Aerosols play significant roles in the climate system by modifying the radiation budget (aerosol–radiation interactions, ARIs) and the hydrological cycle through interactions with clouds (aerosol–cloud interactions, ACIs). Quantitative estimates of anthropogenic aerosol forcing, however, are still largely uncertain because of the complex interactions among aerosols, clouds, and climate across wide spatiotemporal scales . Reducing these uncertainties associated with the effect of aerosol forcing on climate is one of the most challenging issues in climate science .

A key uncertainty arises from the complex response of clouds to aerosol perturbations . Clouds are considered to respond to perturbed aerosols in two opposing ways, i.e., the so-called “cloud lifetime” effect and the “buffered system” effect , in a regime-dependent manner . The cloud water susceptibility to aerosols depends strongly upon cloud type , as well as ambient environmental conditions , which results in non-monotonic cloud responses and therefore diverse impacts on climate .

These observational findings are also supported by process modeling studies using large-eddy simulations . General circulation models (GCMs), however, show a large spread in cloud susceptibility to aerosols , and they tend to overestimate the magnitude of ACI compared with satellite retrievals . This means that current GCMs are not able to reproduce the buffering of cloud responses to aerosol perturbations . Aerosol-induced radiative forcing at the top of the atmosphere (TOA) that includes rapid adjustments caused by ACI, termed effective radiative forcing (ERFaci), varies widely among GCMs . This results in a “best estimate” of global annual mean ERFaci of -0.45 Wm-2 with a 90 % confidence interval of -1.2 to 0.0 Wm-2 , as reported in the fifth assessment report of the Intergovernmental Panel on Climate Change (IPCC AR5). This uncertainty range has remained large e.g., since the early IPCC reports.

As a consequence of the challenges described above, GCMs tend to show more negative ERFaci than that inferred from satellite retrievals even though retrieval errors are considered . This suggests that current GCMs may be missing a compensating warming effect caused by aerosols. The “missing warming” in GCMs may be solved by taking aerosol effects on (i) deep convective clouds and (ii) mixed-phase clouds into consideration, as these effects can modify the ice microphysics due to aerosols and also lead to an adjustment in the longwave component . A recent multi-model analysis demonstrated that simpler GCMs that parameterize the aerosol effect on liquid-phase clouds alone have negligibly small longwave ERF, whereas more sophisticated GCMs that include microphysical adjustments of ice- and mixed-phase clouds as well as liquid-phase clouds produce larger-magnitude ERF values for both the terrestrial (ERFLW) and solar (ERFSW) components. The changes to ERFLW and ERFSW were found to nearly cancel each other out and result in a net ERF (ERFNet) of a magnitude that is similar to that generated by the simpler GCMs. The robustness of this near cancelation, however, largely depends on how microphysical processes in ice- and mixed-phase clouds, which are typically much more complex than in liquid-phase clouds , are represented in GCMs.

Among these processes, precipitation processes involving falling hydrometeors (i.e., rain and snow) are particularly simplified in current GCMs, which is likely to lead to nonnegligible uncertainty in ERFaci . In general, precipitation is treated diagnostically in GCMs (hereinafter “DIAG”), with precipitation being immediately removed from the atmosphere within a single model time step. This over-weights autoconversion relative to accretion to produce precipitation , which results in the pronounced sensitivity of cloud water to aerosols because autoconversion is the only process that directly depends on aerosols . Snow also has significant effects on collection processes among other hydrometeors , as well as on atmospheric circulation . However, snow-induced impacts on ERFaci are much less understood because extremely limited GCMs incorporate prognostic precipitation with the radiative effects of falling hydrometeors (see discussion in ).

This study investigates this unexplored area of ACI, with a particular focus on precipitation (rain and snow) processes and their impacts on ERFaci and with the goal of advancing our understanding of the fundamental linkage of microphysical process representations to their macroscopic climate effects. For this purpose, we use a recently developed global aerosol–climate model, MIROC6-SPRINTARS , which is implemented with a two-moment prognostic precipitation scheme (hereinafter “PROG”) that includes the radiative effects of precipitation . Through a comparison with the traditional DIAG scheme, we use the PROG-scheme model to identify the source of discrepancies in ERFaci between GCMs and satellite observations that are related to precipitation processes. A suite of sensitivity experiments is also performed with the model to isolate the relative contributions of different microphysical processes to ERFaci and to quantify how uncertainties inherent in these processes translate to ERFaci uncertainty. This single-model approach has the advantage of not being affected by varying physics representations, as in the case of multi-model analysis (see “Materials and methods”).

Figure  compares geographical distributions of ERFaci simulated by the DIAG and PROG models. In DIAG, a strong negative ERFaci is observed over East Asia, Europe, and North America where anthropogenic pollution dominates. This is attributed to the cloud lifetime effect caused by anthropogenic aerosols, which increases low warm clouds and hence shortwave reflectance. The global annual mean ERFaci reaches -1.7 Wm-2, which is outside the bound of the uncertainty range (-1.2 Wm-2) in IPCC AR5. The geographical pattern is consistent with other GCMs .

In PROG, however, the majority of the strong negative forcing over anthropogenic regions is reduced significantly, resulting in a reduction of around 54 % in global-mean ERFaci. Although the geographical pattern is somewhat different from previous reports using other GCMs (discussed in the next section), the global mean ERFaci (-0.8 Wm-2) is much closer to satellite-based estimates . The total aerosol ERF associated with ARI and ACI (ERFari+aci) in PROG (-1.1 Wm-2) is only half that generated by DIAG (-2.1 Wm-2).

This significant reduction in ERFaci in PROG results from a substantial weakening of ERFaciSW, particularly over midlatitudes of the Northern Hemisphere, and enhanced warming of ERFaciLW over low latitudes in both hemispheres (Fig. c). The zonal distribution shows that stronger (weaker) ERFaciLW accompanies stronger (weaker) ERFaciSW, which is in line with . To understand the impact of precipitation treatment on ERFaci, decompositions of global mean ERFaci into its SW and LW components are shown for alternate configurations of precipitation in MIROC6 (Fig. ). Figure  confirms that the significant reduction of ERFaci in PROG is contributed to by both increased ERFaciLW and weakened ERFaciSW, in stark contrast to previous CMIP5 model results in which cloud-ice-induced changes to ERFSW and ERFLW cancel each other out to result in few net ERF changes within the DIAG framework. This difference in the present study from previous results is attributed to the snow-induced modulation of ACI newly incorporated into our model.

The impact of snow on ACI can be understood in more detail using the results shown in Fig. , which includes two intermediate versions of PROG, i.e., one that incorporates prognostic rain but diagnostic snow (PRDS) to isolate the relative impacts of rain vs. snow on ERFaci and one that represents prognostic rain and snow but without the radiative effects of snow (OFF / SnwRad). Regarding the LW component, the global mean ERFaciLW of PROG (+0.7 Wm-2) is more than twice as large as those of DIAG (+0.2 Wm-2) and PRDS (+0.3 Wm-2). The OFF / SnwRad simulation also shows weaker ERFaciLW relative to the standard PROG simulation (Fig. ). These results suggest that the warming LW effect comes mainly from adjustments induced by snow together with its radiative effects, in addition to cloud-ice effects included in CMIP5 models as well as our model. The ERFaciLW is significant over the Indian Ocean and Southeast Asia (not shown), which is also similar to the other model, CAM5-MARC-ARG . This is attributable to the increased ice nuclei (IN) due to biomass burning for example, partly supporting the convective invigoration although GCMs do not have the capability to resolve the convective cloud systems. The increased IN results in a faster glaciation and thus enhances snowfall due to the WBF process (i.e., glaciation indirect effect). These mixed- and ice-phase microphysical processes are more elaborated in the PROG scheme, and the associated LW change induced by snow is incorporated only in PROG, which contributes to the higher ERFaciLW across the globe.

The PROG scheme also reduces the SW component (ERFaciSW) relative to DIAG, particularly over anthropogenic regions (Fig. b) in the Northern Hemisphere midlatitudes. A well-known mechanism for the reduction in ERFaciSW is the enhancement of accretion with a smaller contribution from autoconversion, as in PROG (not shown), with only the latter process depending upon the cloud droplet number concentration (Nc) . The smaller contribution of autoconversion in PROG mitigates the excessive cloud water susceptibility to aerosols that occurs in DIAG models . However, Fig.  shows that the replacement in liquid-phase precipitation alone from DIAG to PRDS cannot explain the significant reduction of ERFaciSW from DIAG to PROG, suggesting that ice-phase processes involving falling snow influence the magnitude of ERFaciSW, as discussed in the next section.

This reduction of ERFaciSW in PROG relative to DIAG is also impossible to explain by the response of cloud ice alone, because cloud ice should increase ERFaciSW towards more negative values because of aerosol-induced increases in cloud optical thickness. This is indeed what is happening with the DIAG framework in the CMIP5 multi-model results , in which models with aerosol effects on cloud ice (not snow) show much stronger ERFaciSW, which is large enough to cancel the enhancement of ERFaciLW. In contrast, our PROG model reduces ERFaciSW. We hypothesize that the prognostic treatment of snow plays an important role in weakening ERFaciSW through microphysical processes involving cloud water and snow, as discussed below.

Next, we discuss the role of prognostic precipitation in determining ERFaci by addressing the following two questions raised in the previous section:

Why does the geographical pattern of ERFaci in PROG differ from that of DIAG?

To this end, we first explore how precipitating hydrometeors can modulate the cloud water susceptibility to perturbed aerosols. Figure  shows how the change in the cloud liquid water path (CLWP) relates to changes in precipitating hydrometeor paths, i.e., the rainwater path (RWP) and the snow water path (SWP), through pre-industrial (PI) to present-day (PD) changes in aerosols. The PD-minus-PI change (susceptibility) in RWP is highly correlated (r=0.59) with that in CLWP (Fig. a). We interpret this strong correlation to be the result of the close co-variance of cloud and rainwater through aerosol perturbations, with the cloud water being a direct source of the rainwater. The PD-minus-PI change in SWP is also positively correlated, though weaker (r=0.39), than that in CLWP (Fig. b), suggesting that precipitating snow also co-varies with cloud water through aerosol perturbations. Given that SWP is significantly larger than RWP in our model (Fig. S2; see also ), and that snowflakes, with residence times longer than those of rain, are more likely to interact with clouds, the increased CLWP caused by anthropogenic aerosols can act as an efficient source of snow via interactions among cloud droplets and snowflakes (e.g., riming), likely resulting in the evident robust positive relationship.

These positive correlations between precipitating hydrometeors and cloud water suggest that aerosol-induced increases in cloud mass are caused, in part, by increases in rain and snow in PROG, in contrast to those caused by increases in cloud water and ice alone in DIAG. Given that raindrops and snowflakes are optically much thinner in the SW spectrum than cloud droplets and ice crystals, respectively, increases in precipitating hydrometeors can explain both the stronger ERFaciLW and weaker ERFaciSW in PROG compared with DIAG (Figs.  and ). Furthermore, falling snow is more likely to deplete underlying cloud droplets in PROG, with its explicit representation of the riming process, which can lead to a reduction of cloud water susceptibility to aerosols. This proposed mechanism can also explain the systematic change in the geographical distribution of ERFaci between DIAG and PROG (Fig. ). Indeed, regions with a significant reduction in ERFaci (i.e., over East Asia, Europe, and North America) correspond well to those with large values of SWP (Fig. S3), where the PD-PI increase in CLWP is also reduced significantly (Fig. S4). These results lend further credence to the hypothesis of snow-induced buffering of ACI in our model.

The buffering process, via interactions among hydrometeors described above, depends strongly on the fundamental uncertainty in model representations of various microphysical processes. We therefore now further explore how ERFaci and its buffering by precipitation processes are sensitive to microphysical process representations as summarized in Fig.  (see also Sect. 2.2 for details of experiments). The processes examined here are the autoconversion of liquid droplets; the Wegener–Bergeron–Findeisen (WBF) process; the aggregation of ice crystals, riming, and ice nucleation by freezing aerosols, which are all important sources of uncertainty in GCMs . As expected, the simulated ERFaci is highly sensitive to the autoconversion scheme used, mainly because of its varying dependence on Nc among the various schemes . A different liquid autoconversion scheme with PROG can change ERFaci by 39 %, from -18 % to +21 % (blue bars in Fig. ). The impacts of the autoconversion scheme on ERFaci, however, are smaller than those of the treatment of rain and snow (ca. 54 % change).

The mixed- and ice-phase processes (WBF, aggregation, and riming), represented more explicitly with a larger degree of freedom in PROG than in DIAG, can change ERFaci by 15 %, from -13 % to +2 % (cyan bars in Fig. ). Among the mixed- and ice-phase microphysics processes, the process found to most influence ERFaci is the riming of cloud droplets on snow, supporting the hypothesized mechanism of snow-induced buffering of ACI discussed above. The magnitude of ERFaci is sensitive to ice nucleation processes as well (green bars in Fig. ), because the change in ice number concentration directly controls the size of the crystals and thus the conversion timescale from ice to snow in our model. Although the ERFaci variations with changing liquid and ice microphysical processes do not reach the difference of ERFaci between the DIAG and PROG (i.e., 54 %), both wet scavenging of aerosols and coalescence scavenging of cloud droplets also contribute to the ACI reduction due to the accretion-driven buffering mechanisms , which should explain the remaining part of the ERFaci difference.

In summary, we found that the treatment of precipitation (PROG vs. DIAG) is the most influential factor controlling ERFaci (red bars in Fig. ) among all of the “tunable knobs” associated with the various microphysical processes in our model. It should also be emphasized that the ERFaci change caused by the precipitation treatment (ca. 54 % in magnitude), absent from previous climate modeling studies, has the potential to resolve some of the differences between satellite estimates of ERFaci and GCMs . These findings need to be tested further using other GCMs as they incorporate prognostic precipitation in future studies.

In this study, the sensitivities of ERFaci to various treatments of precipitation and microphysical process representations in a GCM have been systematically examined. As few GCMs incorporate explicit representations of two-moment prognostic precipitation with the radiative effects of precipitating hydrometeors – e.g., CAM6 MG2/MG3 , E3SM , GISS-E3, and MIROC6 CHIMERRA – we used a single model framework to evaluate the sensitivities. This also allowed us to avoid uncertainties from inter-model differences in parameterizations other than the targeted processes.

We found that the treatment of precipitation in GCMs (PROG vs. DIAG) has a significant impact on the magnitude of ERFaci (Figs.  and ), which we interpret to be driven mainly by collection processes among precipitating snow and cloud droplets (i.e., riming). As the SWP is more than twice as large as the RWP in our PROG model, and is in good agreement with satellite retrievals , falling snowflakes efficiently accrete and deplete the underlying cloud water, thus partly canceling the CLWP response to aerosols. Changes in RWP and SWP through PI to PD aerosol perturbations were also positively correlated with those in CLWP (Fig. ), suggesting that snow can co-exist with cloud water to a degree sufficient to buffer the cloud water response to aerosol perturbations (Fig. ). The signatures of the snow-induced buffering are also found geographically over regions with significant reductions in ERFaci (e.g., East Asia, Europe, and North America) that correspond closely to regions with particularly large SWP (Figs.  and S3). Sets of sensitivity experiments, performed both with and without snow radiative effects, did not reveal a significant difference in ERFaci as a result of the near cancellation of SW and LW changes caused by snow. This means that the prognostic treatment of precipitation itself is critical for the buffering of ACI. Accordingly, the impact of a prognostic treatment of precipitation on the magnitude of ERFaci was greater than changes to any of the other tunable knobs inherent to the various microphysical processes (e.g., autoconversion, ice microphysics, and ice nucleation). Notably, precipitation-driven buffering effects (ca. 54 % change in ERFaci) can broadly explain the current model–observation discrepancy in estimated ERFaci .

However, the results presented here are based on a single GCM framework and need to be replicated using other GCMs as they incorporate prognostic precipitation frameworks in the future . This is particularly true because little is known about aerosol influences on mixed- and ice-phase clouds as well as deep convective clouds and cirrus clouds at a fundamental process level, and the degree of microphysical complexity differs widely among GCMs . Although the responses of clouds and precipitation to aerosol perturbations are therefore likely to be model dependent, the sign of the response of ERFaci to the precipitation framework and microphysical processes is consistent with a previous assessment using CAM5/MG2 , suggesting that the major findings of this study will apply across the models. Thus, it is left for important future studies to quantify the inter-model spread of ERFaci sensitivity to microphysical processes and their interplay with precipitation processes as more GCMs begin to include prognostic precipitation. Furthermore, a theoretical approach and idealized process modeling are also urgently required to solidify the process-level understanding of the snow-induced buffering hypothesis, which constitutes our important future work beyond the present study.

This study primarily focused on ERFaci sensitivities to the CLWP adjustment rather than cloud fraction adjustment, because aerosol effects are directly linked to the CLWP change through the modification of the mass conversion rate from cloud water to rainwater that itself relates to the treatment of precipitation (i.e., DIAG vs. PROG). However, it is important in future studies to separate the ACI into the Twomey forcing and rapid adjustments of CLWP and cloud fraction e.g., for better understanding of how the treatment of precipitation influences micro- and macroscopic cloud properties , which relates to the fundamental inter-model spread in ERFaci .