We utilize in situ observations gathered during the Atlantic Tradewind Ocean–Atmosphere Mesoscale Interaction Campaign (ATOMIC) , which was the United States contribution to the international Elucidating the Role of Clouds Circulation Coupling in Climate Campaign (EUREC⁴A) . The NOAA RV Ronald H. Brown (RHB) sampled between 7 January and 13 February 2020 in a region of the ocean between Barbados (60° W) and 51° W, the location of the Northwest Tropical Atlantic Station (NTAS) mooring (Fig. ). Local time is approximately UTC-4. Generally, the RHB stayed between 16 and 13° N. We used thermodynamic measurements made on the RHB to investigate the surface and near-surface conditions, ceilometer measurements to understand the cloud-base height (CBH) behaviors, as well as aerosol measurements and radiosondes launched every 3 h to investigate the thermodynamic profiles of the atmosphere.

Primarily, we utilized the vertical velocity measurements captured by the motion-stabilized Doppler wind lidar from NOAA's Chemical Sciences Laboratory (more instrument details in ). Measurement processing followed the methodology of , as described in further detail below. Vertical velocity profile measurements were gathered at a rate of 2 Hz and with a vertical resolution of 33.6 m. Once every hour, the lidar performed a conical ∼ 2 min scan at 15° from zenith to retrieve horizontal wind information before returning to its motion-stabilized vertical position. Lidar vertical pointing is stabilized within 0.03° (1σ standard deviation) of zenith, enabling accurate measurements of vertical velocity. Our analysis was focused on the profiles of vertical velocity (w) up to CB, a region that was reliably and consistently observed across the campaign. The lidar signal-to-noise ratio becomes too low above this height due to attenuation by clouds or, when clouds are not present, lack of aerosol above the boundary layer.

For a given Doppler lidar profile measurement, cloud occurrence was identified as where the range-corrected intensity (RCI) of a given pixel exceeded a threshold developed following the method outlined for Cu clouds in . The RCI threshold was sensitivity-tested and the RCI behavior was compared to the bimodal distribution used to distinguish cloud and aerosol samples in . The cloud-identified pixels (∼ 0.5 s by 30 m) that occurred lowest in altitude are taken as the CBH. Successive profile measurements where cloud-identified pixels occurred within 10 s and 50 m of each other were aggregated together into a single cloud scene. The reference CBH of the cloud scene is calculated as the 25th percentile of CBH across the aggregated profiles composing the total scene . This method produced a Doppler lidar CBH that agreed well with the ceilometer observations (Fig. 11a, ). The length of a cloud is defined as the duration of the cloud scene scaled by the horizontal wind speed at CB, which is interpolated from the hourly horizontal velocity measurements from the lidar (chord length, LChord). Because we were interested in the behavior of developing or active clouds, we further restricted our dataset to cloud scenes with a CBH within 50 m of the LCL. The LCL was determined from an adiabatic parcel model initialized with thermodynamic observations from the RHB.

Each cloud scene can be thought of as a snapshot of vertical velocity behaviors occurring around a single cloud. Cloud scenes were processed into normalized w matrixes for ease of computations , where length is normalized by total cloud length at CB (-1.5 to 1.5 in x/LChord where cloud occurs between -0.5 and 0.5) and altitude is normalized by CBH (0 to 2 in z/CBH where CB is at 1). Figure a shows an example normalized w matrix for a cloud scene, highlighting a plume connecting from the surface to CB and observable partway into the cloud before attenuation. The sub-cloud and CB updraft region where our analysis is focused is marked: -0.5 to 0.5 in x/LChord and 0 to 1 in z/CBH.

The CB amounts from the individual Doppler lidar cloud scenes and the ceilometer have statistically similar magnitude and broadly similar diurnal cycle shapes (Fig. a, c). Ceilometer CB amount was derived from the 15 s data product for first cloud-base detection restricted to CBH ≤ 800 m (the upper quartile of MO CBH, Fig. b). This is a different sampling rate than the Doppler lidar (which excludes clouds smaller than 10 s in duration) and an approximate CBH cut-off picked to correspond to the more exact LCL restriction applied to the Doppler lidar scenes. These retrieval and resolution differences likely contribute to the differences in cycle details. However, the general statistical agreement and similarity in cycle shape indicate that the sampling of the identified cloud scenes is comprehensive for this period of low, active clouds.

We were additionally interested in the behaviors of thermal plumes sampled by the lidar. A plume is defined as a contiguous region where w > 0.05 m s⁻¹. The plume horizontal length (LPlume) is taken as the maximum diameter of the plume measured in seconds and converted to meters using the interpolated, hourly surface wind speed where the surface is ∼ 60 m, the lowest measured altitude for the Doppler horizontal wind. Because not all plumes are topped by clouds and have varied depth, we utilized the surface wind to be consistent across all plume features. However, the wind shear between this altitude and the CB is negligible so the length conversions are comparable across cloud and plume features with little bias (Fig. ). Note that this analysis is agnostic to the number of plumes contributing to the maximum plume length as, due to the nature of the cross-sectional sampling of the lidar, we are not able to robustly distinguish between multiple overlapping plumes and one wide plume contributing to the contiguous plume feature. Plumes are labeled as cloud-topped if the plumes occur within 1 min of a cloud identification that has a positive updraft at CB (LPlumeCloud). Otherwise, they are assigned to be clear-sky (i.e., “unsuccessful” plumes, LPlumeClear).

We further restricted the cloud identification data to examine clouds with active CB cores (wCB > 0). We defined the CB core updraft for individual clouds (wCBCore) as the mean of CB values that satisfied wCB > 0. The corresponding CB core length (LCBCore) for individual clouds is defined as the total cloud length, LChord, scaled by the fraction of the samples that contribute to the active CB core. In the example w matrix in Fig. a, the plume noticeably strengthens near CB and in cloud, marking the core updraft. In addition to the CB core observations, we analyzed sub-cloud updraft profiles (w >0). We do not include the samples outside of cloud (i.e., < -0.5 and > 0.5 in x, Fig. a) as they may be contaminated by nearby clouds. For this reason we also do not analyze downdrafts as they are likely incomplete, falling largely outside of the sub-cloud range.

We are not able to sample in heavily precipitating conditions. This did not pose a significant issue as there were few instances of precipitation recorded on the RHB . The RHB tended to sample much further upwind compared to the other platforms during EUREC⁴A, observing earlier in the evolution of clouds, before precipitation and the accompanying cold pools developed . This suggests that these observations are particularly well suited for investigating the early stages of Cu clouds in this region, adding context to their early behaviors before the clouds reach the main EUREC⁴A domain and Barbados.

The goal of our comparison is to contrast organized cloud structures with unorganized ones. Distinguishing by organization was a nontrivial task. Mesoscale structures have been variously categorized in recent years (e.g., ) to investigate their influence on the climate system. One such methodology developed a neural network to categorize four archetypal cloud types based on expert assessments (SGFF: sugar, gravel, flowers, and fish) to aid trade Cu investigations (e.g., ). This algorithm, along with hand identifications from the EUREC⁴A team, were applied to GOES-16 and MODIS satellite scenes over the EUREC⁴A region and campaign period to develop the C³ONTEXT dataset . Our initial analysis utilized C³ONTEXT SGFF identifications collocated to the location and time of RHB sampling (see Hovmüller diagram, Fig. ). However, many cloud samples could not be confidently labeled as one of the SGFF types, leading to a large number of unclassified or “none” types dominating the sampling (Fig. ). This presented a problem for analyzing the comparatively sparse campaign sampling, unlike analyses based on multiyear satellite climatologies (e.g., ).

To maximize the number of cloud scene identifications utilized, ensuring more reliable statistical sampling, we developed a more general classification methodology: hand-identifying clouds into two categories, less (LO) and more (MO) organized structures. We relied on MODIS Aqua (13:30 local time, LT) and Terra (10:30 LT) satellite imagery to identify the daily cloud field where the RHB was sampling (see imagery used for classification in Figs.  and ). The LO category (Fig. ) comprised days where only small, scattered structures occurred randomly (30 and 31 January and 1, 4, and 9 February; 1356 lidar cloud scene identifications, 1275 cloud-topped plumes, and 27 040 clear-sky plumes). The MO category (Fig. ) comprised days where more discernibly clustered structures as well as some smaller structures occurred in larger patterns of organization (9, 10, 11, 12 January and 10 and 11 February; a total of 2539 cloud scenes, 2492 cloud-topped plumes, and 29 284 clear-sky plumes). LO days included sugar-dominated cloud scenes and MO included gravel and a few flowers-like cloud scenes.

The LO and MO categorized days, which will be used for compositing throughout the rest of the paper, help us to distinguish between cloud systems that have not been substantially influenced by mesoscale organization and are still occurring fairly randomly (LO, Fig. a) and those that have already been influenced enough by mesoscale circulations generated through the moisture–convection feedback to begin gathering into discernible mesoscale organization patterns (MO, Fig. b). Different LO and MO sample sizes are accounted for by utilizing standard error to determine confidence in the mean and statistical tests for determining distribution differences at 95 % confidence. Days where very large cloud structures associated with the decay of midlatitude cold frontal cloud systems (i.e., fish, ) were not included in our analysis to focus on locally driven cloud systems, as previously noted. We extract the full diurnal cycle based on the daytime cloud identifications to obtain a cohesive signal that can be compared to the corresponding environmental cycles without adding additional complexity in interpretation from day-to-day variations. Although organization varies diurnally , it would be challenging with our small sample size to ensure a complete cycle was captured and appropriately connected to environmental behaviors if classifications were made at a greater temporal frequency (e.g., , where the larger sample size eases these concerns).

To confirm that this coarse LO–MO separation distinguished measurements by organization as intended, we additionally compared these categorizations with two objective classifiers also included in the C³ONTEXT dataset: the mean cloud object size (S) and the organization index (Iorg, which compares the distribution of nearest-neighbor distances between centroids to a random distribution) . This phase space has been reliably utilized to classify trade Cu (e.g., ). Iorg and S were calculated on 10 × 10° brightness-temperature-defined scenes of cloud amount (as in ), approximately centered on the EUREC⁴A region. They were then collocated to the location and time of the RHB sampling (Fig. b). Note that this comparison is approximate since the RHB was often sampling on the edge of this domain, not at its center. The interquartile ranges corresponding to the SGFF archetypes in are shown for reference following . The LO data tend to fall in the sugar quadrant (lower S, higher Iorg), while the MO data tend to fall in the gravel quadrant (lower S, lower Iorg). The apparent greater organization of LO (Iorg > 0.5) follows the behavior of sugar discussed in . This counterintuitive behavior is likely a feature of the brightness temperature thresholds, which are more sensitive to the sparsely distributed deeper clouds in these scenes than the dominate low clouds. Overall, this comparison indicates that (i) the visual categorization described above successfully separates observations by organization to the first order (y axis) and (ii) both LO and MO will span similar cloud structure sizes (x axis). This is fortuitous as it allows us to focus on organization without the additional complexity of structure size contributing to cloud behavioral differences (i.e., as would occur at later stages of organization development when precipitation and cold pools begin to play a role, Fig. c).

Our results will be presented in the context of the current theory on how mesoscale organization influences Cu in the tropical trade winds (e.g., ), as summarized in Fig. . Cu clouds develop randomly when surface-driven, thermal plumes (purple curved arrows) cohere into updrafts (purple vertical arrows) and rise past the LCL, condensing moisture and releasing heat (a, LO). To maintain the tropical weak temperature gradient approximately, gravity waves export heat from the clouds, generating mesoscale circulations (orange to blue arrows). Circulations begin to aggregate moisture in their ascending branches (blue haze, b), gathering more moisture into Cu clouds that can in turn generate more condensational heat and trigger more gravity waves, reinforcing mesoscale circulations and deepening Cu. This moisture–convection feedback transitions non-precipitating cloud fields from LO (a) to MO (b) through generating more organized cloud structures over time. Note that find that the growth rate of cloud layer heating and mesoscale velocities cannot be explained by rapid convective adjustment to surface buoyancy flux anomalies, indicating that this organization is likely not driven by mesoscale surface forcing. Clouds in the descending branches will experience more dry, subsiding conditions and begin to die. Eventually, through moisture–convection feedback developing wider (i.e., length-scale growth of moisture fluctuations, ) mesoscale circulations , clouds will be organized into larger-scale patterns (e.g., flowers, c) that are deep and bright and begin to precipitate . The cold-pool downdrafts associated with stronger precipitation as these features continue to strengthen will eventually disrupt the updraft reinforcement cycle and break this moisture–convection organization feedback (not shown). Results have been incorporated into Fig.  and will be highlighted in Sect.  and .

The central goal of this work is to identify whether there is a detectable difference in cloud dynamics across mesoscale organizational states in the early stages of cloud development (i.e., LO vs. MO in Fig. a, b). The results will inform our understanding of how the moisture–convection feedback affects cloud dynamics as mesoscale organization strengthens. We began by comparing the MO and LO composites of updraft velocity mean (panel a) and variance (panel b) profiles computed for individual cloud scenes (Fig. ). For each composite, here and throughout the paper, the mean and 95 % confidence in the mean (twice the standard error, 2 SE) are shown for each vertical bin (either in altitude or normalized altitude space, z/CBH, as shown here). We apply the Mann–Whitney U test for each bin to test whether the nonparametric distribution of the MO and LO data is statistically likely to have come from the same population at 95 % confidence.

MO clouds have larger mean updrafts that are from a statistically different population than the LO cloud updrafts throughout the sub-cloud layer, CB, and into the observable lower cloud layer prior to lidar attenuation (Fig. a). This organizational difference also holds for the mean variance in the updrafts (Fig. b), indicating that MO clouds have greater turbulence than LO on average in addition to having more powerful updrafts.

In Fig. , all cloud sizes are included in the organization composites. On average structure sizes are similar between MO and LO (Fig. b) but it is worth testing for variations in behavior across cloud size. For example, the mean MO–LO separation could be influenced if there was disparity in updraft region size and the subsequent core area of the cloud. Additionally, if one organization state tended to have larger updraft regions their cores may be more protected, increasing the number of strong updrafts and shifting the mean to larger values compared to the other organization states. To address both of these issues we composited MO and LO clouds by LCBCore ranges (Fig. ), which is also generally informative of cloud size separation (e.g., Fig. d–e) assuming minimal horizontal expansion and tilting with height (e.g., Fig. 14 for sugar and gravel, ). The selection of LCBCore ranges is based approximately on the quantiles of the distribution of cores across the data such that similar amounts of low cloud are captured for each core range. In both organization states, a majority of clouds occur in the 0 to 500 m and 500 m to 1 km ranges and progressively fewer occur in the 1 to 2 and 2 to 7 km (Fig. e). This indicates that the total mean results (Fig. ) are weighted toward the smaller clouds, which makes their substantial differences even more notable.

The resulting profiles for updraft mean (Fig. a, c) and variance (Fig. b, d) composited by LCBCore provide three insights. First, MO updraft mean and variance are larger and significantly different at 95 % than LO updrafts for all LCBCore ranges across the majority of the cloud and sub-cloud altitudes. The few exceptions to this are likely impacted by sampling as they occur at the surface and in the larger LCBCore ranges, exhibiting wider 2 SE bars. For the mean updrafts, the 1 to 2 km range has several bins in the lower half of the sub-cloud layer where MO and LO are not statistically different. For the updraft variance, 1 to 2 and 2 to 7 km composites have several bins in the upper and lower half of the sub-cloud layer, respectively, that are not statistically different. This overlap pattern leads to the second insight. With increasing organization, cloud-topped plumes increase in updraft strength but strengthening appears to occur more in the upper sub-cloud layer and CB region. This suggests that the organization-dependent influence on the updrafts is associated with a cloud layer phenomenon. For example, updrafts may be assisted by the returning branches of mesoscale circulations generated by condensational heat release in the cloud layer. This organizational strengthening could play an integral role in the moisture–convection feedback (Fig. a, b). Third, with progressively larger LCBCore ranges, updraft magnitudes increase (Fig. b, c). The LO composite is an exception to this, staying relatively consistent across the three largest LCBCore ranges. This discrepancy in behavior between LO and MO is intriguing, suggesting that LO wCBCore is less assisted by the cloud layer phenomenon, while MO wCBCore can continue to strengthen with increasing cloud core size. This may indicate greater differences in cloud layer influence or contributions from environmental factors (see Sect. ).

Similarly, we evaluate the behavior of the updrafts across the diurnal cycle (Fig. ). For all times of day, MO updraft mean and variance are larger and come from a statistically different population than LO updrafts at 95 % confidence. Within their respective composites, MO and LO updrafts are relatively persistent and do not have statistically robust diurnal cycles in wCBCore at 95 % confidence. An exception is the distinct peak in wCBCore variance for MO at the end of the night (06:00–12:00 UTC or 02:00–08:00 LT, b; 10:00 UTC or 06:00 LT, f), recovering from a nighttime lull (05:00 UTC or 01:00 LT, f). Mean MO wCBCore echoes this lull and recovery (c). LO may experience an increase in mean and variance at the end of the day, peaking enough sub-cloud and at CB (18:00–24:00 UTC, or 14:00–20:00 LT, c, f) to overlap statistically with MO. These deviations from the constant wCBCore mean and variance suggest that different factors may contribute to MO and LO updrafts over the day: LO appears more tied to the solar influence, while MO may be more influenced by nighttime factors. The order of LO updraft magnitudes also lags between the mid-plume height (0.5 z/CBH, 00:00–06:00 UTC, d) and CB, indicating a potential lag between the surface influences and their impact on clouds. This lag is not apparent for MO, which has similar profile ordering throughout (a, b), suggesting the cloud layer may support the plume strength when surface influences are declining. We will return to this in Sect. .

To gain a more holistic understanding of the updraft differences, we also evaluated the differences in plume and cloud size characteristics across organization states. First, we contrasted the distribution of the plume maximum horizontal lengths in clear-sky (a, “unsuccessful”, LPlumeClear) and cloud-topped (b, “successful”, LPlumeCloud) conditions (Fig. ). LPlume exponentially decays with size for all cases, changing in slope around 2 to 3 km. Unsuccessful plumes occur more frequently than successful ones across all lengths and organization states (a vs. b). This is consistent with relatively low CB cloud fractions. Diurnal mean cloud amount is ∼ 9.5 % for LO and ∼ 13 % for MO, with 2 % 2 SE for both (Fig. a, c). While the mean CB amounts are statistically similar, the lower LPlumeCloud numbers for LO may be associated with the larger diurnal cycle in LO than MO: LO cloud amount is minimized between 08:00 and 16:00 UTC (04:00 and 00:00 LT), while MO holds steady (Fig. ).

For a given LPlume, MO has more successful plumes than LO (Fig. a–c). MO plume success appears to increase more with LPlumeCloud than for LO (i.e., the rightward shift of the MO LPlumeCloud distribution and divergence after ∼ 100 m, b). Notably, this organizational difference is not apparent in the unsuccessful, LPlumeClear distributions (a). We compute the plume success fraction (c) as the ratio of number distributions between the LPlumeClear and LPlumeCloud to clarify this tendency. Plume success increases with LPlume for both organization types, but the MO success rate increases far more rapidly, diverging from LO after 1 km. This organizational difference in success rate is another potential marker of cloud-layer-driven mesoscale circulation, e.g., through the ascending circulation branch cohering plumes into stronger updrafts and lowering the LCL through converging moisture (Fig. a, b). For a given plume width, MO clouds likely have deeper (e.g., Fig. c) and moister clouds due to the moisture–convection feedback. As plumes increase in size, potentially through turbulent enhancement by mesoscale circulations, they can also support larger clouds. Thus, MO clouds for LPlumeCloud > 1 km have an exponentially greater potential to generate heat through condensation release, which will further reinforce mesoscale circulations, plume success rate, and organization.

Second, we evaluated the total (c, LChord) and CB core (d, LCBCore) lengths and found organizational differences apparent in their distributions as well. Specifically, MO clouds tend to be generally larger than LO clouds but with substantial overlap. MO clouds occur more frequently with larger LChord and less frequently with smaller LChord than LO (i.e., MO shifted to the right of LO, c). This also holds for LCBCore with large cores occurring more frequently for MO (d). The CB core cloud amount from LCBCore (∼ 7 ± 2 % for LO and 8.9 ± 0.6 % for MO, Fig. b) is, as expected, smaller than the amount derived from LChord, but both CB and CB core amounts are statistically similar between MO and LO on average and both exhibit persistent MO and an apparent solar burn-off/recovery cycle for LO. Note that LPlumeCloud tends to be wider than the cloud and core lengths they support (i.e., LChord and LCBCore distributions are shifted left relative to the plumes, b–d; relationship is not 1:1 in Fig. a). Recall that LPlumeCloud pertains to the maximum width of the contiguous region, not the plume core, and likely includes turbulent tendrils.

The tendency toward larger LChord and LCBCore for MO clouds is consistent with the expectation that they have undergone more organization. Organizational differences manifesting primarily in LPlumeCloud indicate that plumes may be reinforced by a cloud-associated process, potentially including cloud-layer-driven mesoscale circulations (i.e., Fig. ). The reinforced plumes likely feed back on clouds through supporting stronger cloud updrafts in the ascending branches of mesoscale circulations, increasing cloud success rate and helping to cluster clouds further through the moisture–convection feedback (i.e., by lofting more moisture into cloud, generating more condensation, and strengthening the moisture-converging mesoscale circulations, Fig. a, b).

To understand whether the organizational differences in plume behavior impact CB dynamics and thus CB mass flux, we contrasted the relationships between LPlumeCloud with LCBCore and wCBCore (Fig. ). Because of the different plume width distributions between LO and MO (Fig. b), the CB core variables have been binned into quantiles by LPlumeCloud for ease of comparison. Linear regressions are performed on the underlying data. Hypothetically, for a given LPlumeCloud, a stronger relationship with MO LCBCore would indicate that wider plumes support larger core area, e.g., through more cloud aggregation in ascending circulation branches. A stronger relationship with MO wCBCore would suggest that mesoscale circulations affect the updraft dynamics directly, e.g., contributing dynamically through strengthening the updrafts in ascending branches. Wider plumes associated with more organization could be supporting both of these effects, modifying CB mass flux in two ways.

It is apparent that both LO and MO clouds have statistically indistinct linear relationships between LPlumeCloud and LCBCore (Fig. a; LChord has similar behavior, not shown, with R2 (p) = 0.594 (0.0) and 0.25 (0.006) for MO and LO, respectively). The increased frequency of larger LPlumeCloud for MO (i.e., wider quantile distribution and Fig. b) is consistent with the marginally more frequent occurrence of larger LCBCore (Fig. e). Based on regressions on the quantile binned values, more variance is explained in LCBCore by LPlumeCloud for MO (73 %) than LO (34 %). While this indicates that plume width more directly translates to core size in MO, the similarity in relationships between LO and MO suggests that core size has not been significantly modified by organization effects (e.g., cloud aggregation) at this stage of cloud development (Fig. a, b). More organized stages beyond MO may see a larger impact (Fig. c).

The organizational differences are much larger in the relationship between LPlumeCloud and wCBCore (Fig. b). Best-fit linear regressions are statistically distinct between LO and MO for the range encompassed by the majority of the observations (LPlumeCloud through ∼ 6 km). The variance explained in the quantiles is substantial for both MO (74 %) and LO (60 %). The offset between the LO and MO lines is a clear indication that for a given LPlumeCloud, MO clouds have stronger wCBCore than in LO clouds. These comparisons mark an important finding: for a given plume width, and thus core length, MO clouds achieve stronger CB core velocities and have increased mass flux into the cloud layer. This may be a manifestation of returning cloud-layer-driven mesoscale circulations boosting updraft strength in their ascending branches (Fig. a, b).

Finally, we evaluate these organizational differences in cloud behavior using the classic cumulus-valve phase space from the literature (e.g., : wCBCore vs. LCBCore) (Fig. ). Because the range of LCBCore is more similar than LPlumeCloud (Fig. b vs. e), we can use the same quantiles for LO and MO LCBCore and use the Mann–Whitney U test to statistically evaluate distribution overlap in each quantile. This strategy is also advantageous since the LCBCore vs. wCBCore relationship is distinctly nonlinear. Note that linear regression fits are statistically well separated for MO and LO (Fig. ). Regression on LO and MO quantiles indicates that 49 % and 66 %, respectively, of the variance in wCBCore is explained by LCBCore (separate LO and MO quantiles have lower correlations of 40 % and 59 %, Figure ).

We find, in agreement with prior evaluations (Sect. ), that stronger wCBCore is associated with more cloud at CB (i.e., greater LCBCore). However, there are two important nuances revealed in Fig. : (i) MO clouds have consistently larger mean core updrafts (i.e., the MO curve is higher than the LO curve, consistent with Figs. e and b), and (ii) the relationship between LCBCore and wCBCore increasingly diverges across organization states with increasing LCBCore. The MO and LO populations are statistically distinct at 95 % for the majority of the LCBCore bins, especially for core lengths greater than ∼ 250 m.

In short, we find that MO clouds are more “dynamically efficient”: for a given core size, updrafts are much stronger for MO clouds compared to LO clouds. We hypothesize that this is due to the returning, ascending branch of mesoscale circulations strengthening the updrafts of clouds as well as aggregating moisture and, potentially, widening turbulent plumes (Fig. ). The increasing separation with increasing LCBCore is consistent with the expectation that mesoscale variability contributions become more significant at larger cloud sizes , emphasizing the importance of understanding organization influence on cloud behavior. We further see that, as suggested in LCBCore composites (Fig. ), LO wCBCore flattens out at larger LCBCore, while MO continues to increase (though at a slower rate of increase than LCBCore ≲ 250 m), apparently less assisted by the cloud-layer-driven phenomenon or less constrained by the environment. Whether this inhibition of LO may be set by its environmental conditions is examined in the next section.

In this section, we contrast thermodynamic and environmental conditions across organizational states. This allows us to infer how these conditions may be conducive to supporting dynamically more efficient organized clouds. We focus on cloud controlling factors that are likely influential to Cu: wind speed, heat fluxes, air and sea temperatures, stability, and vertical moisture profiles. We additionally examine diurnal variations for indications of causality in the mesoscale organization cycle. While this is a correlative framework, understanding the evolution of cloud dynamics and their environment across the diurnal cycle is insightful for interpreting some causal aspects of the cloud systems.

MO clouds occur at higher wind speeds throughout the depth of the BL (Fig. a), extending to the near surface (Fig. d, e). MO and LO have statistically different distributions at every altitude at 95 % confidence (a, d). Surface wind speed differences are consistent with climatological sugar (i.e., LO) and gravel (i.e., MO) behaviors . LO has a bimodal shape including a small probability of higher surface wind speeds aligning with the MO mode. This is from 9 February (not shown, see also Fig. 4, ) before LO clouds transition toward MO clouds. However, the LO and MO distributions are still statistically distinct and LO is lower on average (d).

Wind speed increases sharply from the surface, then stays relatively constant until above the mean CBH (horizontal lines) where it begins to decrease with height (a). This indicates there is no shear in the sub-cloud layer 100 m above the surface, allowing plumes to develop with limited interference for both MO and LO. Backwards shear manifests in the cloud layer and above: ∼ 2 m s⁻¹ km⁻¹ between 1 and 3 km (typical for this season in the trades, e.g., ). Shear has a small amount of diurnal variability but is statistically indistinct between MO and LO (not shown). Wind direction is dominated by westward flow with some diurnal variability (not shown). Directional variation is also seen with height , becoming more easterly and northerly (veering north for LO, while MO is maintained more southward).

The LO and MO composite profiles (b–c) and surface (e) winds are statistically distinct for all hourly ranges across the diurnal cycle. Notably, LO surface winds tend to increase gradually over the day, while MO winds are highest at night, declining over the day. This is consistent with 10 m wind speed cycles observed at the NTAS buoy for sugar and gravel, respectively, although MO has a higher magnitude and later peak than gravel . The gradual increase in LO surface wind peaks at the end of the day (18:00–24:00 UTC) before dropping overnight, similar to the subtle pattern in LO wCBCore mean and variance (Fig. c, f). The MO wind speed profile and surface are largest overnight (00:00–12:00 UTC) before being minimized during the day (12:00–24:00 UTC). The end of this extended maximum aligns with the peak in MO wCBCore mean and variance at 06:00–12:00 UTC (Fig. a–c, f).

The stark wind speed differences are translated into the heat fluxes (Fig. ): MO PDFs are shifted to larger latent (LHF, a) and sensible heat fluxes (SHF, b) compared to LO. This is also true for their combined buoyancy flux (Fig. a–b, same PDF and cycle shape as SHF). We find that MO PDFs and diurnal cycle composites are statistically distinct from LO distributions. The shapes of the LHF and SHF diurnal cycles are more distinct than those of the wind speed (Fig. e), peaking for MO (05:00–10:00 UTC) and increasing to maximum for LO (∼ 07:30 UTC) at the end of the night. SHF has a sharper cycle and larger relative amplitudes compared to LHF, maintaining a magnitude offset between MO and LO, but their phases are more alike than for LHF.

The shapes of the LHF and SHF cycles are a result of the conflation between the air–sea temperature difference (Fig. d), surface humidity (Fig. e), and wind speed cycle (Fig. e). Air temperature responds more promptly to the diurnal cycle (increasing to maximum at ∼ 12:30 UTC, Fig. b), while sea surface temperature (SST) lags behind, likely due to a larger heat capacity, and is warmest at day's end and overnight (Fig. d). While the magnitudes are higher, these cycles are generally consistent with the climatological cycles seen for sugar and gravel . Their combination leads to an air–sea temperature difference that peaks at the end of the night and is at a minimum around ∼ 12:30 UTC for MO and LO (Fig. d). LHF has a more gradual cycle and follows wind speed more closely. LO and MO specific humidity at 10 m (q10m) is a minimum at the end of the night (Fig. e), aligning with the peak in wind speed and thus LHF. LO LHF is relatively constant over the course of the day, while MO LHF decreases at the end of the day before its nighttime increase.

MO LHF (SHF) is larger than for LO due to the combination of both larger magnitudes and diurnal cycle phase alignment between U10m and q10m (SST-T10m for SHF). The SHF and LHF early-morning peak matches the location of the MO wCBCore variance peak and the beginning of the recovery from a nighttime lull in mean wCBCore. There may be an accompanying increase in LO turbulence at CB, but it is not statistically distinct. This suggests that there is some boost to the updraft turbulence for MO, and possibly LO, from these turbulent energy fluxes. However, there is evidently another contributor to the MO cycle that helps to maintain the relatively constant MO wCBCore during periods of declining fluxes and winds (Fig. a–c, f). The LO wCBCore cycle, which holds steady but may peak at the end of the day, seems to align more with the larger daytime magnitudes in its associated wind and LHF cycles, suggesting some environmental support (c, d–f).

Thermodynamic profiles often modulate cloud systems' ability to develop and persist. We find that SST and T10m are both larger for LO and statistically distinct from MO cycles (Fig. b, d) and PDFs but not their interquartile range (a, c). These temperature differences are also consistent with the climatological separation between sugar (i.e., LO) and gravel (i.e., MO) . Higher SSTs are often associated with increased dry-air entrainment and reduced low cloud amounts in stratocumulus regions (e.g., ). This appears to hold here as well (Fig. d). The combination of higher SST and smaller air–sea temperature differences and wind speeds, and thus smaller fluxes, likely damps LO plume strength and success (e.g., Fig. b, c). The MO–LO temperature offset extends through the depth of the BL with LO clouds tending to occur in warmer environments, with larger and statistically distinct potential temperatures compared to MO (Fig. a–c). LO cloud development is likely impaired through entrainment of warm, relatively drier air compared to MO (e.g., Fig. a, discussed more later). Note that both MO and LO sub-cloud layers are well mixed (i.e., constant θ with height until after CB), facilitating cloud development through even moisture distribution. MO θ profiles are relatively consistent across the diurnal cycle (b). However, LO above ∼ 1 km has a substantial diurnal cycle with θ increasing over the day (Fig. b), peaking at 18:00–24:00 UTC and being minimized at 06:00–12:00 UTC. This aligns with the T10m cycle but is a much larger magnitude (Fig. b). The relative magnitudes and cycles are consistent with θ profiles derived from reanalysis for sugar and gravel clouds .

The amplification of the LO θ cycle above cloud may be associated with increased shortwave absorption from increased concentrations of absorbing aerosol on most of the LO days (recall LO on 30–31 January and 1, 4, and 9 February). Specifically, identified a layer of absorbing aerosols aloft (29 January–3 February and 9 February) that were a mixture of biomass burning and dust transported from Africa to the RHB between 1 and 3 km in altitude. This aerosol layer gradually mixed into the BL as it was transported toward Barbados, increasing aerosol extinction throughout the atmospheric column and absorbing aerosol concentrations at the surface (e.g., Fig. a–c) . 29–31 January and 1–3 February were the exception to this, with lower correlations between aerosol properties at the surface and aloft . More absorbing aerosols aloft could increase the amount of shortwave absorbed over the diurnal cycle, heating above cloud and stabilizing the BL. found in their LES study that the aerosol layer present on 31 January–2 February also deterred organization through longwave effects. During EUREC⁴A, low-level, longwave radiative cooling was found to depend on the ratio of relative humidity between the BL and FT and was influenced by moist intrusions at the mid-levels .

Sub-cloud air temperature is more complicated to understand. There was no precipitation at the RHB during the LO period (Fig. d–e), enabling the aerosols to persist in the sub-cloud layer. If sufficient shortwave made it through the aerosol and cloud layers aloft, this could contribute to the larger LO T10m cycle (Figs. c, b). LO has lower total BL CF (Fig. d) along with higher surface humidity (Fig. a, c, e), potentially assisting in warming the sub-cloud more than for MO. The result of these effects may be the sharper daytime increase in LO LCL and thus CBH (Fig. a–b, horizontal lines in Fig. c), which would further deter cloud development. In contrast, the MO period was dominated by marine, non-absorbing aerosols that experienced some precipitation removal (e.g., coinciding with times of increased precipitation over the diurnal cycle, Fig. b, d–e), higher CF, and lower specific humidity, potentially damping any atmospheric heating signatures.

No matter the driver of the diurnal cycle in θ, it has a notable effect on the lower-tropospheric stability cycle (LTS = θ700hPa-θSLP, , Fig. e). LO environments are more stable and statistically distinct from MO environments when examined in aggregate (d), consistent with the climatological sugar (i.e., LO) and gravel (i.e., MO) comparisons . The MO and LO cycles are also similar to the gravel and sugar cycles, respectively, but have larger differences in magnitude and a less varied cycle for MO . Notably, MO and LO have the same stability at the beginning and middle of the day (∼ 10:00–15:00 UTC, e). This overlap corresponds to the period of increased winds (Fig. e) and fluxes (Figs. b, d, b) and occurs before re-stabilization of the LO BL, whose inversion appears to have degraded overnight.

Increased stability over the majority of the day presents an additional deterrent to LO clouds developing as their plumes will need to be sufficiently energetic to reach the simultaneously lifted LCL (Fig. a) and form a cloud. The brief drop in stability likely allows LO plumes to have some success early in the day, shortly after the winds and fluxes have peaked and more turbulent energy is available in the system. Similar behavior has been seen in the southeast Atlantic in the presence of smoke in the BL . This also implies that persistent LO clouds are likely sub-selected for stronger wCBCore that can better resist this environmental deterrence (Fig. c).

Even with this sub-selection, the stronger capping inversion for LO clouds may explain the flattening of the wCBCore vs. LCBCore curve (Fig. ). Specifically, greater stability in LO will restrain clouds from deepening as much as in MO conditions, reducing their geometric depth (Fig. c) and liquid amount. We hypothesize that this damps their ability to release heat through condensation, which impairs both the local buoyancy enhancement from heating and the generation of mesoscale circulations, resulting in less updraft strengthening for LO clouds than MO clouds (Fig. a, b). The divergence between LO and MO curves grows with LCBCore, consistent with deeper clouds being more enhanced. The opposite organization stability tendency has been shown in LES where a reduction in stability enhanced the transition from sugar to flowers and produced larger cloud features . Further evaluation of the connection between environmental controls, dynamic efficiency, and mesoscale organization mechanisms is warranted. A more detailed evaluation of the LO evolution during the 30 January–2 February period at the RHB, highlighting the aerosol impacts mentioned here, is in development. The nuances of meteorology–aerosol covariability impacts on organized cloud systems, including the prevalence of LO clouds occurring under mixed dust and biomass burning aerosol conditions, is also worthy of future investigation (e.g., using the long record at BCO and following ).

MO clouds exist in consistently more unstable environments (Fig. d–e), which is conducive to the development of deeper clouds and likely reinforces their moisture–convection feedback through facilitating lofting of humidity into the upper BL. A marker of the deeper MO clouds can be seen through comparing CBH measurements from the Doppler wind-lidar and the ceilometer (Fig. b–c). MO Doppler lidar CBH PDFs are shifted toward larger CBH and statistically distinct from LO. The mean and interquartile range overlap, though, with most CBH falling between ∼ 600 and 800 m for both (Figs. b, also a–c and sonde plots throughout). However, Doppler cloud identifications are restricted to have CBH near LCL (a), effectively removing any information about detraining cloud layers at the top of clouds or cloud edges as clouds tilt with height under shear. In contrast, the ceilometer CBH measurements are not restricted and capture a significant tail at upper levels for both MO and LO (Fig. c). This tail is more extensive for MO than LO, indicating that they consistently achieve greater heights (up to ∼ 2.5 km) compared to their less organized counterparts (∼ 1.5–1.8 km). Geometrically deeper clouds imply greater liquid water in the column and thus larger optical depth. Relative humidity profiles (Fig. ) roughly support these cloud depths (Fig. c): 80 % until ∼ 2 km for MO and ∼ 1.3 km for LO on average. Merging of convective updrafts in organized tropical deep convection enables deeper clouds , consistent with the wider plumes and deeper extent of MO clouds. We also find that the total BL cloud amount (ceilometer measured CBH ≤ 3 km, Fig. d) is statistically larger for MO than LO (40 ± 4 % vs. 22 ± 6 %) despite their similar CB amounts, implying larger MO clouds potentially with detraining layers (a–c). Thus, MO clouds' greater dynamic efficiency likely supports a larger radiative impact due to their greater optical depth as well as their larger cloud amount (e.g., gravel vs. sugar, ). This potential connection between dynamic efficiency and impact on the radiation budget is worth investigating in future work.

To evaluate the prevalence of the moisture–convection feedback for MO cases and the type of air being entrained in both organizational states, we next examine the specific humidity profiles (Fig. ). Unlike other variables, q has distinctly different organizational tendencies in the upper and lower BL (a). Above ∼ 1.3 km, MO q is much larger and from a statistically distinct distribution compared to LO. Moister FT (700 hPa) environments have also been seen for gravel (i.e., MO) than sugar (i.e., LO) . Moisture advection likely does not explain the increased q aloft for MO vs. LO. The 3 d Lagrangian back trajectories of q700hPa are persistently drier for gravel than for sugar up until the final ∼ 12 h before the cloud observation .

In the sub-cloud layer the opposite behavior is true: MO is substantially drier than LO below CB. This is particularly apparent in q10m (d–e). These statistically distinct separations hold across the diurnal cycle at the majority of altitude bins (b–c) and at the surface, with LO recovering more than MO by the end of the day (e). MO has a smaller magnitude than but similar diurnal cycle amplitude to LO at the surface and below CB, with both gaining the most moisture by the end of the day (18:00–24:00 UTC). MO q has a roughly similar diurnal cycle above CB as below (b). In contrast, LO q peaks by the end of the day in the sub-cloud layer and between 0:00 and 12:00 UTC in the middle BL. This may be linked to the period of greatest instability and largest fluxes, facilitating more active lofting of moisture (e.g., a minimum in q10m, e). A drier upper BL for LO clouds means that cloud-top entrainment will introduce more dry air than in MO clouds, additionally deterring LO cloud development. On the other hand, entraining moister air will help to maintain MO clouds for longer. Increased dry-air entrainment for LO clouds is also consistent with the previously discussed expectations associated with the higher SSTs for LO clouds.

The deeper and persistently moister layer aloft for MO is consistent with more moisture lofted into the upper BL through increased wCBCore due to greater dynamic efficiency. This coincides with a critical part of the moisture–convection feedback, increasing aggregation of moisture into the ascending, cloudy branch as organization increases. Our analysis highlights a further nuance: the MO sub-cloud layer is drier than for LO clouds. This presents an additional barrier for MO clouds as it raises their LCL compared to LO clouds (Figs. a, a). The relative decoupling index (CBH–LCL/LCL, ) is marginally higher for MO than LO although it is less diurnally variable (Fig. d). However, the average relative decoupling instances are both well below 0.1, indicating that MO and LO are both well coupled to the surface and that this is a small albeit statistically significant disadvantage. While the challenge for MO clouds to develop may be marginally increased, this could be ultimately helpful for increasing the strength of the CB updrafts (e.g., environments with higher LCL have wider, deeper, and stronger cloudy updrafts, ). In summary, the opposing q tendency in the upper and lower BL across organizational states is potentially an indication of the efficiency of moisture export to the upper BL through the stronger MO cloud updrafts, which more effectively remove sub-cloud moisture over time compared to the LO clouds. Note that on average the more organized systems downwind near Barbados exhibit the theoretically expected moisture enhancement throughout their ascending circulation branches including sub-cloud . Contrasting cloud organization stages and their vertical moisture behavior could provide insights into processes dominating mesoscale evolution, i.e., between early stages dominated by local cloud processes and “dynamic efficiency” (MO, Fig. b) vs. later stages where mesoscale circulations have strengthened enough to enhance moisture throughout the column and restore any previous depletion sub-cloud (Fig. c).