We use the System for Atmospheric Modeling (SAM) cloud-resolving model version 6.10.9. SAM uses a 1.5-order Smagorinsky-type closure scheme to represent subgrid-scale turbulence and subgrid-scale motions. The time step in SAM is adaptive and set based on the Courant–Friedrichs–Lewy criterion, which typically leads to a time step of about 4.5 s for simulations at the horizontal grid spacing of 1 km. Radiative fluxes and heating rates are computed with the Rapid Radiative Transfer Model for GCMs RRTMG;, which is called every 3 min. Cloud and precipitation processes use the Predicted Particle Property P3; microphysical scheme version 3.1.14.

We use a tropical channel setup that is wide in the zonal direction (3888 km) and narrow in the meridional direction (36 km) with double-periodic boundary conditions. The prescribed sea surface temperatures vary sinusoidally in the zonal direction between 24 °C at the domain edge and 28 °C in the middle of the domain. Convection develops over the warmer sea surface temperatures (SSTs) and gives rise to a large-scale overturning circulation in the zonal direction reminiscent of the Walker circulation. Such a “mock-Walker” circulation setup is therefore appropriate for studying the interplay between convection, clouds, and radiation in the tropics . To accurately represent processes in the TTL, we impose a mean large-scale vertical wind based on observations by . Zonal winds increase linearly with altitude from 0 m s⁻¹ at the surface to 5 m s⁻¹ at altitudes above 14 km. Our “mock-Walker” setup follows the one described in more detail by . The simulations are performed for 30 d, the last 20 of which are used for analysis.

For a fairer comparison with model data, which simulates a climate comparable to that of the tropical Pacific, we use airborne data from three campaigns in the tropical western Pacific: the Airborne Tropical TRopopause EXperiment ATTREX;; the Pacific Oxidants, Sulfur, Ice, Dehydration, and cONvection experiment POSIDON;; and the CONvective TRansport of Active Species in the Tropics (CONTRAST) experiment .

Vertical wind data are only used from the tropical western Pacific flights of the ATTREX and POSIDON campaigns. The data are sampled using NASA's Meteorological Measurement System (MMS) instrument , which has a time resolution of 20 Hz. Vertical wind data are corrected by detrending and removing the mean of each flight leg. More details on the processing are described in .

In situ data for ice cloud properties for the three campaigns are taken from the dataset. The measurement resolution is 1 Hz, with an aircraft velocity of 170 m s⁻¹ in POSIDON and ATTREX and 200 m s⁻¹ in the CONTRAST campaign. The lower limit of detectable particle concentration in a given sampling time depends on the aircraft velocity and sampling area of the respective instrument . This limits the measurements for CONTRAST of ice crystals at ICNC < 0.01 cm⁻³ and mean mass ice radii smaller than 35 µm. For consistency, model output under such conditions is not included in the calculation of model performance. Due to the slower aircraft speed and different instrumentation, there is no such limitation for POSIDON and ATTREX. The dataset contains limited measurements within or very near active deep convection (see Fig. ).

We use satellite retrievals from the years 2007–2010 for the tropical western Pacific (TWP) (20° S–20° N, 145–180° E), an almost exclusively ocean-covered area characterized by persistent deep convection throughout the year. In the selected years, all the satellite products used have complete data coverage.

We compare the power spectral density of vertical wind between SAM and four global storm-resolving models (GSRMs) in Sect. . GSRM simulations were run with the Nonhydrostatic ICosahedral Atmospheric Model (NICAM), Global System for Atmospheric Modeling (gSAM), Finite Volume Cubed-Sphere Dynamical Core (FV3), and Icosahedral Nonhydrostatic Weather and Climate Model (ICON), as part of the DYAMOND-1 experiment . The GSRM simulations have horizontal grid spacings of 2.5–5 km and vertical grid spacings of 400–500 m. Vertical winds are analyzed within a square swath in the tropical western Pacific (4° S–4° N, 144–152° E), a region overlapping the aircraft measurements with a climatology comparable to the SAM simulations. Vertical winds are analyzed for model levels closest to 14.2 km, the most common height for ATTREX level-leg sampling.

This section first outlines the step-by-step changes implemented in the ice microphysics scheme. We focus on changes under cirrus conditions, limiting the analysis to temperatures colder than -40 °C and ice water contents larger than 10⁻⁵ g m⁻³, which is close to the detectability threshold of the aircraft observations. The data are first presented in the ICNC–ice crystal mean mass radius (from now on: ice number–radius) space, which provides an intuitive aggregated perspective on ice cloud properties .

Aircraft observations in Fig. a show a peak ICNC between 1 and 10⁻² cm⁻³ for particles with mean mass radii smaller than 30 µm. For concentrations smaller than 10⁻¹ cm⁻³, the observed particle size often exceeds 50 µm. We note that due to retrieval limits, there are limited measurements available for ice radii smaller than 35 µm at number concentrations below about 10-2 cm⁻³ (see the Methods section).

The standard version of the SAM model coupled with the P3 scheme (Fig. b, Sect. ) is strongly biased compared to observations. Most notably, the model drastically underestimates ICNC as it lacks concentrations larger than 3 × 10⁻² cm⁻³. We resolve a large part of the bias by implementing three key changes to the ice microphysics.

We first relax the maximum ICNC limit from 5 × 10⁻² to 20 cm⁻³ (Fig. c, Sect. ). This particularly improves the representation of particles larger than 30 µm, reducing the total variation distance metric a 2D analog to the root mean square error from 0.52 to 0.35. However, the model still strongly underestimates the number densities of small ice crystals in the top left quadrant, indicating errors in parameterizing ice formation under cirrus conditions.

The second key modification addresses the deposition freezing parameterization, which was incorrectly active at temperatures both below and above the homogeneous freezing threshold. Originally calibrated for temperatures warmer than -25 °C (Fig. ), this parameterization extended far beyond its intended range. We restrict it to T > -37 °C and introduce a scheme to account for competition between homogeneous and heterogeneous nucleation at T < -37 °C (Sect. , ). We also include a number of small changes to the microphysics, described in Sect. .

All these changes combined cut the microphysical bias in half, improving the representation of both small and large particles (Fig. d). Nevertheless, some biases remain. The model continues to underestimate ICNC for small ice crystals and overestimate ICNC for larger crystals. These remaining biases largely originate from persistent challenges in representing ice microphysics (see Sect. ) and from vertical wind variance that is too low in the model (see Sect. ).

To provide an alternative perspective, we examine the model's performance by sorting results by temperature. The exponential decrease in IWC with decreasing temperature has improved compared to the standard model version and is well represented by the model (Fig. a–c). Both observations and simulations indicate that a substantial portion of IWC must originate at warmer temperatures and be transported upward by convection. This is evident from IWC values that exceed the maximum expected from local ice nucleation and subsequent depositional growth constrained by the Clausius–Clapeyron relationship. The blue lines in Fig. a and b illustrate these theoretical limits, derived in Appendix . They represent the maximum IWC that can form locally through ice nucleation under supersaturated conditions with respect to ice, followed by vapor deposition until saturation is reached.

Observed ICNC shows large variability but only a weak dependence on temperature (Fig. d). The model reproduces the observed median and spread, improving the agreement substantially compared to the standard model version (only the gray median is shown). Notably, the model still misses some of the (rather infrequent) high ICNC data points at cold temperatures.

Similarly to IWC, the mean mass radius decreases with temperature, with medians ranging from 80 µm at -40 °C to 15 µm at -80 °C. The model simulates particles that are too small at warmer temperatures (between -65 and -40 °C) and comparable to observations for T < -65 °C. Notably, the spread in simulated particle size is narrower than observed, possibly due to the single-mode description of ice microphysics that is too simple (see Sect. ).

In summary, the temperature-sorted model results offer a complementary perspective on cirrus cloud properties, showing generally good agreement with observations, while also highlighting persistent biases in ICNC at colder temperatures and particle size at warmer temperatures. Nevertheless, this agreement represents a substantial improvement compared with the earlier model version (Fig. b), emphasizing processes where further refinement is still needed. The addition of passive tracers discussed in Sect.  helps pinpoint processes requiring further refinement, particularly the representation of in situ ice nucleation at cold temperatures.

To link cirrus cloud properties to the TOA radiative budget, we take an integrated perspective and categorize cloud occurrence frequency and radiative effects based on IWP. We neglect infrequent scenes with low clouds underlying thin cirrus, where low clouds are expected to dominate the CRE signal. Figure a compares the simulated IWP distribution with four satellite-derived datasets, showing that the improved model generally falls within or near the observed range, which also includes deep convective cores typically associated with IWP > 10³ g m⁻², based on. We note that this association is statistical and that individual convective cores are unlikely to span a wide range of IWP values. The apparent overrepresentation of the thinnest cirrus in the model and the corresponding slight underestimation of other cloud types in the normalized PDF should not be interpreted as a bias, as satellite retrievals are known to underestimate the thinnest cirrus clouds . Thin cirrus with IWP < 1 g m⁻² make up 52 % of all simulated cirrus but contribute only 6 % to the LW CRE and 5 % to the SW CRE, indicating that their direct radiative impact at the TOA is modest despite their frequency. Focusing on clouds with IWP > 1 g m⁻², which have a greater radiative influence, improves the model's agreement with satellite observations (Fig. ).

Figure b shows the averaged CRE for high clouds in each IWP bin. The most frequent anvils according to satellite datasets occur at an IWP of 3–30 g m⁻², which corresponds to anvil clouds of intermediate optical depth (1–3) that yield a net positive, LW-dominated CRE . For thicker clouds (IWP > 100 g m⁻²), both model and satellite data show a dominance of SW CRE, with net CRE exceeding -100 W m⁻² for the thickest anvils. However, the model overestimates the SW CRE, a bias partially resolved by averaging the output onto coarser grid scales, similar to the CERES pixel data resolution (36 × 36 km grid boxes). Additionally, a recently identified bug in the ice optics parameterization likely increased the optical depth and radiative effects per unit IWP by ∼ 15 %, but this issue could not be addressed in the current simulations. This SW bias becomes even more apparent when comparing CERES CRE retrieved during daytime satellite overpasses with model-simulated CRE between 13:00 and 14:00 LT (Fig. b).

Determining which type of tropical cirrus is radiatively most important is not straightforward. Are the less frequent but thick deep convective cores and fresh anvils, which have a strong influence on both SW and LW CRE, the most dominant, or do the more widespread anvil clouds of intermediate thickness dominate? Figure c provides an answer by scaling the CRE in a given IWP bin by that bin's frequency of occurrence: the radiatively most dominant clouds have IWP between 1 and 3000 g m⁻², similar to results by . These include anvil clouds with optical depths greater than approximately 1. Nonetheless, thinner cirrus (IWP 0.3–10 g m⁻², Fig. c) also contribute meaningfully to the net CRE, underscoring the importance of studying not only the thickest tropical cirrus but also their continued evolution until they reach an IWP of ∼ 0.1 g m⁻² (optical depth ∼ 0.005). Additionally, the response of thin cirrus to global warming remains highly uncertain and deserves further investigation.

We illustrate the utility of passive tracers in disentangling convective and in situ contributions to cirrus formation using a case study and statistical estimates of cirrus origin in the SAM model that includes all model improvements (Sect. –) and the two passive tracers (Sect. ). Our analysis begins with the model snapshot in Fig. , while a Video Supplement shows the system's evolution over 2 d. The figure depicts a snapshot of a multicore mesoscale convective system with an anvil cloud shield extending approximately 500 km. This system includes a fresh, thick precipitating anvil cloud that evolves into a thinner, aged anvil cloud. Above this, a thin cirrus layer spans altitudes of 15–16 km. Identifying the origin of such thin cirrus from a single model output time step is challenging. Although thin TTL clouds are typically formed in situ , they might also be remnants of TTL-penetrating deep convection.

Passive tracers resolve this uncertainty. The detrainment tracer highlights regions of active deep convection and thick anvil clouds with times since detrainment typically less than 10 h and reveals that the overlying thin cirrus resides predominantly in air undisturbed by convection for at least 24 h (Fig. b). The in situ nucleation tracer confirms this, showing that much of the thin cirrus originates from recent ice nucleation events (within the last 5 h; Fig. c).

Interestingly, the nucleation tracer also indicates that some parts of the anvils experience ongoing ice nucleation. These events, driven by convective gravity waves or within-anvil updrafts, align with previous observational studies . However, their broader significance for tropical cirrus remains uncertain .

To better characterize the microphysical origin of cirrus, we classify them into three categories: in situ, anvil, and dual-origin. Anvil cirrus are defined as clouds where the time since detrainment is shorter than the time since in situ nucleation. In situ cirrus are those that have not experienced detrainment for at least 24 h and where the time since nucleation is shorter than the time since detrainment.

Dual-origin cirrus are clouds detrained within the last 24 h that have a shorter time since in situ nucleation compared to time after detrainment. These typically form within or near anvils or in cirrus clouds close to active convection (Fig. d). To correct for classification biases due to ice crystal sedimentation impacting older anvils and in situ cirrus, clouds that were initially classified as in situ cirrus but located below dual-origin cirrus are reassigned to the dual-origin category. Although dual-origin cirrus are influenced by in situ nucleation, their total ice mass and number are still dominated by convective outflow (not shown).

Our tracer approach confirms previous findings highlighting IWC as a good predictor of cirrus origin . High-IWC tropical cirrus are thought to be of convective origin, while low-IWC cirrus, particularly those at cold temperatures, are more likely of in situ origin . Using a two-dimensional IWC–temperature space, we find that in situ cirrus dominate only for the coldest low-IWC tropical cirrus (IWC < 10⁻³ g m⁻³, T < -65 °C), while high-IWC cirrus and most cirrus at warmer temperatures are of convective origin (Fig. a). In situ contributions range from 10 % (T > -50 °C) to more than 70 % (T < -70 °C). Additionally, we repeat the analysis using ICNC, which leads to less distinct patterns (Fig. b). While the highest-ICNC bins are clearly associated with anvil cirrus, in situ cirrus fractions remain steady at ∼ 10–40 % for ICNC < 0.03 cm⁻³, increasing to over 60 % only at temperatures colder than -65 °C.

Vertically integrated cirrus origin analysis in Fig. c shows that in situ cirrus dominate at IWP < 0.4 g m⁻² and peak at IWP values between 0.1 and 0.5 g m⁻². At IWP > 10 g m⁻², in situ nucleation is highly unlikely due to limited vapor availability and slow depositional growth in the TTL. Clouds with IWPs of 1–50 g m⁻² often consist of contributions from both in situ and convective sources, suggesting that in situ nucleation plays a role in sustaining and prolonging the lifetime of aged anvil clouds. This analysis, together with our previous findings (Sect. 3.2), provides more information about the evolution of tropical cirrus. These can be split into two separate evolution pathways, as qualitatively depicted in Fig. a. Thick anvils originating in high-IWP deep convective towers spread and thin towards IWP ∼ 1 g m⁻², while thinner cirrus (IWP < 1 g m⁻²) predominantly form via in situ ice nucleation. And while ice nucleation is very frequent and may play an important role in prolonging anvil lifetime , dual-origin cirrus generally preserve convective-like properties and could, for simplicity, be merged with the anvil category.

Finally, by multiplying the cloud occurrence frequency in Fig. c with the IWP-binned CRE in Fig. b we estimate the radiative contribution of each of the three cirrus types, summarized in Table . Anvil clouds are radiatively most important, accounting for more than 50 % of both LW and SW CRE of all tropical cirrus. They dominate in the high IWP range, where SW CRE is stronger than LW CRE, leading to a net CRE of about -8 W m⁻². Dual-origin cirrus are aged anvils that dominate moderate IWP. They are still radiatively very important, leading to 40 % of the total LW and 30 % of the total SW CRE. Finally, in situ cirrus radiative contribution accounts only for 7 % of the total LW and SW CRE. The estimate of their SW CRE may be biased high due to the occurrence of lower-lying anvils or dual-origin clouds, likely leading to an overestimation of the SW CRE. A more detailed radiative analysis accounting for cloud overlap see, e.g., is left for future work. We also note that the model-derived in situ cirrus CREs are comparable to satellite-derived CRE for TTL cirrus .

The two passive tracers provide new insights into the distinct microphysical evolution pathways of detrained anvils and in situ cirrus. This perspective helps differentiate and track the life cycle of numerous cirrus clouds formed in our simulation, offering a more comprehensive understanding of cirrus evolution compared to a snapshot perspective. For detrained anvils, deep convection initially injects high concentrations (ICNC > 0.1 cm⁻³) of ice crystals spanning a broad size range, including relatively large particles (Fig. b). The tracers allow for the study of how these crystals evolve over time: within the first 1–3 h following detrainment, there is a rapid decrease in both number concentration and size due to sedimentation, marking the most dynamic phase of anvil evolution (compare Fig. b and c). Beyond this period, the evolution slows, with ice mass and number decreasing gradually as sedimentation and sublimation deplete the ice crystals (compare Fig. c and d). Figure e shows properties of clouds classified as in situ cirrus in Figs. d and .

In contrast, freshly nucleated in situ cirrus ice crystals form at smaller sizes and intermediate concentrations (ICNC: 0.02–0.2 cm⁻³). Homogeneous nucleation events occasionally spike these concentrations but remain transient and thus are barely visible in our frequency figure. Since most of the in situ crystals are smaller than the freshly detrained ones (Fig. g–i), sublimation may be a more important ice crystal sink than sedimentation. This may imply a greater sensitivity to atmospheric thermodynamic conditions, such as temperature and supersaturation fluctuations. Over time, in situ cirrus tend to lose ice number and size, approaching microphysical properties that often overlap with those of aged anvil cirrus. In summary, while both in situ and detrained cirrus retain distinct properties in the first 3–5 h, they become harder to distinguish in the later stages of their evolution.

All results presented so far are based on simulations with a horizontal grid spacing of 1 km, which is sufficient to represent anvil cloud evolution and their influence on mesoscale circulation . To assess the impact of model grid spacing, we include additional simulations at coarser (4 km) and finer (250 m) grid spacings. The 4 km grid spacing aligns with current GSRMs, while the 250 m grid spacing approaches the level of convergence for deep convective updraft strength in idealized and realistic tropical convection setups .

The sensitivity of simulated ice properties to horizontal grid spacing is most pronounced for ICNC, which increases by nearly an order of magnitude from the 4 km to the 250 m simulation, a trend also evident in the 90th percentiles (Fig. b). IWC exhibits grid spacing dependence only at temperatures warmer than -60 °C, with higher values at finer grid spacings (Fig. a). In contrast, ice crystal radius shows little sensitivity to grid spacing (Fig. c). Moreover, SAM underestimates the variability in ice crystal size across all grid spacings and fails to reproduce the 90th percentile of observed particle size.

Overall, simulated microphysical properties at all grid spacings fall within the range of observations. However, the 4 km simulations underestimate both IWC and ICNC across most temperatures, but these biases are gone in the 1 km simulation. The ice number–radius perspective provides a clearer view of these improvements, particularly for the low ICNC bias of ice crystals smaller than 30 µm (Fig. ). Increased model resolution helps identify and understand the sources of model bias. The following section explores vertical wind variability, a likely contributor to these biases and a key factor in the reduced model bias observed for finer horizontal grid spacing.

Ice formation is strongly influenced by the availability of water vapor and the supersaturation with respect to ice, which depends on the dynamical environment. Of particular significance are the high-frequency fluctuations, which are approximately 10 times larger than the slow synoptic-scale motions . Accurately capturing vertical wind variability is thus critical for modeling cirrus cloud formation and properties .

The biases in extreme ICNC values align with the model's underestimate of strong updrafts. This results in a vertical wind distribution that is too narrow and an underestimated standard deviation (Fig. ). Including a subgrid-scale vertical wind term (Sect. ) slightly improves the issue by increasing the standard deviation in vertical wind by 1 %–10 % depending on the model grid spacing (2 %–4 % for 1 km grid spacing). Nevertheless, the model performs well in representing the vertical wind variability near regions of deep convection at all three horizontal grid spacings (4 km, 1 km, 250 m) (Fig. b).

However, updraft variability is strongly underestimated in regions far from deep convection (defined as areas with brightness temperature > 240 K) for simulations at 4 and 1 km grid spacing. The 250 m simulation shows better agreement with the measured winds below 14 km, where anvil cloud coverage is the largest, but underestimates wind variability in the TTL, resulting in a persistent ICNC bias for small ice crystals (Fig. c and f). This also likely contributes to larger microphysical biases in clouds far from active convection compared to those near convection (Fig. ).

Power spectra in Fig. , calculated following , provide further insight into updraft variability across different wavelengths of atmospheric disturbances. Variability at wavelengths larger than ∼ 1000 km corresponds to synoptic-scale motion, while convectively generated gravity waves dominate at wavelengths between 1 and 1000 km. Turbulent processes of various sources dominate at wavelengths below ∼ 200 m.

As already shown by Fig. , wind variance increases with finer horizontal grid spacing. While the 4 and 1 km simulations substantially underestimate wind variance and power, the 250 m simulation reproduces the observed wind variability for wavelengths between 1 and 10 km. Moreover, all SAM simulations under-represent variability at scales larger than ∼ 100 km, which are better captured by GSRMs at horizontal grid spacings of about 4 km. Vertical wind spectra from models flatten out at the model effective resolution, which varies across the different GSRMs and SAM simulations shown in Fig. 10. Aircraft observations are most suitable for investigating vertical wind variability < 10 km and flatten out between 10 and 100 km, possibly due to a smaller signal-to-noise ratio, limited data, and/or changing atmospheric dynamics. Gravity waves, originating from deep convective updrafts, propagate hundreds to thousands of kilometers from their source. However, the narrow channel setup of the SAM simulation may restrict the generation and/or propagation of these waves compared to the real atmosphere or global GSRM simulations. While larger-scale updraft variability fosters favorable conditions for deep convection, its direct contribution to cloud formation is limited and of secondary importance for this study.

The representation of ice microphysics remains a limitation in this study, despite notable improvements achieved with the described computationally inexpensive changes. This reflects the broader challenge of accurately modeling ice microphysics in high-resolution models, where computational constraints and incomplete process understanding often impose trade-offs.

For example, the adopted scheme for ice nucleation, originally designed for coarse climate models, relies on empirical fits to parcel model simulations. While effective, this approach may not capture all the details of ice nucleation at cloud-resolving scales accurately. Future work could address this by implementing a physics-based nucleation scheme, such as the novel scheme, and incorporate a resolution-aware parameterization of wind variability to complement the model-resolved wind variability – similar to approaches being developed for liquid clouds .

Secondary ice formation may be another area of interest, given that recent studies have shown that it plays a large role even at temperatures as cold as -50 °C . Incorporating secondary ice processes into SAM-P3 currently requires additional computationally expensive ice modes, but their inclusion could enhance the model’s ability to simulate anvil cloud evolution. Finally, increased model complexity brings additional challenges, often resulting in hindered process understanding, increased uncertainty, or the problem of equifinality refer toand references therein for a complete outline of these issues.

An alternative approach could involve introducing ice tracers for each ice nucleation process, similar to those used by . This would allow for a more accurate classification of cirrus cloud origins without relying on arbitrary thresholds. Although this would increase the model's computational costs, the expense is substantially lower than doubling the model's resolution, making it a viable path forward.

Our findings confirm that increasing horizontal resolution improves the representation of atmospheric updrafts at scales crucial for cirrus cloud formation. However, this improvement comes at a large cost: a 4-fold increase in horizontal resolution results in a 16-fold increase in data output and a 20–25-fold increase in computational demand. This highlights the importance of exploring less computationally expensive alternatives to improve model performance. For instance, we achieved a large reduction in bias for ice cloud properties through straightforward modifications to the ice microphysics without appreciable increases in computational cost.

Encouragingly, the simulations at a typical GSRM grid spacing of 3–5 km show signs of convergence in updraft variability near active deep convection when compared to updraft observations averaged to same grid spacing, suggesting that the generation of convective gravity waves is reasonably well captured (note that model-simulated updraft variability still increases between simulations of 4 and 1 km grid spacing). However, these waves may not propagate far enough, as indicated by the underestimated updraft variance across all tested horizontal resolutions. Power spectral analyses indicate that grid spacings slightly finer than 250 m may sufficiently capture the scales of motion relevant for cirrus formation and maintenance. This is also the grid spacing at which convergence in cloud macroscopic variables and ice sources and sinks has been observed .

Vertical resolution also plays a crucial role. Previous studies demonstrate its influence on tropical cirrus properties and their responses to global warming , as well as convective self-aggregation . We also perform a sensitivity test that halves the vertical grid spacing from 200 to 100 m, leading to the same model improvement as an increase in horizontal grid spacing from 1 km to 250 m (Fig. g). Doubling the vertical resolution in the upper troposphere increases the computational costs by only a factor of 2.5. This suggests that refining vertical resolution could be a more computationally efficient way to improve model updraft variability than increases in horizontal resolution, and it should be thoroughly investigated in future studies.

Nevertheless, higher resolution alone cannot address all challenges. Processes at microscopic scales, particularly the interactions between ice microphysics and radiation, remain poorly resolved. Efforts to improve ice microphysics in GSRMs are scarce e.g., yet essential for advancing understanding of tropical cirrus evolution. Our study highlights the importance of accurately modeling cirrus evolution, a key factor in determining their microphysical properties and radiative effects.

Passive tracers, as demonstrated in this study, are a valuable tool for tracking the evolution of ice clouds in models. Although these tracers are purely computational, stable water isotopes, measurable in situ or via satellite, could serve as real-world tracer analogs, offering insights into the pathways of ice cloud evolution . Trajectory analysis, such as that in , can provide additional clarity on cirrus cloud evolution. Based on their analysis of cloud source and sink processes along trajectories, they proposed three cirrus cloud life cycles with distinct radiative signatures. In contrast, our analysis focused only on ice cloud properties. However, by enabling the 3D output of microphysical process rates, we could easily perform a similar process-rate analysis that would provide additional clarity on the two tropical cirrus formation pathways. Future work could explore a consistent integration of these approaches in models and observations, linking simulations with aircraft measurements as in and possibly satellite retrievals as well, in an effort to improve process understanding.