-level anvil ice is reduced significantly. Using the lowest value of the threshold size for cloud ice autoconversion reduces the anvil cloud because almost all of the ice is immediately converted to snow (Fig. g). However, using the Thompson ice autoconversion representation in the Morrison scheme significantly reduces the amount of cloud ice in the simulation, and all of the detrained anvil ice is removed (Fig. h). This suggests that for the particular Congo simulation we have investigated, the conversion of cloud ice to snow is the main factor leading to the significant difference in anvil cloudiness between the two schemes and is responsible for the difference in upper-level cloud between the CONGO-M250 simulation and the observations (Fig. ). Indeed, we note that in equivalent simulations performed with the WRF-SBM, the same persistent upper-level ice forms (Fig. S2 in the Supplement). This shows that differences resulting from conversion of one ice category into another is a limitation of any scheme, whether bin or bulk, which uses fixed ice categories. Our results provide further evidence that the use of discrete ice-phase hydrometeor categories may be detrimental to the correct simulation of cloud and suggest that new schemes which do not use such partitioning may give better results e.g..

Our results show little impact of aerosol on precipitation in the Congo basin (Figs. b and ), which is also seen when considering total accumulated surface precipitation (Fig. b), although CONGO-T2500 exhibits weak precipitation suppression under polluted conditions. This may be due to the longer duration of these simulations performed over a larger domain, allowing the interaction of many cloud systems rather than considering the lifetime of a single isolated cloud. However, we also see that although the representation of autoconversion has a significant effect on the vertical hydrometeor structure in the CONGO-M250 configurations (Figs. b, a and c), it has a much weaker effect on total surface precipitation (Fig. b, M250, T250, M250T, T250M, M250noAUTO and T250noAUTO). This is perhaps unsurprising, as the dominant contribution to the accumulated surface precipitation over the Congo domain will be from ice processes in the convective region and not from the liquid-phase cloud. Although the lack of impact of aerosol on precipitation in the Congo simulations may be due to the use of bulk schemes in this study for the reasons detailed in , and perhaps a different response would be seen using a bin scheme e.g., other studies using bulk and bin–bulk schemes have identified aerosol impacts on precipitation of up to about 15 % e.g.. Indeed, even studies using bin schemes have been shown to have little impact on total precipitation, although they induce a shift in rainfall rates . Therefore, we note again that the choice of microphysics scheme, rather than aerosol response in either scheme, is the dominant contribution to uncertainty in the total precipitation.