 Good morning, my first speaker is Yosvan Handerbeck from Ciena, Ciena, or Ciener, if you want to say it in original language. So, experiences from tuning high-resolution climate modeling with this year's people. Yes, so what I want to show you is some of the things we have done in the past years regarding mainly tuning a climate model, namely ECEOT. ECEOT is based on IFS, so I think it will be of interest for this community. And, well, I will come to the main points. Of course, the point is that ECEOT, of course, is a climate model. So, we are interested in longer time scales and what we would be interested in for weather forecasting. So, first a few words on ECEOT, which has already been shown and mentioned by many speakers in these days. ECEOT is actually, it's mainly a global climate model. It's quite recent one in the sense that development has started in the, well, basically 10 years ago, in 2006 or 2007. The first idea came out. It was the first version, ECEOT 2. But already then, the current version we are working on, which is version 3, has started development. Actually, I just checked this morning and asked Uber who actually did that. The first commits were from 2009. To give a reference, ECEOT 3.0, which is basically, what we are working on now is basically continuation of that version was released in 2012. It's based on IFS. It's based on IFS Cycle 36. Our idea is actually to move on to open IFS in the next future. So, ECEOT will prepare this year a new science and implementation plan, which will also state clearly this plan. The idea is actually that ECEOT version 4 would be based on open IFS. This will have many advantages among others to have a closer interaction with open IFS community, to have development cycles which are closer, faster updates, etc. And I think also there's a lot to be gained from parallel development and basically the same issues not having to be solved over and over again. Of course, as a GCM, ECEOT needs an ocean. It uses Nemo. It makes a particular choice for the IC model, which is Limb 3. And, well, for land, of course, coming from IFS, it's using HTSL. And this is the GCM. But ECEOT is also, or has a mission, and it is also being developed as a nurse system model. So, other components at the same time are ready and have been coupled to the model. And so we have a dynamic vegetation model. And there's an interactive atmospheric aerosol and chemistry module, which is based on TM5. Dynamic vegetation instead is LPJ gas. And there's a biochemistry model for the ocean, pieces. So basically all these components have been under heavy development and being coupled to ECEOT in the past years. The goal is actually to develop an overall system model which is ready to do CMIP6 integrations. And I think, well, time schedule is always a bit too optimistic, but I think things are well underway for this goal. And finally, ECEOT is also a consortium because ECEOT is based, sees a contribution of a large community of European institutions which are contributing. This slide may not be completely updated. So if you're involved in ECEOT and you don't find yourself here, don't be offended, this is an older one. So basically we talk about something like 22 European institutions actively contributing to the development of ECEOT. The original plan of ECEOT, well, one of the motivations is also this idea of seamless prediction that you have a model which you start from a weather model. You can apply it also longer time scales and then you end up with climate. And the idea is that of course if you make a good model for one, the physics being the same, why shouldn't you be able to have a good model also at longer time scales. So the tuning of the model, well, I will go more in detail what I mean by tuning of course, but the goals of the tuning of the model are mainly to improve some certain things in the model and to make it ready to be competitive with other models in CMIP6. Obviously it's all aimed at having a more realistic model. And the first thing we looked at and actually we worked a lot and actually a great part of my presentation will be on this was actually on looking at energy fluxes, energy balances, mass fluxes and mass balances. And the reason for that is actually that it's quite different to tune a weather model from a climate model in the sense that suddenly you need to take everything what you certainly need to have all the parts which you would look at in a weather model. And actually being based on IFS, you would say that part is already well tuned. I mean, ECMWF did a lot of effort in doing this to the best to state of art. So actually why would we need tuning? Well, the problem is that if we look at longer time scales, tuning a climate model poses new problems and new issues. There are some things like mass conservation and energy conservation which may be negligible over shorter time scales and may become very important over long time scales. So the small imbalance in P-C, precipitation minus evaporation, may be completely irrelevant at short time scales or even at seasonal time scales. But over 100 years it may change your sea level heights by meters and meters. And so obviously it will be very relevant and so you don't want that in your model. And the other thing as we will soon see also is that actually energy and mass conservation are sort of related because obviously mass conservation also implies or missing mass conservation also implies missing energy conservation. If you don't conserve water vapor in the atmosphere, well sooner or later this water vapor will condense releasing Latin tea and again so that missing mass conservation becomes a missing energy conservation. And these are issues which may not be relevant at shorter time scales. Of course it's not all about these radiative fluxes and so of course you're also interested that the long term averages or statistics of certain fields are as realistic as possible. And of course you have the problem of having good references to compare with. For ECOS we, the trick is actually to compare with era intering which is sort of cheating because we compare the model sort of with itself. But it's helpful because at least it gives a clear reference what to compare with. ECOS should at least be close to era inter if possible. Of course it's not only about fields, it's also about model variability and so that's another thing you look at. And all these needs and all these comparisons of course need some, if you need to do this process of many variables, different groups and you need to compare what you're talking about, you need to have some common measures, some common way to talk with each other. So indices and by index I mean any number which can be sort of shared and which gives you a measure of skill become extremely important in this sense. So they are, well, the simplest one is this rifling key indices which were actually developed now quite some time ago which is basically adjusted with mean square differences compared to some reference fields. And then of course a more refined thing is actually to look at the regional features, regional dynamics, etc. This whole exercise actually, as I mentioned, is necessary also for coordinated experiments which is involved in to have a model which is well suited for this experiment. And the reason is actually for each experiment, say you do CMIP6 or you do some other experiment like we're involved in a coordinate experiment called high res MIP, which is actually part of CMIP but where you compare the high resolution version of the model. Every time the protocol changes a bit, the forcing fields change a bit and so of course you need to check again if your model is sort of well tuned. Because some of the tuning you do may be specific for having made so much assumptions about the forcing fields, say ozone or aerosols, etc. So I mentioned the radiative imbalance, well there are many, there are very good estimates of these, of course recent ones also from satellite, etc. There is some recent literature about this. All these numbers have uncertainties which are still unfortunately quite large. Basically it's always something like plus or minus zero five in all of these numbers. But still we can use the central values as a reference and we would like to see the model to be quite, to be similar to what is observed in some of these numbers. And the most important numbers of course for, while usually you would judge coupled model in general and climate model from the top of the atmosphere fluxes, basically you just look at the balance between what is going in in terms of energy, what is coming out. It's not zero because our planet is warming. What is actually more relevant for a couple models of course is surface fluxes because that is where the adjustment between the atmosphere and the ocean and the land surfaces are curing. I would say mainly ocean because the land surface has much reduced heat capacity so what really counts is the ocean. And so of course it is this fluxes at the surface which will be a particular important. So as, I wasn't here the morning of the first day so but I was told that actually Franco also mentioned this for IFS, that the first thing we discovered when we started looking at this to our dismay was that actually if we took the top of the atmosphere fluxes, the net flux at the top and compared it to the net flux at the bottom, well this should be the same because the atmosphere has no significant heat capacity. So there's no way, average of a long time. So there's no way it could possibly store energy somewhere or release energy significantly. So if you, on average, those are the difference between what you get at the top and what you get at the bottom, you really should get zero. And any different numbers in indication that the models either producing or consuming energy. And we did that and well the number was quite significant, what's significant, very high. And we got this huge number of minus 2.5 baht per square meter. All these are at our reference resolution, standard resolution. At different resolutions unfortunately you get even different numbers. So there's also the sensitivity of this number to the model resolution. This had been tested, et cetera, and basically it's quite robust result. And then of course, depending on the forcing fields, you can shift where this energy goes. So like in IFS, the convention of course is that incoming downward radiation has positive signs. So minus means actually outgoing. And in this case, this is top minus surface. So a minus really means outgoing from the atmospheric slab. And well, so the atmosphere loses energy. That's okay if it were cooling. But if you look at, by looking at this experiment, the atmosphere was not cooling significantly. So nothing which could justify this flux of energy. Of course, depending on the forcing, this could be distributed differently. You just change say you do an atmosphere-only experiment. And of course then, depending on your sea surface temperatures, the ocean may be too cold or too warm. And so of course these fluxes may either go inside the ocean or be released at the top of the atmosphere. But some is always the same. So nothing to be done about that. And you can fiddle as much as you want with your parameters of the model. And this difference usually is quite robust. As long as you stay at the same resolution, as I will also mention. So basically, this is a, nominally, this is an internal heat source in the model. And it's a large heat source because it has to be compared to, mainly basically to what we're interested in in a climate simulation, which is anthropogenic forcing, which is something around 06. So 2.5 is embarrassing. And you could live with that, but the problem is that you could just, if we were sure it's always the same, it's always the same, and we understood what causes it, we can of course live with that. And we can also make future projections, et cetera. But we do not know, if we don't know the processes, which are at the base of this, in this case energy production, how can I trust a future projection if I have no idea if maybe these processes may change, and may change this in a different setup. And in fact, there is a state dependency of these imbalance, which is unfortunate as we show, so that's important. So the first thing is, which we quickly, well, took us some time, I must say, realized was that we were making a mistake in our balance. So because nominally, the model saves all these fluxes. So you have your sensibly heat, your latent heat from evaporation. You have your incoming shortwave radiation at the surface. And you have your thermal heat, longwave heat, net at the surface. These all get stored nicely in grip files. You see them in the output also of OpenFS. And it would be natural to sum them up and say, okay, that's a net surface balance. And that should be, well, 06 if you are in a transient present-day simulation. The problem is that the latent heat actually is computed only based on evaporation and sublimation. It does not include one important player, which is not obvious immediately, because it's actually a mass flux. It doesn't include snow. Snowfall actually carries with it a latent heat content. And it's quite easy to understand, the atmosphere actually, to create the snow, takes droplets, freezes them, and to simplify, freezes them and basically extracts heat by doing this. So the atmosphere is gaining heat. The droplets, then, the snow falls into, say, the ocean. And the ocean will use heat to melt them again and bring them to SST. So basically, the atmosphere has gained heat. The ocean has lost heat. This is a heat flux to the atmosphere. It's a negative heat flux because the sign is upward. And, well, based on the average, there's no fall in the model. Well, it's something around 0.23 meters a day. You multiply that by 334 kilojoules per kilogram. And, well, you get that that's actually equivalent to minus 0.88 Watt per square meter. So that's huge. In a sense, if the numbers are below 0.1, 0.2, I would start worrying less. But as long as we're talking about numbers around 0, comparable to the anthropogenic forcing, I would be quite worried. So this clearly has to be added. And the same, actually, also is true for land because also the land surface uses heat to melt the snow. And this, of course, will cool the land surface, quite natural. And part of the snow, instead, is sort of excess snow, which at least in the third becomes a flux called calving. So it's basically a representation of glacier calving. So it gets transferred to the ocean. And then it melts again. So it's a flux which, in the end, goes to the ocean. So, basically, it's the same story also for the land surface. So this we had to remove. OK. And so basically, based on this, we have to look at the revised version of our top of the atmosphere in the surface. But, well, it's still not a full 2.5. So we haven't solved the problem at all. We only have reduced it significantly. Then the other thing which we found is that comes actually from solving a completely different problem. It's that the model is not mass conservative. P minus E is actually 0.3 millimeters a day. Was actually 0.03 millimeters a day. This may seem not very much. It's actually quite a lot if you integrate it over 100 years or so. So it will significantly change your sea level heights because this water ultimately will end up in the oceans. And so it also poses a problem in that sense. And in fact, it has to be corrected. And the initial actually admit we correct this by then it's unacceptable that the water would increase. So in the end, any residual mass imbalance like this, we actually correct. In the end, by slightly adjusting river runoff. But apart from that, this actually could be identified to be caused obviously by the advection scheme in IFS. So of course, it's an error which comes due to the numerics of the advection scheme. It's a well known problem. Actually a lot of fixes have been developed by using the graph basically in later versions of cycles of IFS. And actually it's significant also from an energetic point of view because 0.03 millimeters per day, well you do your mass and you end up with a number like this. So almost 1 W per square meter of energy. Because again, it's the same thing. You advect water vapor, but you create new water vapor. And this, once it condenses, releases heat in the atmosphere. So voila, you have energy production in your atmosphere. But okay, so we implemented the simplest possible fixer we could which was simply a proportional fixer. That was easy actually to write ourselves. And later we actually backported one from Cycles 38. We could not backport more sophisticated ones because the core structure of IFS has changed too much. So it would be significantly too much work to do that. It's actually a nice product. And well, doing this, P minus E actually was reduced, actually became negative, but at least half of what it was before in amplitude. And so this residual minus 0.016 is, you probably listen to something else, some other numerical things. But the main point is, so basically we reduce it by, instead of reducing it by 003, we reduce it by 0045 or something like that. And in fact, we had a significant improvement in the top of the atmosphere minus surface imbalance, actually 1.4, which is corresponded with this change in the mass balance. And so basically at the end of the story, with all these things put together, now our residual top of the atmosphere minus surface imbalance is only minus 0.27 W per square meter, which is, but we decided we can live with that. Even if it's obviously still, it's still a minus sign, and so it's still an energy production. Okay, something similar while doing this, at a certain point we also work with stochastic physics. And so investigating stochastic physics, I said also the stochastic physics scheme was not energy and mass conservative. And so basically this was leading to a very strong negative precipitation and vibration imbalance, actually significantly strong, because this now starts being a very huge number, comparable to the average snowfall basically, which might do us the 0.23. And it was also corresponded to very large top of 30 atmospheres minus surface dead flux imbalance in the model. So again, in this case, we implemented a simple proportional, since we had that tool, proportional fixer. So basically making sure that before and after the application of the SPPT scheme, these quantities would be conserved. Of course it has a cost in terms of communication in the model. But it solves the problem, because basically after that, if you use SPPT or you don't use it, the P minus E balance becomes exactly the same. So basically that problem is solved. And then working actually with Antje and others at the ICNWF, the same thing has also been implemented even in IFS itself, as documented in this memo. So having solved that, we did a lot of series of experiments to actually start fixing in general the fluxes to actually tune the model and get fluxes to our liking. And we need some tools to do that. And these are specific parameters which we identified for IFS, which we could actually modify. It's sort of classic or standard for all communities, all climate model developing communities to have to modify basically convection and basically parameters related to the water cycle mainly, or related to clouds and to the microphysics. And so this is a list of the parameters which we experimented with. So all these parameters will have... Yeah, sorry, this one here, which is not really a parameter, it was a code modification which we also experimented. Problem was that when we first did this, we had very unrealistic fluxes at the surface even when we forced the model with present-day sea surface temperatures. So basically the model was decided that it really did not like the temperature of the ocean which we were giving it, no matter what we did. And it was even by changing all these parameters within ranges which are sort of reasonable, all these parameters are not fixed. They are not physical. They have no real physical value. They have recommended ranges also, which you can derive basically by talking with people who implemented these parameterizations. And this actually, I believe, a thesis work in Oxford which has been done for development of the stochastic schemes which also discussed this a bit. And then... But basically we were not able to shift enough basically our fluxes, mainly our surface fluxes. So one thing actually then by discussion which came out by discussion which is in the ref was also that there is one change which was specific for cycle 36 but then had been changed again in cycle 38. Actually it was a reversing because it was there before and then came back afterwards which is a specific condensation limiter, a specific way how to implement this. And so we went back also in ECR to the old scheme and this is actually, we could use it as a tool to shift surface fluxes by at least 1.5 watt per square meter. So we had an additional tool we could use to use it. Here I talk only about global values but of course after doing all this in the end you have to go and look if your model climatology also regionally in terms of fields, etc. is still reasonable and of course this has been done. So anyway we have all these parameters and at that point what can we do? Well you need to know the sensitivity of the model to these parameters. So we did a series of experiments relatively short, these are all six year experiments bearing these parameters. So starting from the reference value and then increasing and decreasing them. So this graph is a bit older but actually each of these actually had five points in the end. And so you change each, so you see these are all the parameters, you change each parameter and see what happens if you increase them slightly or decrease them slightly within the range of what is reasonable. And then you can make similar plots for all possible radiative fluxes you're interested in. So this is a TOA flux but you can do that for cloud forcing, long wave cloud forcing, you can do that for short wave at surface, etc. So you get a bunch of plots like this. And the lines actually are, believe in this case the lines here are where we would like these fluxes to be. But what we really were interested here were the slopes. And what is interesting is that most of these are reasonably linear so for all purposes we decided that these are actually linear so by fitting a line we can actually derive these sensitivities. Don't look at the numbers, I'm just showing this to tell you to give you the idea. So for each parameter you would get the sensitivity of a specific flux in terms of the parameter. So what happens if you change the parameter how much does it change if you change the parameter by one unit. And then you can put all these together and so we made basically a software tuning simulator to do that. Actually we were lazy so it's a big Excel spreadsheet so don't worry but it works nicely. And with that tool actually what you can do is to plan your experiments. So you can say okay I don't like these fluxes how they look like, you can even implement a score. You could say, and this we did actually, you could say I value a lot the net surface flux so I give it weight 4, I value a little bit more, little bit say cloud forcing, long wave cloud forcing so I give it a certain weight and make a score like that and then I play changing these parameters and try to reduce to find a minimum of my score. So to get a model configuration which is the closest possible to what I would like. And the nice thing is that this way you can plan your experiments, do that, then make the experiment actually for say a reasonable time, say 20 years, get new averages and then see how we have done and actually this converges quite efficiently so in a couple of iterations you get actually where you want to go so that's a very useful tool. So we were able to do that, we combined these and actually we decided we actually focused on a few fluxes which were particularly important for us and we tested different combinations. This is actually an exercise we did a couple of years ago so this is not what we would do now, what we are doing now but it's illustrative what we did so these are all different combinations of parameters and this is the end result in terms say of TOA fluxes of what is it, yeah, of the different fluxes which we're interested in. This for example is this famous TOA minus surface difference which as you can see is pretty stable no matter what we do never changes, there's nothing to be done about that by changing these parameters but all the others are quite sensitive and so you can actually find combinations which really give you what you would like and the green bands are actually the desiderata where we would like the model to lie based on observations and it's interesting that at least this was an exercise we did one and a half years ago, one year ago and that was based on the previous version of this year with EME brand it was interesting that in the end the winner was actually a combination of parameters which was pretty similar to what a later cycle of IFS cycle 340R1 was using with some modifications but many of the parameters in the end we ended up changing in a similar way maybe a chance but it was pretty interesting so of course this has been repeated more recently with more recent versions of the model now using actual CMIP6 for things and so this is an exercise which is undergoing and well in the end you can get really what you want and still then you can check the model climatologies the dynamics in the model, etc and check that still everything is okay and make sure that the model is that you didn't ruin anything of course I do not have the slide here at the right place but basically in order to check to keep quickly under control that you do not do any damage by doing this you also use other indices so we have this Reichelin-Kim performance index I mentioned before which basically quickly gives you some numbers to tell you if you're getting too far from observed climatology and so basically as I said you shouldn't ruin your scores and still get the decent score even if you make these changes okay so this is of course not the end of the story because this was AMIP so if you now couple the model to the ocean different problems come out of course the first problem of course will be that you may have a problem that actually you should tune the model for AMIP and then maybe since the ocean will have a different idea on what the SSTs will be you could be tempted to also do a different tuning for the coupled model and that depends on your strategy of course it would be desirable to have the same model for everything but actually when you couple the model to the ocean new things come out and this was actually a recent one which we discovered and discussed which is actually the fact that Nemo takes actually additional initiatives and adds further heat which you wouldn't have expected it should add or remove further heat which you shouldn't expect to do the idea is that the atmosphere is coupled to the ocean it sends I can use actually this slide which says exactly what I need the idea is that the atmosphere is coupled to the ocean it sends from the point of view of heat it sends a solar flux which is distinguished from non-solar fluxes mainly because the solar flux is penetrative so it goes deeper and then it sends the surface fluxes which are the latent heat flux the sensibly heat and the longwave fluxes which all act on the first layer so basically IFS is sending to the ocean everything it needs and then it also sends mass fluxes and Nemo of course knows it would be surprising if it didn't that it has to spend heat to made snow so it will add that latent heat for you so that's all nice problem is that then Nemo but most other ocean models I'm told also has to take care of other things like you add a mass of water to a bucket of water you're diluting that water so it becomes important to know what is the temperature of the water I put in it makes a difference to put well obviously it's probably not a good term it's actually a wrong term but there is a sensibly heat content in addition to the latent heat content say of snow there is a sensibly heat content in the temperature of rain say falling into water or in the temperature of river runoff or the fact that calving is at zero degrees all these actually play a role and you can estimate them there is also some literature recent which has tried to estimate this and you always end up to a number which is similar to what we also get in Nemo which is basically we evaluated this and basically this is worth well it's not insignificant it's worth minus 0.23 watts per square meter so again it's a number we could be worried about I must say though this is an average of the ocean so globally it's a bit less of course and so this has to be taken into account and why is it important because IFS does not take into account of this heat IFS did not spend energy to warm up the rain whereas the ocean uses this the warm rain and extracts heat you see there is a problem here because obviously there's a non-conservation of energy in all this and so the first attempt you could make is to say okay all mass fluxes entering the ocean are all at the same temperature which is good and nice but obviously it's not very good for say calving which is basically snow and ice which just melted which obviously will be at a much lower temperature so you'll make a mistake there and so in the end basically what we decided to do was to leave it as it is and to live with the fact that there's an energy sink in the coupling basically and there's nothing else you can do actually and the energy sink of course is less than this because well you have to convert it to a global average so that would be minus 0.16 Watt per square meter and there's another thing which happens in the ocean which we have on and which actually happens also of course in nature is geothermal heating personally I thought it was not very relevant but then if you look at the number it is it's well it's this number you may decide that's important or not but it adds up to all the other guys so in the end you also have this 0.06 Watt per square meter of geothermal heating in the ocean just that's physical you can decide if to represent it or not if you have that represented but of course you have to take care of that and remember that when you do your balances if you decide if the fluxes which you get to your ocean are realistic or not because then you would have to add this number to your balance and I just make a few final caveats points which came which one has to remember which are many of these things are actually state dependent so say all other fluxes too but say this top of the atmosphere minus surface in balance actually has a dependence on the state this was a series of AMIP experiments where we manipulated the temperature of the ocean and immediately this reacts and reacts quite a lot so that is that's why it's very important to avoid them because in a climate projection where the sea surface temperatures may change this may mean that this difference may also change so again it makes you worry sorry it's the same thing until I don't understand what is the cause of an effect I'm worried that it may do uncontrolled things in my simulation the radiative fluxes themselves are state dependent so this shows you the same sensitivity plots I showed before but done at present time and done in a future scenario say in the year 2100 main point it's the most extreme one so the main point is that it's much warmer planet then and you see well you could say that these are different and that the sensitivities are different and again we have a problem then and finally there's a technical thing how many minutes one minute two minutes the final thing is that there's dependence on the time I just show you the final thing which is this how do you actually proceed to do the tuning of the coupled model actually and what is very useful in this case is to use plots like this where you plot some measure of your radiative net top of the atmosphere or surface fluxes versus temperature either global average temperature and what is nice is that all models actually have a typical sensitivity and while they're wandering the space they do that so following an approximately constant linear line obviously if it's short enough ECOD does the same and the nice thing then is that you don't need to run your model for hundreds of years to guess where it will end up because then you can do an exercise like this you can take say ECOD estimate this from a long run you have done but maybe not the run with the final forcing the final setup etc you spent maybe months to do this and now you want to get as much information as you can from this well the slopes are very important because this way now I set up the model with the right forcing, the right tuning everything now is ready and I just do a short run a few years, ten years or so coupled which may already be expensive I look where I am and then I can sort of guess where I end up basically what the system will do is we'll try to set to zero the flux at the interface between the atmosphere and the ocean if you are in a permanent run say you're doing a pre-industrial run and so this can be done and this is actually an attempt to do this now with high resolution ECOD which is quite costly and the details are not exact but it gives you the idea that you can sort of guess from already 50 years of run what your equilibrium temperature will be and then you can decide if you like that or you don't the two sets of points are doing this either at top TOA or at surface and I think that's it I mentioned all the processes so we also worked on improving some dynamical processes in the model but I think we'll skip this