This research integrates four observational datasets. The Modern-Era Retrospective analysis for Research and Applications, Version 2 (MERRA-2; ) provides gridded fields at 0.5° resolution with 72 vertical levels and 3-hourly temporal sampling. The Clouds and the Earth's Radiant Energy System CERES SYN1deg-1Hour Edition 4.1 product (CERES; ) provides gridded radiative fluxes and cloud properties at a 1° spatial and hourly resolution. These fields are derived from retrievals across 16 geostationary satellites, calibrated against Moderate Resolution Imaging Spectroradiometer (MODIS; ) collection 5.1, and adjusted for radiative budget consistency using the Fu-Liou algorithm, with noted seasonal and diurnal biases . Comparisons with MODIS collection 6 are used to evaluate aerosol optical thickness and cloud properties. Precipitation is provided by the Integrated Multi-satellitE Retrievals for GPM precipitation on a half-hourly basis on a 0.1° grid (IMERG; ). These datasets are collocated with the pre-calculated Lagrangian trajectories to investigate the cloud evolution, provide collocated satellite and reanalysis data, which are averaged over a 1° by 1° grid-box, and provided hourly along the trajectory. Specifically, CERES and IMERG provide geolocation data (latitude, longitude, altitude, land fraction) and collocated satellite fields, while meteorological variables are obtained from MERRA-2 and MODIS.

The Large-Eddy Simulation (LES) model SAM (System for Atmospheric Modeling; ) is used for the simulations. To address the impact of aerosols on tropical cloud transition, simulations are conducted using an idealized Lagrangian framework , based on the mean trajectory derived from each initiation location in the observational data. In doing so, we do not aim to exactly reproduce the observed mean evolution, acknowledging the nonlinear relationship between individual trajectory behavior and their ensemble-mean response, but to represent observational-based idealized evolution. The model simulates the emerging cloud evolution under two vertically uniform CCN levels of 800 and 20 cm⁻³, assuming a supersaturation of 1 %. This wide range of aerosol conditions is useful for establishing physical understanding and is not intended to mimic the observed difference. The simulations use a two-moment bulk microphysics scheme and the RRTM radiation scheme . It is important to note that the aerosols are not prognostic in our simulations.

The domain size is chosen to be 57.6 km × 57.6 km with 147 stretched vertical levels extending up to 33 km. Vertical grid spacing is finer (tens of meters) in the lower atmosphere to better resolve boundary-layer processes and gradually increases with height. Horizontal grid spacing is 200 m × 200 m. We use a time step of 2 s and a radiation time step of 30 s. This configuration balances the need for a sufficiently large spatial domain to resolve mesoscale cloud structures and dynamics while still resolving small-scale cloud processes at LES resolution, enabling us to capture the full tropical cloud transition . To assess the sensitivity of the results to domain size, we performed an additional simulation with a larger domain (102.4 km × 102.4 km) at the same horizontal resolution (200 m × 200 m), as shown in Fig. S62. The results are qualitatively consistent with the baseline configuration, indicating that the chosen domain size does not substantially affect the main conclusions. This larger domain is comparable to the 1°×1° grid box of the observational data.

We apply small temperature perturbations (O(0.02K)) near the surface at the beginning of the simulation to initiate boundary-layer turbulence and allow initialization of convection. To ensure cloud presence at the start of the simulations, we initialized the model with a supersaturation of 1 % at the midpoint of the cloudy layer during the first hour of the simulation. Humidity increased linearly from the surface up to 1 % supersaturation at the cloudy layer midpoint, then decreased linearly back to the background value at the top of the cloudy layer. This step is not intended as a physical representation of supersaturation but as a practical initialization procedure to produce early cloud development consistent with the observed conditions along the Lagrangian trajectories.

To estimate water-vapor mixing ratios, qv, in a way that accounts for temperature differences between polluted and clean groups, we first compute saturation quantities. The saturation vapor pressure over liquid water (es(T)) is given by: 4es(T)=611.2exp⁡17.67(T-273.15)T-35.85, according to (Eq. 10), where T is the air temperature. The corresponding saturation mixing ratio (qs(T,p)) is given by: 5qs(T,p)=εes(T)p-(1-ε)es(T),ε=0.622, where p is the air pressure. For the polluted group, we include a uniform offset, ΔT, based on the location's SST difference, to represent the observed warmer temperature background. This procedure is applied only to the polluted group to normalize it by the ΔT and compare it with the clean group.

The reconstructed water vapor mixing ratio for each trajectory is obtained as: 6qv(grp)(z,n,t)=RH(grp)(z,n,t)qsTref(grp)(z,t),p(grp)(z,n,t), where the relative humidity (RH) is used in fractional form and the result is expressed in g kg⁻¹. Here, “grp” denotes the trajectory group (clean or polluted). This approach isolates the role of relative humidity (RH) from that of temperature: polluted-clean differences can thus be attributed either to the background temperature (via ΔT) or to humidity anomalies independent of temperature.

We start by analyzing the observational dataset to identify patterns in the evolution of polluted and clean cloud regimes. Specifically, Fig.  shows the evolution of key environmental variables and Fig.  shows cloud and radiation properties. Both figures represent the NEP1 dataset. Equivalent figures for all other initiation locations are provided in the Supplement (Figs. S2–S17), which exhibit generally consistent behavior across locations. In case differences arise, they will be explicitly mentioned.

Across both polluted and clean groups, the trajectories reflect a gradual warming of the SST (Fig. a), indicative of movement from the cooler subtropics toward the warmer tropics. Simultaneously, a steady decline in lower tropospheric stability (LTS; defined as the difference in potential temperature between the 700 hPa level and the surface; Fig. b) suggests a destabilizing boundary layer environment. Specific humidity at 850 hPa (qv) increases markedly along the trajectories (Fig. c), indicating moistening of the lower free troposphere and a shift toward conditions favoring deeper convection. These thermodynamic trends are consistent with a progressive deepening of the cloud layer, as evidenced by the increasing cloud top height (CTH; Fig. d). Before considering aerosol differences, we note that these mean trajectories reflect the canonical tropical cloud transition.

The evolution of cloud fraction (CF; Fig. c) also reflects the gradual transition between cloud regimes. Initially, CF is increasing rapidly during the first day, representing the formation of extensive Sc decks during days 2–3, then declines as these decks break up into scattered Cu during days 4–5. Finally, as deep convective clouds develop in the last three days near the deep tropics, CF increases again.

As the air mass advects equatorward, LWP increases during the shallow-to-deep transition, and radiative properties respond accordingly. Specifically, as the air mass moves equatorward, the TOA emitted longwave (LW) radiation decreases steadily, reflecting the rise of cloud tops into higher and thus colder layers of the troposphere. The reflected TOA shortwave (SW) radiation is initially high (during mid-day), then declines and increases back again as deep convective systems form later in the trajectory, following generally the pattern of CF. This baseline progression provides the physical context for interpreting the differences between the clean and polluted groups. We note that during the final stage of the trajectories (days 6–8), clouds become substantially deeper, as indicated by increasing CTH, and may include mixed-phase or glaciated cloud tops. In this regime, MODIS liquid-phase retrievals such as re and LWP may not fully represent the cloud column and should therefore be interpreted with caution.

Having established the baseline tropical cloud transition, we now examine how aerosol loading modifies this evolution by comparing the clean and polluted groups. The polluted group exhibits generally smaller cloud droplet effective radii (re; Fig. a) compared to the clean group from the second day forward, consistent with the expected microphysical signature of the Twomey effect . This reduction in droplet size coincides with a persistent enhancement in LWP (Fig. b) for polluted trajectories, especially during the last few days. The average LWP difference along the entire 8 d is 9.3 g m⁻².

Total CF (Fig. c) is generally higher in the polluted group around the third day and onward, indicating more extensive cloud cover in the later phases of the cloud transition. This is accompanied by a faster and more pronounced increase in CTH (Fig. d), with polluted trajectories reaching higher cloud tops earlier in the transition and ending with deeper convection than the clean group, with a mean difference of about 1.3 km over time.

In terms of the radiative effects in the different groups, the polluted group shows enhanced reflected TOA SW radiation (Fig. e) compared to the clean group, consistent with increased cloud optical thickness from smaller droplets (manifested by the decrease in re), and higher LWP and CF. This difference in reflected SW radiation is pronounced in the latter half of the trajectory and has a time-average difference of about 19 W m⁻². The emitted TOA LW radiation (Fig. f) is lower in the polluted group, especially in the later stages of the trajectory, with a time-mean difference of about -14 W m⁻², reflecting the combined influence of higher and colder cloud tops, and increased CF. These trends are consistent across all different initiation locations (Figs. S10–S17; see also Fig.  below).

Figure  provides an observational perspective on the robustness of aerosol-related cloud adjustments. Across all basins (NEP, SEP, and SEA), a generally consistent signal emerges: polluted trajectories tend to have smaller re (Fig. a; with the exception of SEP3), higher LWP from day 3 forward (Fig. b), larger CF (Fig. c), and higher CTH (Fig. d) compared to their clean counterparts. The reflected SW radiation at the TOA (Fig. e) is higher on average for polluted trajectories across all locations, reflecting the combined effect of the generally higher LWP, greater CF, and smaller droplet sizes. Correspondingly, emitted LW radiation (Fig. f) is lower on average in polluted cases, consistent with higher and colder cloud tops and enhanced by a larger CF. Yet these differences cannot be attributed solely to aerosol impacts, since the two groups also differ in their underlying thermodynamic environments (Figs. ; S2–S9), reflecting potential confounding factors, i.e., co-variability with meteorological state .

The opposite re response in the first 5 d seen in SEP3 may reflect the influence of the coastal aerosol environment off northern Chile. This region frequently experiences strong offshore gradients in cloud microphysical properties and episodic enhancements in sulfate outflow , which could shape the observed signal. Furthermore, satellite-retrieved AOD may include contributions from elevated aerosol layers above cloud top, which can increase column aerosol optical depth without directly affecting in-cloud microphysical processes, potentially weakening the observed cloud response .

The co-variability between aerosol concentration and meteorological conditions is demonstrated in Fig. . While both groups follow similar overall trends fitting with the tropical cloud transition: warming SST (Fig. a), declining lower-tropospheric stability (Fig. b) and increasing qv at 850 hPa (Fig. c), a systematic offset emerges between the polluted and clean groups.

Specifically, polluted trajectories exhibit slightly warmer SSTs, particularly during the mid to late stages of the cloud evolution, potentially providing a more favorable thermodynamic environment for deep convection. Here, the SST differences between the groups reflect co-variability with the large-scale meteorological and seasonal background state rather than a direct aerosol effect on SST. LTS decreases more sharply for polluted cases, implying a deepening of the inversion layer and with that, an earlier onset of deep cloud growth. qv at 850 hPa is generally higher in the polluted group, indicating a moistening of the lower free troposphere that could enhance and sustain deep convection. These environmental differences suggest that at least part of the cloud development differences between polluted and clean trajectories (Fig. ) may be supported or explained by variations in the environmental thermodynamic conditions. The combination of higher SSTs, reduced stability, and increased lower-tropospheric moisture in polluted trajectories creates a background state more conducive to rapid cloud deepening and larger radiative impacts.

These environmental differences between clean and polluted conditions (Fig. ) are robust across the NEP locations (Figs. S2–S3), while the SEP locations show even larger differences, with polluted trajectories consistently warmer, less stable, and more moist (Figs. S4–S6). In contrast, the SEA locations display the opposite behavior: polluted trajectories experience slightly cooler SSTs, and more stable and drier conditions than the clean group (Figs. S7–S9), which could be driven by seasonal biomass burning from the African coast (; Figs. S33–S35). Notably, despite the opposite thermodynamic differences between clean and polluted conditions in the SEA region compared with the other basins, we still find a similar cloud and radiative response. Specifically, polluted conditions are characterized by smaller re, higher LWP, CF, and CTH, more reflected TOA SW radiation, and reduced TOA LW emission, as in the other basins (Fig. ). The fact that the correlation between aerosol concentration and thermodynamic conditions differs between regions, yet the pattern of the polluted-clean differences in cloud and radiation properties remains consistent, suggests that these signals cannot be fully explained by environmental variability alone.

Importantly, we also find that the two groups differ in the seasonal composition (Figs. S18–S26), reflected in the number of trajectories originating from each season. These differences are indicative both of the inherent seasonality of aerosol loading in the full dataset and additional imbalances introduced by our filtering process (Sect. ), which in turn manifests as differences in incoming solar radiation between the two groups. These seasonal differences could partly explain the observed environmental differences, even though the spatial advection paths of the trajectories remain broadly similar across seasons (Figs. S27–S35). To test whether the differences between polluted and clean groups could arise from seasonality differences, we conduct sensitivity analyses using a seasonality-controlled bootstrap. The polluted-clean differences persist even after accounting for the seasonality (Fig. S61).

The robust agreement in sign of cloud and radiative properties differences for clean and polluted conditions across diverse meteorological regimes and ocean basins, which also have different seasonality (Fig. ), supports the interpretation that the cloud microphysical and macrophysical responses to pollution: smaller droplets, enhanced LWP, CF, and deeper clouds are robust features in our satellite record.

Using AOD as a proxy for aerosol loading has known issues . Hence, sulfate aerosol mass concentration (SO4) at the 910 hPa level was also tested as an alternative proxy for aerosol loading (as was suggested in ). Both aerosol proxies (AOD and SO4) yield generally similar cloud adjustments between polluted and clean trajectories as reported above (smaller droplet sizes, higher LWP and CF, and deeper clouds for polluted trajectories). However, using AOD as an aerosol proxy produces more pronounced differences in CTH and radiative properties across the different locations (Fig. S60). The main discrepancies between the two methods arose in the correlation between aerosols and environmental variables: for SO4, the polluted group shows lower SST, higher LTS, and lower qv – opposite to the AOD-based grouping (Fig. ). Despite these environmental differences, both aerosol proxies led to consistent cloud adjustments, suggesting that the signal is generally robust to the choice of aerosol metric.

To better isolate and understand the direct influence of aerosols on cloud and radiative properties, we turn to model simulations where, by construction, the aerosol influence is decoupled from the confounding meteorological factors present in the satellite data. Unlike the observations, where aerosol and meteorological effects are intertwined, the model simulations isolate the impact of aerosols by holding environmental conditions the same between polluted and clean runs. This allows the simulated cloud adjustments to be more directly attributed to aerosol perturbations.

Figure  shows Hovmöller diagrams illustrating the simulated evolution of cloud cover, showing the transition from Sc clouds to deep convection. At the early stages, clouds are mostly confined to the boundary layer, where shallow Sc clouds dominate below ∼ 1 km, with high CF. With time, around day 2 to 4, moistening of the lower free troposphere and gradual destabilization of the boundary layer support the breakup of Sc into scattered shallow Cu with lower CF and slightly deeper clouds, marking the transition to a more convective regime. As the lower troposphere becomes increasingly humidified, convection deepens, and cloud tops rise steadily into the mid- and upper-troposphere. During the last two days of the simulation, the domain is characterized by deep convective clouds, extending above 12 km. This progression is consistent with the canonical subtropical to tropical cloud transition.

The polluted simulation (Fig. a) exhibits higher CF through much of the column, particularly in the mid to upper troposphere during the later stages of the trajectory. The clean simulation (Fig. b) also shows cloud deepening, but with reduced vertical extent and lower mid-tropospheric coverage. Figure c highlights these differences, with positive anomalies (red) dominating above the boundary layer, indicating earlier and more pronounced vertical development in polluted cases. This is especially pronounced during the first two days of the cloud evolution. The earlier onset of mid-level cloudiness in the polluted group suggests a faster erosion of the capping inversion and enhanced detrainment aloft, consistent with the microphysical suppression of precipitation and associated increases in LWP and CF (Fig. ; ).

Figure  shows the corresponding hourly evolution of cloud and radiation variables from the model for the same simulations presented in Fig. . The simulations reproduce the main observed trends: increasing LWP during the deepening of convection, decreasing CF during the Sc breakup, and recovering with the onset of deeper convection, and CTH rising as cloud systems grow deeper. These cloud changes with time are accompanied by radiative changes, including enhanced mid-day reflected SW and reduced emitted LW as cloud tops ascend into colder levels. While some differences in magnitude and timing exist compared to observations, the model captures the essential progression of tropical cloud transition. The time evolution of the simulations from the different initial locations is presented in Figs. S44–S51, and demonstrates a generally consistent behavior.

The simulated aerosol impact reproduces many of the observational signals: higher LWP, especially in the latter half of the simulation (Fig. b), largely similar CF (Fig. c), and slightly higher CTH (Fig. d) in polluted conditions compared to clean trajectories. The radiative responses in the model also align with the observed cloud adjustments, with polluted trajectories showing enhanced reflected SW (Fig. e) and slightly reduced emitted LW (Fig. f), which is amplified by the vertical cloud deepening (Fig. ) and the increased high-altitude cloud cover. Together, these are indicative of optically thicker clouds with higher and colder tops, consistent with the observed trend (Fig. ). As expected, we can also see a large enhancement in Nd (Fig. a), which reflects the imposed increase in CCN in the domain.

Across all basins, the model produces a coherent signal: polluted trajectories exhibit substantially higher Nd (Fig. a), which propagates into enhanced LWP (Fig. b), larger total CF (Fig. c), and higher CTH (Fig. d). These cloud adjustments highlight the systematic influence of increased aerosol concentrations on both the horizontal and vertical extent of cloud systems. Radiative responses are similarly consistent across basins. Reflected SW at the TOA (Fig. e) is higher on average for polluted trajectories in all locations, consistent with optically thicker and more extensive clouds. Conversely, emitted LW at TOA (Fig. f) is on average lower under polluted conditions, in line with higher and colder cloud tops. The magnitudes of these differences are generally comparable to those derived from the satellite observations (Fig. ). However, the observation-model comparison is not straightforward due to the different aerosol perturbations, with model aerosol differences larger than the observed differences.

The consistency in both the sign and the general magnitude of these simulated responses across the NE Pacific, SE Pacific, and SE Atlantic is interpreted here as the cloud and radiation changes are robust outcomes of aerosol perturbations during marine subtropical to tropical cloud transitions.

The interpretation of ACI from observations is complicated due to the co-variability between aerosol concentration and meteorological conditions. Air masses differ not only in aerosol loading but also in their thermodynamic environments (e.g., temperature, stability, humidity), making it difficult to determine whether observed cloud differences arise from aerosols or from pre-existing environmental variability . Consequently, it is often assumed that the apparent cloud adjustments are strongly shaped, or even dominated, by background meteorological variability rather than the correlated aerosol influence itself .

Here, we argue that while it is well established that aerosol-meteorology co-variability can strongly influence apparent cloud responses, an additional aspect, previously discussed in the literature , deserves further attention.

Environmental changes are not only an external confounder that co-varies with aerosol loading; they may also partly arise as a consequence of aerosol impact on clouds. Previous studies have shown that aerosols can modify cloud microphysics and vertical development in ways that subsequently feed back into the surrounding atmospheric structure . According to this interpretation, differences in thermodynamic structure for different aerosol conditions may reflect the combined influence of large-scale meteorological variability and aerosol-induced cloud adjustments that reshape humidity, stability, and energy transport. In this context, the environment is not only a background influence on clouds, but may also be shaped by cloud responses to aerosols. Aerosols, therefore, might act as an internal driver contributing to cloud and moisture adjustments alongside the external modulation imposed by the large-scale environment.

Figure  demonstrates this point by showing vertical profiles of the differences in qv between polluted and clean conditions from both observations and model simulations. Figure a presents the qv profiles after accounting for SST differences between polluted and clean trajectories. Thus, they represent the humidity change, which is decoupled from the SST differences (Sect. ).

In the observational record (Fig. a), polluted trajectories show enhanced moistening near the top of the boundary layer and into the lower free troposphere, with the signal reaching higher altitudes compared to the model results. This vertical structure is consistent with cloud deepening and detrainment of moisture aloft. In the observations (Fig. a), the early days (days 1–3) show relatively shallow moistening confined near the boundary layer top (∼ 1–2 km), whereas later on (days 5–8) the moist anomaly strengthens, extends and peaks at 4–6 km. This progressive deepening of the moist layer is consistent with a transition from stratocumulus-dominated conditions to more convective regimes.

The model generally reproduces this qualitative vertical structure (Fig. b): moistening near and above the boundary layer top, though with a smaller magnitude and a slight drying near the surface. The general agreement between observations and model simulations (where, by design, the aerosol impact is isolated from natural co-variability with meteorological conditions) is consistent with the interpretation that aerosols may contribute to shaping thermodynamic structure through cloud–environment interactions, although we cannot fully rule out the influence of co-variability with large-scale meteorology in the observational analysis. Importantly, we suggest this as a possible interpretation. We do not wish to suggest that this effect is conclusively demonstrated by our analysis.