 Okay, great, thank you. Well, thank you so much to Adrienne and Simona for inviting me to speak as part of this seminar series. And today I am gonna be presenting some results on convective self-aggregation and climate sensitivity in our RCE intercomparison that Adrienne mentioned. And so I need to add many more names to this title slide to reflect all of the people that have contributed to that effort, especially my graduate student, Catherine Stoffer, as well as Tobias Becker who led the climate sensitivity proportion of this work. So I'll start by means of motivation by pointing out that sometimes tropical convection is isolated or scattered as in this image that I took from the Otrick Field Campaign where we have these sort of individual convective cells surrounded mostly by a clear sky. But often convection is not isolated in that way, whether it's shallow convection, like shown here in the satellite image or a deep convection in this photograph of a mesoscale convective system from that same field campaign. So convection in Earth's atmosphere organizes on a variety of spatiotemporal scales from the mesoscale, like features like squall lines and mesoscale convective systems to the synoptic scale with phenomenon like tropical cyclones and equatorial waves in the tropics and even to the planetary scale with phenomenon like the Madden-Julian Oscillation. And this organization occurs by a variety of different mechanisms including vertical wind shear which helps organize squall lines, sea surface temperature gradients which organize some of the large scale modes of organization like the walk or circulation as well as a whole host of dynamical distributions like some of those equatorial waves that I mentioned before. Now this organization of convection is not just a spatial reorganization that makes for pretty pictures but it actually has a profound influence on the large scale environment. So observations show that a particular cloud regime associated with organized deep convection contributes about half 47% of all total tropical rainfall as well as a significant fraction of cloudiness. Additionally, these observations show that most of the regional increases in tropical precipitation are associated in an increase in the frequency of this regime of organized convection. We also find evidence and observations that the mean state under more aggregated conditions is drier considering both individual convective scenes as well as monthly variations at the scale of the tropics. And so on the left panel here as we go from red to black we're going to more aggregated scenes of convection and seeing a drier relative humidity. In the right panel, this is using a metric of tropics wide aggregation and on the right as it's more aggregated the mid troposphere humidity is lower for monthly timescale variations. We also see effects on cloudiness and so aggregated states as we go from the black to the red have fewer high clouds and often are also observed to have more low level clouds as we go from the blue to the red they're becoming more aggregated though this is not necessarily true at all scales. These effects on the precipitation the humidity and the cloudiness mean that organized convection has a big effect on both the tropical and climate and global kind of energy budget. And so if we again are considering sort of individual scenes of convection from satellite imagery regardless of whether we're considering 10 degree or three degree boxes we find that as the scene is more aggregated so with going towards the left on these plots the outgoing long-grip radiation is greater as those sort of drier less cloudy areas are able to cool more to space. These more aggregated scenes also have more tropospheric rate of cooling as well. And in addition if we again consider monthly variations of tropics wide aggregation we also see a signal in the actual top of atmosphere energy budget with more net top of atmosphere cooling when conditions are more aggregated. Though the effect on the top of atmosphere energy budget does seem to have some sort of scale dependence in it as well but nevertheless this is strong evidence that in observations organized convection of many different types has an impact on the large scale environment and energy budget. Now one type of convective organization is known as self-aggregation. And self-aggregation is a spontaneous transition from randomly distributed like in the picture on the left to organized convection that occurs in idealized numerical modeling simulations despite those simulations having totally homogenous boundary conditions and forcing. And so these are snapshots from a cloud resolving model simulation and it has constant SST and constant solar insulation. And at the beginning of the simulation on the left as indicated by OLR the convection is sort of quasi-randomly distributed in space and time. But later on in the simulation it becomes organized into in this example one single intensely precipitating moist clumps surrounded by clearer, drier cloud-free slowly subsiding air. And this process results from interactions between convection and its environment. Again, there's nothing that says just from the boundary conditions that we need to have convection here there's no large scale forcing. And so the only way that this organization can occur is through feedbacks involving clouds, water vapor, radiation, surface fluxes all of these sort of interactions between convection and its environment. And this type of localization spontaneous localization of convection was first seen in a paper by Held et al in 1993. And I've written a couple of reviews on it in the last couple of years. One of the tools that we use and then I'm going to use in this presentation to investigate self-aggregation are simulations of radiative convective equilibrium. And this schematic on the left here shows the different components of radiative convective equilibrium. We have shortwave radiation from the sun which heats the surface and atmosphere the surface and atmosphere then cool by emitting longer radiation. And if that was the only thing going on we'd be in a state of radiated equilibrium but it turns out that that state is strongly unstable to convection. And so we have turbulent convective motions and fluxes of latent and sensible heat from the sea surface to the atmosphere. And then vertical convective heat transport in the rising warm moist air parcels in clouds. And so what radiative convective equilibrium is then is a statistical equilibrium where on average we have a balance between the net radiative cooling of the atmosphere and heating from deep convection. And this framework is really I like this quote from one of Chris Bretherton's papers it's a time honored idealization for understanding the tropical atmosphere and its sensitivity to relative to relevant forcings. And it really is the simplest possible way to phrase many important questions about climate including questions about the response of convective clouds to warming and questions about aggregation of convection and climate. But there's not necessarily a standard way to set up these radiative convective equilibrium simulations and with this great renaissance of using RCE over the last 10 years many people are using this framework but configure it in slightly different ways which makes it difficult to interpret and compare different studies and results to each other. And so we felt like it was really the time to establish a common baseline for comparison and that motivated us to organize this radiative convective equilibrium model intercomparison project where we really would then be able to truly assess the robustness of the RCE state and address these questions about clouds and aggregation and climate. One of the advantages of RCE as a framework is that it's accessible by many different types of models. So there's a wide variety of model types in which you can simulate that. And so we really leverage that possibility in our RCE MIP ensemble which I'll describe in a moment. So the questions that we wanna address both in RCE MIP overall and in this talk in particular there's several things to address. One, the kind of most basic is do the models actually agree on what the simulated RCE state is if they're provided with the same boundary conditions, the same radiatively active gas concentrations things like that same SST, do they simulate the same climate? Do they exhibit convective self-aggregation? Do they agree on how strongly convection self-aggregates? And then if they do self-aggregate how does that self-aggregation impact the mean state in the models? We're also interested especially because of its potential influence on climate on how self-aggregation may change with warming which is something that is sort of not fully understood in the literature currently. And then we wanna link this all together with climate and climate sensitivity. And so ultimately we like to know if self-aggregation impacts climate sensitivity or the response, how the climate would respond to an external forcing. But first we need to actually assess what the climate sensitivity is in these RCEMIP simulations and then ask whether or not the inter-model spread in self-aggregation, if there is one, if that can explain the inter-model spreading climate sensitivity. So first I'll describe the RCEMIP protocol. Again, as RCE is accessible to many different types of models we were able to have participation from about 30 models in this project. And they include models of a lot of different types. So both GCMs and single column models with parametrized convection as well as models with explicit convection like cloud resolving models, global cloud resolving models and large eddy simulations. In all cases we configure the models in their aqua planet framework. So no continents, no land, just ocean everywhere with uniform insulation. So every grid point receives the same amount of insulation and a uniform fixed sea surface temperature. These simulations are not rotating and the motion is just initialized from random, small perturbations, random noise in the boundary layer temperature field. We then have two sets of simulations. The first are our so-called small domain simulations which are 100 kilometer square for cloud resolving models which are run at one kilometer resolution for this configuration or for LES 200 meters. And it's then the equivalent for the parametrized convection models is either a single column or a very small earth. And we do this at three different values of the fixed sea surface temperature 295, 300 and 305 Kelvin. And these simulations, which an example is shown on the right for the icon model with the cloud resolving intermediate version that has more vertical levels and then the LES on the right. These tip simulations as I'll show typically are expected not to have aggregated convection in them. We then have a large domain configuration which is the global domain for our global models is shown here for an example for a GCM and for a global cloud resolving model which has a smaller earth radius. And so that's it to scale down here. And then for cloud resolving models we use a 6,000 by 400 kilometer rectangle with three kilometer grid spacing. We also have one additional configuration which uses the Worf model in the sort of rectangular domain but with course resolution and parametrized convection. And so that's sort of bridging the gaps between these two different regimes. All right, so now I'd like to show you what actually the simulations look like in these different scenarios. And so these are the small domain simulations for the SAM model. On the left we have the cloud resolving configuration at one kilometer resolution. The middle has about double the number of vertical levels and then the right is the LES configuration that has that higher number of vertical levels as well as 200 meter horizontal resolution. The top row shows the simulations at 295 Kelvin, the middle at 300 and the bottom at 305. And then in each the left is OLR showing where our deep convective clouds are and then precipitable water. So as you can see and I'll play this again, the convection is pretty much randomly distributed in space and time in all of these simulations. There's not really any noticeable differences between the different configurations other than that, there's obviously more detail in the higher resolution ones. The precipitable water looks pretty boring because this is hourly average precipitable water and it's pretty uniform spatially because the convection again is randomly distributed. And again, we have this kind of configuration for all of our different cloud resolving models. So going to all of those different cloud resolving models for this small domain configuration, this is just a snapshot towards the end of the simulation for again OLR on the left, precipitable water on the right. And we see that obviously there's variability in the cloud structures but nearly all of these small domain simulations are unaggregated, which you can see if you played movies of all of them or just kind of by the fact that the precipitable water is pretty, pretty uniform. The one exception is this one particular version of the UK Met Office model, which does actually sort of semi-aggregate, which was surprising that it did so on such a small domain size. Again, just emphasizing the value of having such a large ensemble of models so that we have really a better grasp on the robustness of these sort of features. So if we then turn to our large domain simulations, we have on the left here, the cloud resolving model configuration for SAM, one of the models. Again, 295 on the top, 300 in the middle, 305 on the bottom, OLR and precipitable water. And then one of the global cloud resolving models, so SAM, the same model in its global form. And you can see here that now the behavior looks quite different than in those small domains. It's apparent, especially in the precipitable water where we have these regions of very persistent dry in the yellows and green sort of areas and moist areas in the darker blues, especially you can see that in the ones on the left where it's kind of these alternating moist and dry bands. And so this is convective self-aggregation where our convective clouds are limited to these sort of moist bands and then it's clear in between. And there's obviously still a great deal of variability within those kind of moist envelopes, but they remain sort of spatially and temporally coherent. So they move around, they have variability but the convection is still localized in those different areas. And these are really fun movies to watch. You can also see, of course, that there are some differences in the structures as we look at cross-differency surface temperature regimes. So looking at that view for all of our different clover solving models, so these are just the CRMs on the rectangular domain. Again, OLR on the left, precipitable water on the right, just a snapshot from the end of the simulation. We here now see that nearly all of our large domain simulations are aggregated where convection is in these sort of alternating moist and dry bands, which again is very apparent in precipitable water, but we can also sort of see in the OLR. The one exception is this one particular configuration of the Worf model, which with this particular choice of physics packages that was used, actually did not really aggregate on this domain. And so this sort of suggests that aggregation is quite a robust phenomenon. I mean, it occurs across, I don't know how many there are, more than 10 models here, but it doesn't necessarily occur strictly all of the time and so that's good information to know. But even though it occurs across most of these models, there is quite a great deal of diversity in the structure, the scale, how variable the aggregated regions are. And so there's really a rich sort of spectrum of possible sort of representations of the aggregated convection. This is what it looks like in one of the GCMs, one of the general circulation models. This is on the left cam five, in the middle cam six, just a little later version. And then a super parameterized version of cam on the right, again, simulations on the top at 295, middle 300, bottom 305 Kelvin. And so these simulations too are aggregated, but on somewhat larger scales, again, we can see differences with temperature and with the different versions of the model in terms of, again, the spatial structure and sort of visually, at least apparent, the degree of aggregation perhaps in these different simulations, some of them have kind of one giant cluster, some of them, it's more like you have interspersed tri-regions. And if we then kind of look at that over all of the different global models, we again see this rich diversity in the structure and patterns of convection that emerge, again, anything from basically one hemispheric spanning aggregated region to having multiple sort of smaller ones and every sort of possible thing in between, some of them have more kind of squall line structures. It's really dynamic and fascinating. These little tiny ones here, those are the global cloud resolving models which are to scale, but then blown up here so that you can see the different patterns and all of these global models are aggregated to some extent though, the extent varies which we're gonna quantify later on. So with regards to our first question, do models agree on the simulated aggregated state? We can assess that by looking at mean profiles of different quantities across all of these different models. So these are domain averages of temperature profiles where the black line here shows the ensemble mean, the blue shows the full range of the ensembles and the yellow line shows the interquartile range. And as we can see, the temperature structure looks roughly similar across all the different models. You can't tell it from this figure, but the temperature profiles are in fact very close to a moist adiabat though they are slightly systematically cooler than the moist adiabat as we would expect if convection is following a sort of zero buoyancy and training plume sort of framework. But there is quite a degree of spread across the different models. If we look at one particular height, there can be a spread in temperatures of up to 10 Kelvin which is a pretty big difference for a sort of tropical simulation. If we look at relative humidity, we see a very widespread across the models. I should have mentioned the small domain simulations are the top panels here and the large domain simulations are the bottom panels here. In our small domain simulations on the top, the relative humidity in the mid-troposphere can be anywhere from 25 to 90% across the ensemble. So that's a pretty big difference and relative humidity is something that of course has a lot of importance in climate and so the fact that the models differ so much is really fascinating. If we go to the large domain simulations on the bottom, which is where most of them are aggregated, there's a bit more of a consistent shape to the relative humidity profiles and somewhat better model agreement but actually there still is quite a large spread. So it says that the presence of aggregation and its associated circulation might help constrain the models a little bit versus in the small domain simulation where the convection really is pretty much just random and they can really disagree quite a lot on what the resulting average relative humidity profiles are. These are all also averages over the last 25 days or so. Well, everything except the first 75 days of the simulation. So last 25 days for cloud resolving models and last year or two for the GCMs. What about other variables? If we look at clouds, because that's also the main thing that these types of models are wanting to simulate correctly, also a huge spread. So the left panel here shows cloud fraction which is diagnosed either from a cloud scheme of the model employs it or a threshold value of cloud condensate. And across the small domain simulations, we really have clouds that are ranging from very small amounts of say high cloud to domain filled by high cloud. Though these are typically very, very small amounts of cloud condensate, very thin clouds. Over the large domain simulations, a bit better agreement, but still a widespread. And we see these widely varying amounts of cloud in terms of the actual average amount of cloud condensate as well, which is shown on the right panel here. These are profiles of total cloud water. Again, just really big differences across the models. We can also separate this out in terms of splitting up parameterized from explicit models. And we've done that for all these variables, but I'm not showing it because there aren't really any consistent differences. Meaning the models with parameterized convection differ from each other pretty much as much as the models that with explicit convection do. So just resolving convection does not all of a sudden make things agree better. One thing that model, the one thing that models do agree on more is how things change with response to warming across our simulations at different SSTs. And so across the majority models, we find that anvil clouds rise at a rate about 0.3 kilometer per Kelvin of warming. They warm slightly at a rate of about 0.4 Kelvin. So this is the change in the anvil cloud temperature. So an increase of 0.4 Kelvin per degree of SST warming, that's roughly consistent with what you would expect from the proportionally higher anvil temperature hypothesis. And we also find that across 70% of models, there is a decrease in the anvil cloud fraction. And so here the anvil cloud fraction is plotted as an anomaly from its value at 300 Kelvin. And we can see that not all, but most of the models have a decrease in that with warming, which is a sort of interesting result because there's been a lot of work in the literature recently about how anvil clouds change with warming. So with regards to this first question, we found that there's a widespread in the simulated RCE state, but better agreement in the response to warming and quite a bit of diversity in the spatial structure of convection. And so we're now going to move to talk more about aggregation specifically. And so in order to address whether the models agree on how strongly the convection self-aggregates, we need to really quantify it. And we use three different metrics to do so. The first is subsidence fraction, which is the fractional area of the domain covered by sinking air where we consider like a hundred kilometer size averages and daily averages of the vertical motion. And this is related to the transition of the velocity distribution to this situation where you have small areas of strong assent that are persistent in time surrounded by large areas of weak subsidence. We also use IORG, which is a clustering metric. And here we identify convective entities using an OLR threshold. We then compute the nearest neighbor distance for each of those convective entities and then compare that distribution of nearest neighbor distances to the theoretical expectation from a random distribution. And so for both the IORG metric and the subsidence fraction, if its value is greater than 0.5, it means the convection is aggregated and clustered and the larger the value is, the more clustered the convection is. We also use a metric related to the spatial variance of column relative humidity, which reflects the clear signature of self-aggregation and broadening the moisture distribution. And we use these three different metrics because they all capture kind of different aspects of the self-aggregation. And unfortunately at this stage, we don't actually really have an agreed upon metric in the community as for which is the best one. And so absent that, we thought it was prudent to use multiple and compare them to each other. So we've gone on ahead and calculated all of these metrics for all of the different large domain simulations. And this is what we find. So there's a lot of things on here. Let me explain the plot. We have all of our models listed on the bottom and we've organized it to have the clover solving models on the left half and the GCMs on the right half. The red circles show the subsidence fraction metric, the values of which are on the left axis here. And we've ordered this to sort things from low to high values of subsidence fraction. The blue squares show the IORG metric and its access is over here on the right. And then the green triangles show the column relative humidity variance whose access is here on the right. Again, for subsidence infraction and IORG greater than 0.5 means aggregated for the column relative humidity variance, there's not really a hard threshold. It's more just that larger numbers mean more aggregated conditions. We also have box and whiskers plots here that are showing the range of these different variables. So as you can see, these points are pretty much all over the place. And so there really is a wide spread in the degree of aggregation. And relative to sort of the average values of the metrics, there's a bit more variability in the column relative humidity variance than in the other metrics, but all of them have a quite large spread. But we do notice a few things. One is that the GCMs generally have lower values of IORG than the cloud resolving models do. These blue squares are a bit lower than here on the left. However, we don't really see that relationship in our other metrics. And so it's more likely an artifact of the fact that this IORG clustering metric is kind of difficult to calculate for models that have like course resolution and parameterized convection rather than an actual real indication that the GCMs are less aggregated. In terms of how the metrics agree with each other, we find that across all models, subsidence fraction and IORG, the blue and red, are weakly correlated with each other. But if we consider the CRMs separately, then all three metrics are strongly correlated with each other. And this is really reassuring because it says, even though these metrics don't necessarily always agree with each other, typically if one metric says that a simulation is strongly aggregated, the other metrics agree that it is also strongly aggregated, at least relative to the other models. And so that gives us confidence in their utility in this setting. If we look at just the GCMs, we see that the subsidence fraction and IORG metrics are strongly correlated with each other. There's not a correlation with the cloud resolving, sorry, the column relative humidity variance. However, if you look at that, it seems like this sort of group of models for the green triangles are a bit of an outlier here. And those are all these different versions of WARF that are parametrized, but actually on a rectangle rather than a sphere, which is sort of a strange kind of configuration. So if we kind of remove those, then we see that across the true GCMs, there is a correlation across all three of the metrics, which is good. So do the models agree? Well, they agree that convection aggregates, but again, IORG can range from very weakly aggregated to very, very strongly aggregated values of like 0.9. So there's a very widespread in the degree of aggregation across these different models. How to solve aggregation impact the mean state? We can address this by comparing pairs of small and large domain simulations first. So for a model that did both, we can look at the difference between the two because the small simulation is unaggregated and the large simulation is aggregated. So the difference then tells us the impact of self-aggregation. So if we do that for relative humidity, here we find that the mean state is drier when aggregated consistent with what has been seen in prior studies and in observations as well. And this is true both for profiles of relative humidity. So here again, we're taking the difference between aggregated and unaggregated pairs of simulations, where all the different colored lines are all the different models and they all lie to the left of the zero. So all the aggregated ones are drier. We see that as well in color-integrated metrics like precipitable water. So the domain average precipitable water is less when the convection is aggregated. If we look at changes in cloudiness, we see that there's a reduced high cloud fraction when convection is aggregated on the left panel here. The changes, oops, sorry, the changes in total condensed water are a little bit less clear. So the right plot here shows the condensed water path. So this is the vertical integral of all of the cloud condensate and the distribution of the differences between aggregated and unaggregated is sort of spanning zero. And that in part is because we have some compensating changes in high clouds and low clouds. So the high clouds decrease, but most models have more low clouds when they're aggregated, though it's a little bit more variable. And we can see that as well here in the actual profiles of cloud condensate, there's some variability, of course, but we have decreases aloft and increases in the middle and low levels. And these, again, changes are consistent with what we've seen in observations and other simulations as well. Now one theory for why a more aggregated state has fewer high clouds is given by the stability Ivers hypothesis of Boneydall from 2016. And according to this hypothesis, if you're in a warmer environment where you have increased upper level stability, that requires less clear sky subsidence per amount of radiative cooling to balance it. And so that results in less radiatively driven divergence and thus less anvil spread. And so that might actually explain why more aggregated simulations have fewer high clouds because our aggregated simulations are systematically warmer than the unaggregated simulations by several degrees in the upper troposphere. And this then results in the tropospheric lapse rate being more stable compared to the unaggregated simulations. And so this would then be consistent with these stability Iversite arguments. And so the decrease in the high cloud fraction in aggregated simulations is probably tied to the fact that those simulations are warmer, which is consistent with the fact that in an aggregated state, you're convecting from larger values of boundary layer entropy or also consistent with thinking of in-trainment being less effective at sort of reducing the buoyancy of your in-training plumes when you're convecting in a sort of moisture aggregated sort of situation locally. If we kind of then put these things together and look at the effective aggregation on the energy budget, we find that with aggregation we go again comparing aggregated simulations to unaggregated simulations. We have an increase in the column radiative cooling. Consistent with that, we have an increase in surface enthalpy fluxes. Again, this is for domain average. And thus we have an increase in domain precipitation. And so here this is showing the large domain average precipitation compared to the small and it is larger consisting with this change in the energy budget. In terms of the top of the atmosphere energy budget, we find that there is less net radiation into the top of the atmosphere. And that means that there is more cooling to space with the sign convention that these are defined as. So the more aggregated simulations, they open up these kind of bigger like clear sky, dry windows in which they can cool more effectively to space. And that is very important for the possibility of aggregation affecting climate and climate sensitivity. So with regard to how self-aggregation impacts the mean state, we find that models agree that self-aggregation dries and warms the atmosphere, reduces high cloudiness and increases the cooling to space. And so before we get to how convection affects climate sensitivity, one factor that might affect that is whether or not aggregation systematically increases or decreases with warming. And we can address that by using our metrics of aggregation to actually see what the rate of change of those metrics are across our simulations at 295 to 300 to 305 Kelvin. And so now similar plot to before, but what's plotted is the rate of change of each of the metrics rather than their actual values. And the basic thing to take away from this is that roughly half of these points are above the zero line, the dash line here and half are below. So there's really no consensus across the models in how aggregation changes with warming. Half of them have an increase in aggregation with warming and half have a decrease. And that's true regardless of which metric we look at. If you look at the box and whiskers plots here, all of them are pretty much centered around zero. And so this is a little disappointing because this was one of the big open questions about aggregation and we were really hoping that this RC intercomparison would tell us what was happening, but it sort of suggests that maybe the null hypothesis that aggregation does not change with warming is maybe still our kind of best answer. We do see some interesting differences in particular comparing the left and right hand side of the plots on the right hand side. The GCMs tend to have more of an increase in aggregation with warming than the cloud resolving models do. And that's one of the few really notable differences between the population of explicit and parameterized convection models. And yeah, something sort of interesting to point out which is gonna become relevant shortly for the climate sensitivity discussion as well. And you can see that again, most of these points on the right are more weighted above zero than on the left. One other thing to point out is that metrics can disagree with each other. So for example, here are a couple of models where some metrics say there's an increase in aggregation with warming and some say that there's a decrease which really does say that if you're examining any one model, it is important to use multiple metrics because they could lead you to different conclusions about how aggregation is changing with warming. Okay, so we've shown that across the RCMEP ensemble, there is a widespread in the changes in self-aggregation with warming. And so that brings us to our last question, does self-aggregation impact climate sensitivity? But first we have to ask, well, what actually is the climate sensitivity in these simulations? And since these are simulations at fixed sea surface temperature, we can estimate the climate sensitivity by calculating a SES type net climate feedback parameter by looking at how the net top of atmosphere radiative fluxes change with surface temperature across these simulations. And this can be converted into an approximate estimate of climate sensitivity by using a sort of typical value of radiative forcing. And we do this only just to give you a kind of number to compare against because I think usually we're more familiar with looking at climate sensitivity rather than this feedback parameter value. Of course, the actual radiative forcing is model and state dependent. And so if we look here for our small domain simulations which do not have aggregation, we see that their values of this climate feedback parameter would be consistent with pretty high values of climate sensitivity. But one interesting thing to know is, well, there is a spread across models. The different colors here are the different model simulations. The spread across models is typically larger than the spread across setups. That is to say our LES models with 200 meter resolution have a similar spread than their coarser resolution counterparts. And so that's again sort of interesting that, going at least from one kilometer to 200 meter resolution doesn't really decrease or increase the spread of responses across the models very much. However, when we go to our large domain simulations which do have aggregation, those have a substantially larger spread in the values of this climate feedback parameter. And some of the models even have positive values which would indicate an unstable climate. This spread increases even more if we consider our models with parameterized convection, the GCMs here on the left, with really quite an extreme range from climate sensitivities less than one to being infinite basically. For reference, the kind of current estimates of climate sensitivity for, seem it kind of six models are somewhere in the middle here around like three or four, five kelvin, something in the middle here. One other thing to note is that the GCMs have a lower average climate sensitivity as indicated here by this black circle than the models with explicit convection too. And that is a pretty noticeable difference which we'll explain in a moment. So what we've shown kind of collectively is that these simulations have a wide spread in self-aggregation. The self-aggregation impacts the climate but there's a wide spread in how aggregation changes with warming and there's a very wide spread in climate sensitivity. And so do all of these things explain each other? Does the intermodel spread in self-aggregation explain the intermodel spread in climate sensitivity? Well, we can assess that. And it turns out that in about 70% of the RCMIT models, though heavily weighted towards GCMs, the aggregated simulations have a more negative climate feedback and less lower climate sensitivity than their unaggregated counterparts. And so here, if we compare basically the solid circles to the open circles, the solid ones are the large domain simulations that are aggregated and then the open circles are the unaggregated small domain. And this is just showing the estimates using all of our different combinations of temperature ranges. And so we do see that the large ones have a lower climate sensitivity, more negative feedback parameter than the unaggregated small domain simulations. But we really mostly see that in the red, which is the GCMs, the blue, which is the explicit models, they don't really differ that much. And so that's sort of a little suspect maybe. We also find that these differences between the large and the small domain simulation are not explained by changes in the cloud rate of effect. And so we can calculate the cloud rate of effect as the difference between the net top of atmosphere, all sky and clear sky fluxes, and then look at how that changes with temperature as a sort of very crude approximation of the cloud feedback. And here, the cloud feedbacks really are not very different between the solid large domain simulations and the filled, sorry, open circle small domain simulations with a few exceptions. One other thing to note though, which is somewhat of an aside, but I think very interesting is that here the whiskers are showing the intermodel spread. And the intermodel spread in our estimate of cloud feedbacks is smaller in these blue models with explicit convection than it is in the red models, which are the ones with parameterized convection. And so that is, I think a little bit hopeful in that it says that even though many things about the explicit models do still disagree strongly with each other, at least with regard to this estimate in cloud feedbacks, maybe explicitly resolving convection will help us reduce the uncertainty in that. So this again, it suggests that aggregation is having effect on the climate sensitivity, but in terms of explaining the intermodel spread, there's actually no consistent correlation across models between the average amount of aggregation and climate sensitivity, at least as represented here by the average value of IORG. So here we have the climate feedback parameter on the y-axis and IORG on the x-axis. The blue models are all the cloud resolving models and then the red are the GCMs. And so we see a correlation within the GCMs, but not in the CRMs and not across all of them. And then furthermore, if we use a slightly different assumption with how we calculate IORG in the GCMs, we end up with these golden figures here, which gives us an actually opposite relationship. And so this all really provides no strong evidence for the fact that there's a relationship between the average amount of aggregation and the climate sensitivity. However, there is a relationship between changes in aggregation and climate sensitivity. And so now we again have the CRMs in blue and the GCMs in red. And what's plotted is the climate feedback parameter and the change in this IORG metric with warming. And we find that models that have more aggregation with warming have lower climate sensitivity. And this is actually why the GCMs have on average lower climate sensitivity as indicated by the red cross here than the CRMs do. And that's because the GCMs have on average an increase in aggregation with warming, whereas the cloud resolving models really don't. They're really this kind of cloud of points right around zero. And so that sort of explains this systematic difference in climate sensitivity between the two classes of models. If we probe this a bit further, we find that this relationship between more aggregation and lower climate sensitivity is mostly due to enhanced clear sky outgoing along great radiation due to mid tropospheric drying in the simulations that are more strongly aggregated. So if we repeat the calculation using just the clear sky version of the feedback parameter, we get a very similar relationship. And then if we repeat it again, using just the long wave component, we again get a simulation. So it's really the long wave clear sky component that is dominating here. And then if we kind of relate that to changes in humidity, we find that the models that have more of an increase in aggregation, these ones down here have lower climate sensitivities because they have a decrease in the mid tropospheric humidity with warming because they're becoming more strongly aggregated. And we know that affects the humidity distribution. Now aggregation is not the only thing going on here. We do also have contributions from changes in shallow clouds that contribute to the spreading climate sensitivity. So here on the X axis, we show the change in shallow cloud fraction. And this is really mostly a factor in the GCM simulations in the red lines where GCMs that have bigger increases in shallow clouds also that also contributes to lower climate sensitivity. There was ones that have a decrease in shallow cloud fraction would have a higher climate sensitivity. It's less prominent in the CRMs but does contribute a lot in the GCMs. So if we put these two things together, it turns out that the changes in aggregation and the changes in shallow clouds are really pretty uncorrelated from each other. And so we can combine them in a multiple linear regression. And if we do so, we're collectively able to explain about 70 to 80% of the intermodal spread in climate sensitivity by this combination of changes in aggregation with warming and changes in shallow cloud. And this is really exciting. And this I think really shows the purpose for doing these kind of idealized intercom comparisons because the whole hope with an idealized setup is that you're going to be able to actually explain a bit what's going on. And so we're really excited that we were able to explain the spreading climate sensitivity in these simulations. So with that, I'll wrap up and summarize. We've covered a lot of different things here in this really awesome data set of RCE simulations. In terms of the RCE state, we found that temperature, humidity and cloudiness very substantially across models, which maybe was surprising to some, maybe not surprising to others. It may be something that's a little bit disappointing that even in this simple setup, things disagree a lot. But I think it actually really what it does is it reflects the fact that radiative convective equilibrium is relatively unconstrained, meaning we have to have energy balance between the rate of cooling and convective heating. But in terms of how the models get their convection set up to make that balance work, they have a lot of freedom in how they can do that. And I think that this sort of spread in the representation of the RCE state that we see here is large, but it really reflects the true spread in terms of how convection can be represented across models. Whereas in other more realistic configurations with large scale circulations, those kind of constrain more what the models can do and might have been hiding what the kind of, the true differences across models are, whereas that's really revealed here, which is maybe scary, but it's very important information to know. Models do however, agree better in terms of how things change with warming with a rising and slightly warming of anvil clouds and in 70% of the models decrease in extent. With regards to self-aggregation, nearly all models aggregate to some extent in the large domain, but there's a wide variability in the structure and degree of aggregation and no consensus on how they change with warming. When it comes to climate sensitivity, there's a large spread and a big proportion of that spread in climate sensitivity is explained by the spread in the change in aggregation with warming. And again, the GCMs have lower climate sensitivity than the cloud resiling models because the GCMs have more self-aggregation with warming as well as a greater increase in shallow clouds. And so with that, I'll end and I'm happy to take any questions. I also would just like to kind of, I don't think I added a slide for this yet, no. Two announcements, one is that the RCEMEP data is publicly available for anyone to use to investigate all sorts of interesting questions about tropical clouds and convection and climate. And if you just Google it, you should be able to find it, but you can also email me and I can send you the link to the data which is generously hosted by the DKRZ. And then the other announcement is that we have a special collection across AGU journals on studies using radiative convective equilibrium. And so really would encourage you, if you're working in RCE to consider submitting a paper to our special collection, it does not have to use the RCEMEP, it can just be any sort of RCE study. So thank you again for your attention and thanks again to Adrienne and Simona for inviting me. Thanks. Very much, Alisson, that was great. So just a reminder to everybody that wants to ask a question, just to send it to us in the chat. So the first person who wants to flag the question is Sergio, so I will ask Sergio to unmute. Hello, thank you for your presentation, Professor Alisson. I was wondering about the self-abrogation during the warming environment. What do you mean? What is the mean of warming environment? Is this a specific semi-experiment considering some feedbacks? I'm sorry, but that was not clear for me. Thank you. Yeah, so we did three simulations. We did three sets of simulations at different values of sea surface temperature. So the sea surface temperature is fixed in uniform in these simulations, but we did three separate sets. So we have one at 295 Kelvin, one at 300 Kelvin, and one at 305 Kelvin. And so we can look across those different values of SST and see how the aggregation is different. And so it's an SST-based warming that's affecting the aggregation here. Okay, thank you very much. Next question is from Kerry, Kerry Manuel. Hi, Alisson, wonderful talk. I see the old MIT proverb that listening to an MIT talk is like drinking from a fire hose, but it was a very productive one. It turns out Adrienne and I had the same question. And Adrienne asked me to ask it to you, which is, and this may be a little subjective, but how much of the differences in among all these models and their degrees of aggregation and particularly in their climate sensitivity, you think may be traceable to different, there are different formulations of cloud micro-physical processes? That's a good question. I mean, my speculative answer is probably a lot of it, but that's sort of more of a gut instinct than any proof that I have. It's really difficult to try to entangle the differences right now because the models differ a lot from each other in many different ways. They're all configured to have the same SST, the same insulation, the same resolution, the same like carbon dioxide concentration, that sort of stuff, but they have different micro-physics schemes. They have different radiation schemes. They have different turbulent schemes. They have different boundary layer schemes. They have different dynamical cores. And we can kind of like slice and dice the ensemble different ways. Like there are some models that happen to be the same, except have a different micro-physics package. And so we can show that that does have an impact. Then we have ones that are the same, but have a different deep convective parameterization. And we can show that that has an impact. We have ones that have the same parameterization but different dynamical cores. And that has an impact. And so everything has an impact. And so it's sort of hard without doing more experiments where we like maybe constrain things a bit further to be more similar. And then like specifically change specific parameters to really see what the biggest factor is. But my suspicion is that the micro-physics is a big contributor. Especially like for example, in that one UK Met Office version that actually aggregated in a small domain, which was the only one. The thing that was different between that and the other versions of the UK model was the micro-physics package. And so that's a pretty big difference to have it aggregate versus not on like a hundred kilometer domain. So yeah, I mean, I think we're hoping to do maybe a second phase of RCEMAP where we maybe simplify the physics packages more and then kind of have fewer degrees of freedom between the models, which maybe will help answer that question. Yeah, in fact, just to follow that up, I mean, one of the things that stood out when you talked about the anvil cloud fraction reducing, I noticed that when you looked at the stability, all of the models except for one show the increase in stability with aggregation. And yet for the anvil cloud fraction, that wasn't the case. The majority showed the reduction, but there was probably on the order of a quarter of third that actually showed an increase. And so I was wondering if, you know, those particular models used a particular type of ice micro-physics scheme. The large eddy model you were referring to, that was the one actually years ago I was using for my PhD. And that in particular had a strong reduction in cloud fraction due to the micro-physics. There was a kind of a hardwired temperature formalization in one of the ice micro-physics processes that caused a big sensitivity to the tropospheric temperatures in which that process was taking place. Yeah. Yeah, we haven't, I mean, we've looked a little bit at kind of the ones that have, that don't have that decrease in anvil cloud fraction. And there's not anything that sort of stands out as being obviously different about them. But I can really quickly show like an extra slide that I had, which was working on trying to kind of assess whether the radiatively driven divergence stability iris hypothesis does explain the changes of anvil cloud fraction with warming in these simulations. And for the large domain simulations, which are on the right, it seems like it does. And so here we've got anvil cloud fraction on the y-axis, the anvil cloud peak, and then the radiatively driven divergence peak on the x-axis. And it's all normalized about 300 Kelvin. So the little small dots are the simulations at 295, and then the bigger ones are 305, and then the one is 300. So basically, if you're going from this top right quadrant to the bottom left, then that says both anvil cloud fraction and radiatively driven divergence are together decreasing with warming. And that's true across most of the large domain simulations, not all of them, but most of them. This one I would ignore that has sort of a spurious behavior of its clouds. But in the small domain simulations, we really don't see that so much. Like they have a consistent decrease in the divergence, but that doesn't necessarily lead to a consistent decrease in anvil cloud fraction. So I think it sort of suggests that these stability-based arguments are playing a role, but they're not the only thing going on and there might be other contributions from micro-physical effects, like you're mentioning that can add additional variability to this relationship. Okay, thank you. Well, just one final question. I'm just waiting to see if anything else comes in. I was, you mentioned at the end that the next phase might focus on, should we say, simplifying the physics. So there are less degrees of freedom across the ensemble. Where else might you be taking RCU map in terms of, should we say, investigations? Are there aspects where you think, oh, I really wish we had included that in that phase one. We really need to look at that in terms of maybe experimental setups, or so on, and are there other areas that you don't want to? Well, one of the areas that I personally have been interested in would be, it kind of continued to try to better constrain this effect on the changes in aggregation with warming and its effect on climate sensitivity. One thing I'd like to do is to maybe configure some simulations where either aggregation is suppressed, for example, by homogenizing the radiation or like doing like a mock walker type setup where you can kind of control if it's aggregated and how strongly aggregated it is and allow us to have like more ability to control that and then see, given, once you can control for that, then what is the sort of spreading climate sensitivity? I think that another, an additional avenue exploration could also be turning rotation on in these simulations and having rotating RCE, which is gonna form lots of tropical cyclones, which opens up a whole other interesting avenues of exploration. It's a little trickier to figure out how to do that consistently for clover-zombie models and GCMs because ideally you would want, like clover-zombie models, you could do an F plane that you just kind of get one or two TCs or something. I don't know, it's harder to figure out how to do it, but it would be cool. So yeah, I have ideas and we're still also trying to kind of continue analyzing what we have from the first phase, but I hope that this is something that continues in time and we kind of keep this community together and in train more people as well. Okay, so that's fantastic. Is there anybody else? I didn't see any other questions on the list. So next week, thank you very much again, Alison, for the presentation. Next week, the hosting passes back over to Trento. Oh, Caroline is just coming with a question. Let me, there you go, that's on mute. Sorry, everyone's muted, you see, just do it. Here we go, Caroline, you should have it. Thank you, thank you, Alison. This was wonderful. Thank you very much. I did have a quick technical question perhaps on the profiles, the relative humidity profiles that you showed, I think you had the ensemble, it was hard to read the caption, but you had the ensemble average and then the ensemble minimum and maximum. And it was close to a hundred percent on a bunch of levels. Is it the maximum mean profile of the mean profiles of all models? There's been that there's a model with a hundred percent relative humidity. I'm guessing not, but if you could just, maybe elaborate on this. Yeah, so these are domain average profiles, averaged over the last something 25, 30 days of simulation. So these are mean profiles. And here we compute relative humidity with respect to liquid. If it's below freezing and with respect to ice, if it's above freezing. And so it's actually common to have even super saturation with respect to ice in the upper troposphere. And that's actually supported by observation. So yes, there are a relative humidity of 100% in the upper troposphere and even above, but that's over ice, which is not unusual. Okay, so does that mean that the model has a cloud spanning the whole domain or does it mean that there's super saturation allowed and it's actually not forming a cloud although the rest humidity is above a hundred? Yeah, it depends on the model. Some of them allow super saturation. Some of them do not. Some of them do have domain filling very, very thin amounts of cloud. Okay, yeah, that was surprising. Okay, okay, that's clear. Thank you. The ones that do have that domain filling high cloud, it's very, very thin cloud. And so if you just increase your threshold for defining a cloud a little bit, then that kind of goes away. And it's sort of the optical thicknesses are not very large. So even though it might be diagnosed a cloud, it doesn't necessarily have a strong radiative impact. And we're actually working on digging into the cloud profiles more and changing our definitions of cloud fraction and reassessing it and computing optical depth and things like that to try to better understand what's going on. Yeah, it is a little surprising, I agree. Yeah. Okay, thank you very much. Thanks, Caroline. Thanks very much, Caroline.