The Met Office Unified Model (UM) vn13.0 with the Cloud AeroSol Interacting Microphysics Module (CASIM; Field et al., 2023) is used in this study. Simulations are conducted using the Regional Atmosphere and Land (RAL) 3.2 configuration (Bush et al., 2023). The UM employs the ENDGame non-hydrostatic dynamical core with a semi-Lagrangian formulation (Wood et al., 2014). Subgrid cloud fraction is diagnosed using the bimodal cloud fraction scheme of Van Weverberg et al. (2021) for coupling to radiation in the RAL3.2 configuration, and no convection parameterization is active. Subgrid mixing in the boundary layer is represented by the Lock et al. (2000) scheme. Further details are provided in Bush et al. (2025).

Cloud microphysics are represented using the CASIM scheme, configured in two-moment mode to prognostically predict both mass and number concentrations for five hydrometeor species: cloud droplets, raindrops, ice crystals, graupel, and snow. Particle size distributions follow a generalized gamma distribution. In CASIM, the cloud droplet number concentration (Nd) can be either prescribed as a fixed value or diagnosed from background aerosol. To systematically examine aerosol sensitivity and isolate the effects of prescribed droplet-number conditions, we adopt the fixed Nd approach in this study. As in RAL3.2, a vertical taper is imposed on the fixed Nd to avoid unrealistically high droplet concentrations at upper levels and the associated excessive homogeneous freezing (Bush et al., 2025).

Rain is generated through auto-conversion, accretion, and melting of ice-phase particles, while losses include evaporation and freezing by ice. The separate treatment of mass and number concentrations allows for variability in hydrometeor size and fall speed, which directly affects sedimentation and evaporation. In contrast to single-moment schemes where particle size is fixed for a given mass, CASIM allows dynamic adjustment of size distributions based on aerosol and microphysical evolution.

Heterogeneous ice nucleation is parameterized using the default scheme of Cooper (1986), which prescribes ice crystal number concentrations as a function of temperature. Immersion freezing of raindrops is parameterized following Bigg (1953), with freezing probability depending on drop volume and temperature. When freezing occurs, rain mass and number are converted to graupel, providing a bulk representation of raindrop freezing in CASIM (Field et al., 2023; Miltenberger et al., 2020). In addition, rain–ice interactions are represented through accretion/collection processes in CASIM, and rain collecting ice can contribute to snow or graupel formation depending on the hydrometeor state and model configuration. Homogeneous freezing of cloud droplets is parameterized to occur below -38 °C (Field et al., 2023). In this study, aerosol perturbations are represented by prescribed Nd. INP concentrations are not varied and are not linked to Nd; therefore coupled CCN–INP effects are not represented. In CASIM, aggregation and accretion among ice-phase hydrometeors are represented through parameterizations for self-collection and inter-species collection processes involving ice, snow, and graupel. The associated collection efficiencies are prescribed in the model and are temperature-dependent.

In addition to primary ice nucleation, this study incorporates multiple secondary ice production (SIP) mechanisms. The default SIP process in CASIM is rime-splintering (RS), which produces ice splinters through the riming of supercooled droplets by graupel and snow between -2.5 and -7.5 °C, with peak efficiency at -5 °C (Hallett and Mossop, 1974; Mossop, 1985). The production rate is based on the cloud liquid accreted by graupel and snow, using a default efficiency of 350 ice splinters per milligram of rimed liquid. In our previous study (Sun et al., 2025), we implemented three additional SIP mechanisms into CASIM: ice-ice collisional breakup (BR), and droplet shattering during symmetrical and asymmetrical freezing (Mode 1 and Mode 2). The BR parameterization follows the theoretical formulation of Phillips et al. (2017b), and accounts for collisions between ice-phase particles (graupel, snow, and ice crystals). The production rate depends on collision kinetic energy, rimed fraction, temperature, and particle size. For droplet shattering, we adopt the two-mode parameterization from Phillips et al. (2018). Mode 1 (M1) describes the collisions between frozen drops and smaller ice particles, with splinter numbers fitted to laboratory data using Lorentzian curves, peaking around -15 °C. Mode 2 (M2) involves collisions between freezing raindrops and more massive ice particles, where fragmentation is governed by the ratio of collision kinetic to surface energy. An empirical factor of 0.3, representing the ice-containing fraction of fragments, is applied following James et al. (2021).

Radiative transfer is handled by the SOCRATES scheme (Edwards and Slingo, 1996; Manners et al., 2024), which employs the two-stream approximation for fluxes and spherical harmonics for radiance. The single-scattering properties of ice crystals are derived from both the ice water content and the number concentration of ice-phase hydrometeors predicted by CASIM. This allows changes in ice particle characteristics associated with different SIP mechanisms to influence the simulated cloud optical properties. Turbulence is represented by the blended boundary layer scheme (Boutle et al., 2014), which incorporates both local and non-local mixing processes.

Simulations were conducted for the 6 February 2006 case from the Aerosol and Chemical Transport in tropIcal convection (ACTIVE) campaign (Vaughan et al., 2008; Connolly et al., 2013). The ACTIVE campaign was carried out in the Darwin region of northern Australia and aimed to investigate aerosol-cloud interactions in tropical convection. This case, which occurred during the break period of the monsoon season, featured a relatively weaker convective storm over the Tiwi Islands (Fig. 1a). In contrast to the vigorous pre-monsoon convection, the environment was characterized by a dry mid-troposphere and limited convective instability, with a convective available potential energy (CAPE) of approximately 1010 J kg⁻¹ (Fig. 1b). The sounding also suggests a warm cloud base and a relatively deep warm-cloud layer, which likely favor warm-rain formation during the early stage of storm development.

The cloud-resolving simulations are conducted at a horizontal resolution of 1.5 km, using a domain of 900 × 900 grid points centered over the Darwin region (12.4° S, 130.8° E). The vertical grid contains 90 levels, extending up to 40 km, with finer resolution in the lower atmosphere (52 levels below 10 km, 33 below 4 km, and 16 below 1 km). Each simulation starts at 18:00 UTC on 5 February 2006 and runs for 36 h. Boundary and initial conditions are prescribed from the ECMWF Reanalysis version 5 (ERA5) reanalysis data, updated every 6 h. A spin-up period of 6 h is applied, with a model time step of 75 s. In this configuration, the semi-Lagrangian dynamical core permits relatively large time steps. To avoid hydrometeors crossing multiple vertical levels within a single time step, sedimentation is treated using an approach based on Rotstayn (1997), as implemented in the UM (Wilson and Ballard, 1999; Field et al., 2023). To better represent surface features, we incorporate the Shuttle Radar Topography Mission (SRTM; NASA, 2013) dataset as ancillary input. For analysis, we focus on a 1° × 1° subdomain over the Tiwi Islands (indicated by the white box in Fig. 1a), selected to minimize potential boundary effects.

To investigate the sensitivity of mixed-phase convection to secondary ice production (SIP) mechanisms and aerosol conditions, a series of simulations was conducted by combining three prescribed in-cloud droplet number concentrations (Nd) with different SIP configurations (Table 1). A fixed Nd approach was employed to provide controlled droplet-number conditions and to allow the microphysical, radiative, and precipitation responses to SIP to be compared consistently across three different aerosol environments. The rates of the different SIP pathways are linked to Nd through the prescribed droplet field and its influence on warm-rain production, providing the basis for assessing SIP impacts across the three prescribed Nd conditions. Nd was specified at 200, 400, and 800 cm⁻³, corresponding to low, intermediate, and high droplet-number conditions (hereafter referred to as “clean”, “moderate”, and “polluted”, respectively). These values span the range reported during ACTIVE, where cloud droplet number concentrations of approximately 150, 500, and 900 cm⁻³ were associated with the monsoon-break, pre-monsoon, and burning-period environments, respectively (Connolly et al., 2013). The three prescribed values therefore provide a consistent framework for examining how the SIP response varies across plausible aerosol conditions.

For each Nd value, six SIP scenarios were tested: all SIP mechanisms active (allSIP), no SIP process enabled (noSIP), rime-splintering only (RS), and three combined schemes involving RS with one additional mechanism: RS with Mode 1 (RS+M1), RS with Mode 2 (RS+M2), and RS with ice–ice collisional breakup (RS+BR). For each prescribed Nd, the noSIP simulation at the same Nd provides the matched comparison for the corresponding SIP experiments, so that SIP effects are diagnosed with Nd held fixed. The matched comparison between allSIP and noSIP is used to quantify the combined effect of SIP, while the RS, RS+M1, RS+M2, and RS+BR simulations are used to examine the contributions and aerosol dependence of individual mechanisms. RS was retained as the baseline SIP configuration because it is the default SIP process in CASIM. This design enables a systematic assessment of how aerosol conditions influence the impacts of SIP on cloud microphysics, radiation, and precipitation. Of these simulations, allSIP-400 is used as the reference simulation, that is, the complete model configuration, with all SIP mechanisms active, that provides the most observationally defensible representation of the observed storm. Its prescribed Nd is selected by comparing the three complete-model configurations (allSIP-200, allSIP-400 and allSIP-800) with the satellite, radar and aircraft observations (Sect. 3.1). Although the ACTIVE monsoon-break value (∼ 150 cm⁻³; Connolly et al., 2013) lies closer to the clean configuration, allSIP-400 agrees more closely with the observations than allSIP-200, and the uncertainty in Nd is represented by the 200 and 800 cm⁻³ experiments. The remaining simulations are treated as perturbations of the SIP treatment and of Nd (Sect. 3.2–3.4).

The modeled reflectivity fields were evaluated against observations from the C-band polarimetric radar (CPOL), located at Gunn Point (12.25° S, 131.05° E). The CPOL radar provides volumetric scans every 10 min, with a vertical resolution of 0.5 km and horizontal resolution of 2.5 km. In the model, radar reflectivity was diagnosed from the simulated hydrometeor particle size distributions using a forward operator based on a simple Rayleigh scattering assumption, consistent with Field et al. (2023). To facilitate comparison with the model, we used the Statistical Coverage Product (May et al., 2008), which represents the time–height distribution of the fraction of a defined domain exceeding 10 dBZ, a threshold chosen to capture the spatial and temporal characteristics of convective precipitation. The same 10 dBZ threshold was applied to the simulated reflectivity fields for comparison. The analysis domain is indicated in Fig. 1a.

Simulated rainfall accumulations were also compared against CPOL radar–derived precipitation datasets, available at 10 min intervals and 1 km horizontal resolution. Although residual uncertainty may arise from reflectivity calibration and C-band attenuation, CPOL quantitative precipitation estimates generally provide a reliable representation of tropical rainfall (Louf et al., 2019; Jackson et al., 2021). Aircraft observations from the ACTIVE campaign were also considered, with the NERC Dornier sampling the lower troposphere up to approximately 4 km and the Egrett sampling the upper-level anvil outflow at approximately 12–14 km (Vaughan et al., 2008). Available in situ measurements of IWC and ice particle number concentration were confined to upper-level clouds and did not directly sample the mixed-phase region where SIP is expected to be most active. These upper-level observations are compared with the reference simulation in Fig. 4. Radiosonde profiles from Darwin Airport station (12.42° S, 130.89° E) were used to characterize the thermodynamic environment. Additionally, top-of-atmosphere radiation was evaluated using equivalent outgoing longwave radiation (OLR) calculated from Multifunctional Transport Satellite (MTSAT) brightness temperature, with spatial and temporal resolutions of 1.25 km and 30 min, respectively. The observed and simulated OLR are compared in Fig. 2.

On 6 February 2006, the Hector storm developed over the Tiwi Islands during the break period of the monsoon season. MTSAT imagery (Fig. 2a–c) shows shallow cumulus initiating along the sea-breeze front (Connolly et al., 2013), developing rapidly between 07:30 and 08:30 UTC, and reaching maturity by 09:30 UTC. The coldest cloud tops, with OLR below 150 W m⁻², were concentrated over the central island, with an expanding westward anvil. With the simulation time shifted by -1 h to account for the earlier development of the simulated storm, allSIP-400 reproduces the location of the storm and its westward anvil expansion (Fig. 2d–f), although its minimum OLR remains higher than observed. In the domain mean (Fig. 2g), allSIP-400 captures the observed decrease and recovery of OLR. Its mature-stage minimum is lower than observed by ∼ 7 W m⁻², and OLR is underestimated before and after the mature stage by up to ∼ 15 W m⁻².

The CPOL reflectivity CFAD (Fig. 3a) shows echoes exceeding 10 dBZ extending to about 15 km during the mature stage. allSIP-400 reproduces this vertical extent, with frequent echoes above 10 km (Fig. 3b). The simulated rainband structure and heavy rainfall cores (> 100 mm) are broadly consistent with the CPOL accumulation (Fig. 3c, d). The peak domain-mean precipitation rate in allSIP-400 (∼ 1.35 mm h⁻¹) is close in timing and magnitude to the observed peak (Fig. 3e). Accumulated precipitation is underestimated, at ∼ 3.54 mm compared with ∼ 4.8 mm observed (Fig. 3f).

Egrett observations sampled the anvil outflow at about 12–14 km. Within this layer, the ice particle number concentration and IWC of allSIP-400 are broadly consistent with the observations (Fig. 4; noSIP-400 is also shown, to indicate how far the SIP response extends into the sampled layer). These observations do not sample the mixed-phase region between about 5 and 9 km, where SIP is most active. However, the upper-level ice population reflects the cumulative history of the convective ascent, including the supercooled liquid lofted to homogeneous-freezing levels, which is itself influenced by SIP lower in the cloud (Phillips and Patade, 2022). The other two complete-model configurations agree less well with these observations, which is the basis for choosing allSIP-400 as the reference. allSIP-200 produces a higher minimum OLR (∼ 165 W m⁻²), a lower peak precipitation rate (∼ 0.95 mm h⁻¹) and a smaller accumulation (∼ 3.25 mm). allSIP-800 produces a lower peak precipitation rate (∼ 0.94 mm h⁻¹), a smaller accumulation (∼ 3.11 mm), and a more fragmented anvil and rainfall pattern than observed.

Overall, the reference simulation reproduces the main observed features of the storm: the timing and location of convection, the depth of the radar echoes, the organisation and peak intensity of rainfall, and the upper-level ice properties. The main remaining differences are the OLR minimum, the OLR before and after the mature stage, the accumulated rainfall, and the absence of in situ observations in the mixed-phase layer. These differences provide the context for interpreting the sensitivity experiments that follow.

Figure 5 shows simulated OLR from the allSIP simulations at the three prescribed Nd values (allSIP-200, allSIP-400 and allSIP-800), with the corresponding noSIP fields in Fig. S1 in the Supplement. The underlying microphysical mechanisms are discussed in Sect. 3.4.1.

The impacts of SIP on radiation were first examined by comparing the spatial distribution of OLR between each allSIP simulation and its corresponding noSIP case at 07:30 UTC (Fig. S1). For Nd = 200 cm⁻³, allSIP-200 shows a weaker OLR reduction relative to noSIP-200 (Figs. 5b, S1b), with a minimum OLR of ∼ 165 W m⁻² and a smaller increase in the low-OLR coverage (∼ 10 %–15 %). For Nd = 400 cm⁻³, allSIP-400 produces a minimum OLR of ∼ 150 W m⁻² (Fig. 5e), about 25 W m⁻² lower than noSIP-400 (∼ 175 W m⁻²), and expands the area with OLR < 200 W m⁻² by around 25 %–30 % (Fig. S1e). At Nd = 800 cm⁻³, allSIP-800 still exhibits reduced OLR relative to noSIP-800, but the low-OLR region becomes more fragmented (Figs. 5h, S1h).

The CFADs (contoured frequency by altitude diagrams) of reflectivity (> 10 dBZ), as shown in Fig. 6, further link these radiative patterns to storm structure. For Nd = 200 cm⁻³, the difference between allSIP-200 and noSIP-200 is smaller and mainly confined to mid-levels (Fig. 6a, b), consistent with the weaker radiative impact (Fig. 5a–c). For the Nd = 400 cm⁻³ pair, allSIP-400 sustains a deeper vertical extent of reflectivity > 10 dBZ above 10 km than noSIP-400 (Fig. 6c, d), consistent with the expanded low-OLR region. At Nd = 800 cm⁻³, reflectivity > 10 dBZ reaches upper levels (Fig. 6e, f), but the echoes are more vertically dispersed and occur with lower frequency at each height than in the Nd = 400 cm⁻³ pair. This suggests a deeper yet optically thinner distribution of hydrometeors, which may explain the less concentrated low-OLR region (Fig. 5g–i). Overall, the intermediate Nd simulations produce the most persistent upper-level reflectivity, particularly when all SIP mechanisms are included. The response to SIP is weaker under low Nd, while the high Nd simulations exhibit a more vertically dispersed reflectivity structure.

The temporal evolution of domain-averaged OLR is shown in Fig. 7a, and the SIP-induced changes are summarized in Table 2. In all simulations, OLR decreases after approximately 02:00 UTC as convection develops, reaches a minimum during the mature stage, and recovers towards the end of the analysis period. At Nd = 400 cm⁻³, allSIP-400 reaches a substantially lower minimum than noSIP-400. The SIP-induced change is smaller at Nd = 800 cm⁻³, and small and of opposite sign at Nd = 200 cm⁻³. Quantitatively, the strongest reduction occurs at Nd = 400 cm⁻³, where the allSIP case drops below 250 W m⁻² and the mean allSIP–noSIP difference reaches -6.86 W m⁻² (-2.5 %). At Nd = 200 cm⁻³, allSIP-200 produces a slight increase of +0.61 W m⁻² (+0.2 %) relative to noSIP-200, whereas at Nd = 800 cm⁻³, allSIP-800 produces a decrease of -3.34 W m⁻² (-1.2 %) relative to noSIP-800. The reductions under the intermediate and high Nd conditions are consistent with stronger anvil development and enhanced upper-level ice content.

In contrast, outgoing shortwave radiation (OSR; Fig. 7b) increases after 02:00 UTC and peaks around 06:00 UTC, but its response to SIP is more variable. At Nd = 400 cm⁻³, SIP enhances OSR relative to noSIP, with peak values exceeding 500 W m⁻² compared to ∼ 480 W m⁻², and a mean increase of +7.95 W m⁻² (+2.0 %). At Nd = 200 cm⁻³, however, SIP reduces OSR by -6.55 W m⁻² (-1.8 %), while at Nd = 800 cm⁻³ the reduction is smaller at -4.05 W m⁻² (-0.9 %). This contrast suggests that SIP has a robust and consistent impact on longwave radiation by maintaining colder, more extensive anvils, whereas its shortwave effect is weaker and dependent on CCN concentration. In addition, OSR differences may be further influenced by variations in cloud phase, particle size, and the timing of solar insolation (Yin and Porporato, 2017).

The spatial distribution of accumulated precipitation from 04:00–10:00 UTC (Fig. 8) shows that SIP affects both the intensity and organization of rainfall, with responses depending on CCN concentrations. At Nd = 200 cm⁻³ (Fig. 8a, b), allSIP-200 slightly increases the number of rainfall cores relative to noSIP-200 but does not substantially change the overall coverage, indicating a weaker response in this case. allSIP-400 produces a more organized rainband with more heavy rainfall cores (> 100 mm) than noSIP-400 (Fig. 8c, d), indicating stronger organization when SIP is included. Under Nd = 800 cm⁻³ (Fig. 8e–f), allSIP-800 broadens the precipitation area relative to noSIP-800, but the rainfall pattern is more fragmented than in allSIP-400, resembling the more dispersed pattern in the OLR distribution (Fig. 5h–i).

Domain-averaged precipitation time series (Fig. 9) further quantify these effects. The peak precipitation rate in allSIP-400 (∼ 1.35 mm h⁻¹) is notably higher than in noSIP-400 (∼ 0.80 mm h⁻¹; Fig. 9a). In the Nd = 400 cm⁻³ pair, SIP enhances the production of ice-phase particles aloft, accelerating mixed-phase processes (Sun et al., 2025). This promotes melting and sustains a robust rain region, as shown in Fig. 8c. For Nd = 200 cm⁻³, allSIP-200 (∼ 0.95 mm h⁻¹) produces a lower peak rate than noSIP-200 (∼ 1.20 mm h⁻¹). In clean conditions, early precipitation formation likely involves a stronger contribution from warm-rain processes. SIP accelerates glaciation during the early storm stage, shifting part of the condensate from liquid to ice phase and likely reducing the liquid water available for warm-rain collision–coalescence. This is consistent with the lower peak precipitation rates under clean conditions. The precipitation duration remains comparable, resulting in similar accumulated totals between allSIP-200 (∼ 3.25 mm) and noSIP-200 (∼ 3.27 mm; Fig. 9b). For Nd = 400 cm⁻³, allSIP-400 (∼ 3.54 mm) yields a higher accumulated total than noSIP-400 (∼ 2.81 mm, +25 %; Fig. 9b), consistent with its stronger precipitation peak. At Nd = 800 cm⁻³, high aerosol loading delays precipitation onset due to smaller cloud droplet sizes and reduced collision–coalescence efficiency. In this case, SIP increases ice water content, leading to greater late-stage precipitation in allSIP-800 than in noSIP-800 (∼ 2.30 mm). However, the total accumulation in allSIP-800 (∼ 3.11 mm) remains lower than in allSIP-400 (∼ 3.54 mm), as the limited warm-rain formation under high droplet number concentrations cannot be fully compensated by the enhanced ice-phase process. The simulated domain totals span approximately 50 % across the combinations of SIP and Nd (2.30–3.54 mm), demonstrating the sensitivity of precipitation accumulation to both prescribed droplet number and SIP configuration. This provides a clear motivation to represent and constrain these processes more accurately in precipitation prediction.

Overall, these results suggest that SIP can intensify rainfall and produce a more concentrated precipitation pattern, with the strongest response in allSIP-400 relative to noSIP-400. At Nd = 200 cm⁻³, large cloud droplets allow warm-rain processes to occur efficiently even without SIP, so the additional ice-phase pathway has only a small impact on precipitation. At Nd = 800 cm⁻³, enhanced cloud droplet numbers delay warm-rain processes and lead to more dispersed precipitation; allSIP-800 increases totals relative to noSIP-800 but does not reach the allSIP-400 total.

The experiments show that SIP alters the radiative and precipitation characteristics of the 6 February 2006 storm, but its impacts depend on CCN concentrations. Under moderate aerosol loading (Nd = 400 cm⁻³), SIP produces the largest reduction in outgoing longwave radiation (OLR), expansion of low-OLR anvil areas, and the strongest enhancement of organized precipitation. By contrast, SIP effects are weaker under clean (Nd = 200 cm⁻³) and polluted conditions (Nd = 800 cm⁻³), and the OSR response is overall modest and varies in sign with CCN. These results highlight a non-linear SIP–CCN influence on storm structure and radiative forcing.

Agreement between the reference simulation and the available observations (Sect. 3.1) does not exclude compensating errors. The observations constrain the storm-scale radiative, reflectivity and precipitation evolution and the upper-level ice field, but not the mixed-phase layer in which SIP operates. Offsetting biases in, for example, primary ice nucleation, raindrop freezing or aggregation therefore cannot be ruled out. The SIP responses diagnosed from the matched experiments should thus be interpreted as model responses relative to an observationally consistent reference state, rather than as validated estimates of SIP in the observed storm.

The underlying mechanisms can be linked to the modulation of cloud microphysics by CCN. At low CCN, large droplets form and precipitation develops early through warm-rain processes, depleting supercooled liquid water and leaving little available for SIP-induced ice growth. As a result, OLR reduction and precipitation changes are limited. At high CCN, numerous small droplets delay warm-rain processes and sustain supercooled liquid water at higher altitudes. The reduced collision–coalescence efficiency leads to the lowest accumulated rainfall in noSIP-800 among the three CCN levels (Fig. 9b). When SIP is activated, this additional supercooled liquid water provides a source for secondary ice production. In this case, the extra ice particles enhance riming and graupel formation, and subsequent melting increases surface rainfall, so allSIP-800 yields more rainfall than noSIP-800. The SIP-related increase remains smaller than at Nd = 400 cm⁻³, because at moderate CCN the mixed-phase layer contains a more favorable combination of supercooled liquid water, raindrops, and large rimed particles within the temperature ranges where the SIP mechanisms are active, allowing more efficient ice growth and melt out. At moderate CCN, the SIP mechanisms therefore act together most effectively, leading to stronger upper-level ice growth, thicker and more organized anvils, a larger reduction in OLR (Figs. 5, 7), and enhanced rainfall (Figs. 8, 9). Previous studies also suggest that SIP impacts on precipitation can depend strongly on the parameterization used, and may increase ice particle number while suppressing localized or heavy precipitation (e.g., Dedekind et al., 2021; Grzegorczyk et al., 2025b). Thus, SIP partly compensates for warm rain suppression under polluted conditions, but its strongest combined radiative and precipitation impact occurs at moderate CCN, where the microphysical environment most strongly supports secondary ice production and conversion to surface rainfall.

This non-linear response should also be interpreted in the context of the weak-CAPE environment of this case (CAPE ≈ 1010 J kg⁻¹; Fig. 1b). With limited instability, the balance between warm-rain and ice-phase processes could be more sensitive to CCN perturbations. In contrast, our earlier pre-monsoon case (Sun et al., 2025), characterized by higher CAPE and stronger convection, showed larger reductions in OLR, stronger increases in OSR, and suppressed precipitation in allSIP compared to noSIP. This enhanced radiative signal arose from deeper and more persistent anvils formed under high CAPE. Vigorous updrafts transported large amounts of hydrometeors into the upper troposphere, strengthening both longwave cooling and shortwave reflection (Feng et al., 2011; Hartmann and Berry, 2017). At the same time, stronger glaciation converted more supercooled liquid water into ice particles, enhancing the formation of precipitation particles in the upper levels. However, much of this ice was retained in anvils or sublimates/melts before reaching the ground, leading to reduced surface rainfall despite greater ice production (Han et al., 2024; Grzegorczyk et al., 2025a). By contrast, in the weak-CAPE 6 February 2006 storm, OLR responses remain robust but OSR is only weakly affected (Figs. 5, 7). SIP enhances OSR at Nd = 400 cm⁻³, but this effect reverses at Nd = 200 cm⁻³ and nearly vanishes at Nd = 800 cm⁻³. This weak and non-monotonic OSR response likely reflects the combined influences of changes in anvil coverage, optical depth and particle size, and the timing of SIP-induced anvils relative to the maximum solar insolation. As a result, shortwave fluxes exhibit the SIP signal much less clearly than longwave cooling in this case. This behavior is consistent with Finney et al. (2025), who showed across 12 convective cases that changes in cloud droplet number, ice nucleating particles, and secondary ice configurations can modify anvil albedo and high cloud area, but that these shortwave responses are generally small compared with the case to case spread, highlighting the strong role of environmental variability and the need for caution when interpreting single case results. Meanwhile, precipitation increases in allSIP relative to noSIP (Figs. 8, 9). Under weak instability, cloud depth and anvil extent are limited compared to the high-CAPE case. SIP generates additional ice particles that likely promote riming and subsequent melting, which may provide an extra source of surface rainfall. This may help explain why SIP increases precipitation in the 6 February 2006 case, whereas the pre-monsoon case showed the opposite response.

These results underscore the importance of considering environmental conditions, such as CAPE, when assessing SIP-CCN interactions in tropical convection. Across both radiation and precipitation responses, the strongest enhancement consistently occurs under moderate CCN, indicating that SIP plays an important role in amplifying aerosol effects on tropical convection, particularly under weakly unstable environments.

The aerosol response of secondary ice production is mechanism dependent and non-linear. RSM1 is suppressed under polluted conditions (Nd = 800 cm⁻³), RSM2 is inhibited under clean conditions (Nd = 200 cm⁻³), and RS together with RSBR show limited CCN sensitivity. The combined effect produces the strongest signals at moderate CCN (Nd = 400 cm⁻³), including enhanced upper-level ice, higher peak updrafts, and the largest reduction of outgoing longwave radiation (Figs. 12–14; Table 3).

For RSM1, the diagnostics indicate a limitation in the warm-branch supply. Mode 1 depends on the freezing of supercooled raindrops and their collisions with small ice particles. Under high CCN, cloud droplets are more numerous and smaller, collision–coalescence proceeds more slowly, and the delivery of raindrops into the 0 to -20 °C layer is reduced. Consistently, RSM1 exhibits less ice number between 7 and 9 km and the smallest ΔIWC at Nd = 800 cm⁻³, while crystal sizes remain nearly unchanged across CCN levels (Fig. 13; Table 3).

For RSM2, Mode 2 depends on encounters between raindrops and large rimed ice particles. In clean conditions, warm rain forms and precipitation begins earlier, which limits the overlap of supercooled liquid and large ice in the 7–9 km layer. The diagnostics show the lowest upper-level ice number and the smallest graupel radii at Nd = 200 cm⁻³, whereas moderate to high CCN maintains larger graupel and stronger upper-level ice production, with ΔIWC maximized at Nd = 400 cm⁻³ (Fig. 14; Table 3).

Our simulations further demonstrate that the allSIP configuration produces the strongest enhancement at moderate CCN (Nd = 400 cm⁻³). This non-linear behavior reflects the need for supercooled raindrops and sufficiently large rimed ice to co-exist within the mixed-phase layer. Under clean and polluted conditions, one of the droplet-shattering pathways is limited: Mode 1 at high CCN due to the reduced supply of supercooled raindrops, and Mode 2 at low CCN because collisions with large rimed ice particles are less frequent in the mixed-phase layer. At moderate CCN, supercooled liquid is available, and raindrops occur with larger rimed ice, allowing Mode 1 and Mode 2 to operate concurrently on top of the relatively CCN-insensitive RS and BR processes. As a result, the combined SIP contribution is largest at intermediate aerosol loading, yielding the greatest ΔIWC and clear dynamical and radiative responses (Figs. 7, 9, 10). This peak at intermediate aerosol loading is consistent with previous observational and modeling studies (Rosenfeld et al., 2008; Fan et al., 2016), and our results identify SIP as a process pathway that helps produce this non-linear CCN dependence. This response likely arises because the amounts of supercooled liquid water and rimed ice, which are controlled by Nd, do not change in a simple way across the temperature ranges where each SIP mechanism is most active. Different SIP parameterizations, with other temperature ranges or production rates, could change this relationship and in some cases produce a more monotonic response to Nd. However, the general result that aerosol control of liquid and ice in the mixed-phase layer affects the strength of SIP is expected to remain.

Beyond this CCN dependence, our results also indicate that droplet shattering mechanisms (Mode 1 and Mode 2) contribute more strongly to ice production than ice–ice collisional breakup (BR). This suggests that interactions between raindrops and ice particles can play an important role in sustaining secondary ice production under the simulated conditions. However, Grzegorczyk et al. (2025c) found that BR and RS were the most effective SIP mechanisms in reproducing ice numbers in deep convection, while droplet shattering played a minor role. Previous tagging-tracer analyses in continental deep convection also found a dominant contribution from BR to total ice number, whereas droplet shattering contributed comparatively less (e.g., Waman et al., 2022). Part of this difference may arise from differences in the microphysics schemes, as well as from the environmental conditions in our simulations, where graupel sizes and collision energies are relatively modest (Sun et al., 2025), limiting the efficiency of BR. One specific example is the treatment of raindrop freezing in the two modeling systems. UM-CASIM employs a formulation based on Bigg (1953) with modified parameter values, which is represented separately from the heterogeneous ice nucleation scheme and is not explicitly linked to INP activity. This treatment affects the number of freezing raindrops and graupel particles formed in the mixed phase region where droplet shattering can occur. This differs from recent approaches in which raindrop freezing is related to INP concentrations, such as Waman et al. (2022). In comparison, the bin microphysics model DESCAM uses the original Bigg (1953) scheme (Grzegorczyk et al., 2025c), which tends to produce fewer freezing raindrops (e.g., Morrison et al., 2005; Fan et al., 2009; Sullivan et al., 2018) and thus fewer opportunities for droplet shattering, so that BR and RS account for a larger share of the simulated secondary ice. A further uncertainty concerns the representation of heterogeneous ice nucleation. In the present study, this follows Cooper (1986) and may affect the relative enhancement attributed to SIP. The magnitude of the simulated SIP contribution should therefore be interpreted with caution. Environmental differences, such as CAPE and liquid water content, may also contribute to the SIP responses. These results highlight the uncertainties in quantifying the relative importance of SIP mechanisms. Consistent parameterizations and further observational constraints are still needed to separate the contributions of droplet shattering and collisional breakup in deep convective clouds.