 Good morning. My name is Franco Moltenia. I also work at the European Center. And my role for a long time has been mainly in predictability studies and long-range forecasting. And since the main topic of this particular workshop is the application of the Open AFS to seasonal forecasting, I thought you might want to hear more about some recent projects that we are running and also some operational developments regarding long-range forecasting and the simulation of climate variability. So we'll present results from different projects and, of course, as usual, the TCM that we have, there are many colleagues involved in these exercises so I would like to thank and acknowledge the contribution of a number of colleagues of mine. So as Erlander has anticipated, I will say a few more words on the new seasonal focus system that is going to be operational in autumn 2017. This is a higher resolution system with respect to the one we are running at the moment, but not as high as the resolution we use for the medium range ensemble. So we wanted somehow to have an experimentation about what advantage we could get on those timescales if we could afford to run the same resolution in both the atmosphere and the ocean on the sub-seasonal and seasonal timescale. And so in collaboration with the Center for Ocean and Land Atmosphere Studies, which is hosted by George Mason University in Fairfax, Virginia, we made a joint application to the NCAR Accelerator Scientific Discovery Program for a set of experiments that would basically run our seasonal and sub-seasonal focus system at the resolution that we actually are using in the medium range. And this is the so-called Metis project. And I will say a few words. We are just basically finishing the integrations. And so this will just be a brief outline of the project. I don't have yet results. We still have to get the results from this project just to make you aware of what will be available in the near future. Then I will mention the runs that we are doing in the Primavera project, which is a EU funded project under the Horizon 2020 framework. And this is about simulation of climate variability with high resolution couple models. And again, high resolution has a different meaning when you talk about multi-decadal timescales. And so here the reference are basically the current resolution used in typically, I assume, five experiments, which is of the order of 100 kilometers for both the atmosphere and the ocean. And in Primavera, we are running a new generation of climate models at the resolution of the order of 25 kilometers, so about a quarter of a degree. And as I will explain this project, we have a sort of historical component, so simulation of the climate up to the current times, and then an extrapolation up to 2050 CEM that we have is only concerned with the historical part. But this will allow us basically to look at the performance of our model on basically long integrations that we've never done before. So let me start with the seasonal forecast system. This will be called System 5. As Erlan mentioned, we started during operational ensemble predictions in 97-98. We were convinced, you know, we managed to convince our council to allow us to do so, thanks to the very good simulations of the 97-98 El Nino that were done with the first version of the CMWF couple system. And so this will be the fifth realization of this system. It will use a very recent model cycle, the one which is operational at the moment. We will also use the new approach to the numerics that Erlan has described, so the use of an octahedral grid with a cubic type resolution, which means that we can actually compute exactly up to a third order terms on the grid. So although in terms of the spectral truncation the difference is not very large, so System 4 was running with a spectral truncation of 255, the new system will run with a spectral truncation of T319. However, the resolution of the grid used has improved substantially from 80 kilometers to about 35. The number of atmospheric levels is unchanged. We tested the version with 137 levels, but somehow for reason of availability of computer time, we had to limit ourselves to the 91 version. The ocean resolution has increased substantially, so we still use the Nemo model in a more recent version, but instead of using the one degree configuration, we are now using a quarter of a degree, and this allows us to represent the fronts and the sharp gradients in the SST in a much better way. Another big improvement is that in System 4, we didn't actually predict dynamically the evolution of the sea ice, we were just starting from a self-condition and assuming that over basically the time of one season, the sea ice will evolve in a way similar to what happened in the past five years. So we were picking up sea ice distributions from the previous five years. Now, with System 5, we will have a dynamical sea ice model, it's the Lim 2 model which has been extensively used by the Nemo community, and so we will have dynamical predictions of sea ice evolution. You're probably aware that when we run seasonal forecast system, we have to run a re-forecast in the past to calibrate our system. For System 4, the re-forecast period was 30 years from 1981 to 2010. For System 5, we are just extending it basically to the present. The re-forecast ensemble size will also increase from 15 members to 25 members. So we are actually starting running this operational re-forecast. We started with the re-forecast for the summer season, then we'll go into the autumn season, the first ones that will be needed to issue operational forecast in the autumn. So, as Ellen mentioned, one of the big improvements that we saw with this new model was the strong reduction of the biases in the tropics and in particular in the equatorial cold tone. In System 4, this resulted from a combination of two strong trade winds with the atmospheric model and also some deficiency in the ocean model. Now, there have been advances in both directions. I would say that both components have improved and the result now, as you can see from the panels on the right-hand side, is that the strong cold bias that we had, particularly in the Nino 3 region, has now been replaced either by a slight warm bias during the northern winter in the JF at the top or a much, much reduced cold bias along the equator in the summer season. Another area where we were expecting to see improvements are the regions of the western boundary currents are regions with sharp gradients that are not well resolved by the A1 degree model. You can see here that there is, in fact, a reduction in the bias in the region of the Gulf Stream. We're perhaps hoping that the reduction will be even stronger. We still have some problems to solve in this region, but technically you can see that the amplitude of the bias has been reduced. Some other improvements have come from improved mixing. In the ocean mix layer, we have developed a new parameterization of the mixing which also gets information from the wave model in an interactive way. This has been instrumental in reducing the warm bias we had in the North Pacific and the North Atlantic during the summer season. I will go briefly with this slide because you've just seen it. One addition is actually the stop diagram that Erlen has not shown and that is the impact that this bias reduction has had on the amplitude of El Nino. You have seen these two graphs and you have seen that we have a slide with significant improvement in the anomaly correlation but actually a much bigger improvement in the RMS error. The reason is that because of the cold bias, system 4 was overpredicting the amplitude of El Nino. You can think that because El Nino comes from the oscillation of the Pacific thermocline basically the maximum event you can get is where the thermocline is almost horizontal. So if in your mean state the thermocline is already too tilted, then basically you have the possibility of creating very large positive anomalies when the thermocline almost flattens. This is in fact what happened in system 4. We tended to overrepresent the amplitude of the big El Nino events. This was particularly evident in the range from 1-2 to 1-5. So you can see in this diagram here the curve for the amplitude of system 4. You see that it reached almost 40% more than the observation. System 5 is much closer to the unity, so it means that the amplitude is much closer to reality. So basically the reduction in the RMS error comes from the fact that although the sequence of the time revolution was also reasonably well predicted by system 4, sometimes this happened with too large amplitude and this particular error has been now substantially reduced. We were expecting improvements because of the use of a dynamical CIS model and certainly this has been verified. So if you look for example at the anomaly correlation of CIS cover during the northern winter, you see that even with this sort of basic system we had some predictability in the variation for example in Decara and Barency which are quite important regions for the connection to the North Atlantic oscillation, but clearly we get a much better simulation of these variations with the dynamical CIS model. This is one error that I think will be explored by some of the people here with the Open IFS. Now, not everything has improved. It's a common experience when you change a very complex model. Some of the biases remain similar in character. For example, we still tend to have an overestimation in the amount of precipitation. We model from the tropics so you can see here the biases in precipitation with respect to GPCP, system 4 again on the left, system 5 on the right and you can see that the character of the bias has not changed dramatically. Perhaps in another area where I think we have to understand why things have not changed very much are the teleconnections associated with convection over the Indian Ocean and especially from MJO activities. It's well known that enhanced convection in the Indian Ocean which is usually associated with phase 2 and phase 3 of the MJO tends to produce a positive NAO signal after about 10, 15 days. If we actually look at the observation and if we composite the anomaly of geopotential height at 500 hectopascals 10 days after an MJO in phase 3 we see this very strong NAO signal here. If you actually look at the two top maps that come from similar analysis in system 4 and 5 you see a positive NAO signal but I would say roughly half. The amplitude is about half of the observation and this is an aspect which has not been substantially improved in the new system although I have to say the teleconnections from the core NINU regions have actually improved due to the better representation of NINU. Another important phenomenon for Europe is blocking particularly during the winter period when it can be associated with cold spells quite intense cold spells. Here is a plot of the frequency of simulated blocking using the Pivaldi-Molteni index which is now 30 years old but still in use. And you can see that there are some areas of improvement like in the Northwest Atlantic these are the so-called Greenland blocks and the Pacific. Not so much in the actually on the eastern side of the Atlantic. In fact, the area of improvement are those when the blockings are more associated with large scale variations in planetary waves. The European blocking is a much more non-linear process due to the interaction with the transient deaddies. So we still have some work to do. I would say that usually the winter seems to be for our model the most critical season we had actually much better results from autumn and spring. Okay, just a few mention because we have not got results yet from this Metis project but again to restate this is a collaborative project with COLA funded by the anchor Accelerated Scientific Discovery, it's a collaboration with also with the anchor in the sense that the anchor is doing similar kind of experiments with their model. So the team from COLA was led by Ben Cash and Jim Kinter and on our side in addition to myself, Roberto Buizza was also involved in the planning and Sami Sareen actually produced a version of our model that could run on the anchor computed and Sami and Decrimer as provided the data for initial conditions for these experiments. So this is what we can say we have actually done most of it is a combination of six months and two months experiment. The six months experiment only had two start dates, May and November. They cover a 30 year period, 1986 to 2015 with ensemble of 25 members and the forecast go up to six months. In addition to that for each of the two seasons we also run shorter two month integration starting from June, July, August for the summer December, January and February for the winter and these experiments are done with the resolution of again in order 100 kilometers as a reference and these are done with the cubic grid at T199 but the core experiments are the ones with the cubic grid at T639 which gives a resolution of about 18 kilometers. This is a very high resolution for seasonal predictions and so we hope to see some good results from this set of experiments so we will be able to assess the impact of using the high resolution in both the atmosphere and the ocean throughout the integration. Also plan are some integrations with basically the highest resolution we are running operationally at the moment. We could only afford to do two months with 15 members and actually this is what was planned. In total this project requires almost 60 million core hours that we were allocated. Actually what we managed to do were all the low and sort of high resolution runs. For the very high resolution runs we have been able with this allocation to do the winter runs and the summer runs will actually be done in the near future with an extended allocation for this project. So I don't have yet result from this project but I have some result from the Primavera Horizon 2020 project and this is the main goal of this project is to develop a new generation of high resolution global climate models. The hope is that this high resolution will be able to simulate the climate with higher fidelity than what we get from the sort of semi-five generation. So for us you know ECMWF is not in the climate change business but of course we want to evaluate the quality of our couple system to study its attractor. You know when we do sub-seasonal or seasonal predictions our couple system basically drifts from the initial conditions which are on the real atmospheric attractor towards the model attractor but because the ocean has very long time scales in here and we actually never seen the seasonal prediction the asymptotic climate of the model. So we can study the start of this climate drift in our couple system but just from seasonal prediction we cannot see what the asymptotic state of the climate is. So this was our motivation for participating in Primavera and so in Primavera the goal is to do simulations that cover a hundred year period as I will show in the next slide. We are only doing the historical part and we are doing again with two different resolution again TCO 199 as a reference again as I said typical semi-five resolution and then well given the length of the time scale we cannot afford to do TCO 639 but we are using the TCO 399 which actually gives us a grid resolution of about a quarter of a degree so this is a good match to the ocean resolution or the Nemo model we are using in this experiment. So this is the schematic of the so-called stream one integrations that will have to be delivered by October this year. So we have a set of amic type integrations so only atmospheric and land. SST is prescribed basically from the head ISST 2 data set that has been adopted for CIMI-6 experiment and so the historical part will cover 1950 to 2014 with CIMI-6 radiative forcing and then this will be extended with forcing provided by one of the CIMI-6 scenarios up to 2015. Similar runs will be done in couple mode so starting from 1950 but because as I mentioned the ocean has some long adjustment time scale we need to run a spin up period. For this particular experiment we chose not to run the very long hundreds of years spin up which is done by most of the CIMI participating groups but rather to do a shorter spin up 30 to 50 years with constant 1950 forcing and then to continue actually this run as a control experiment and also at some point to branch off start to applying time varying radiative forcing again provided by CIMI-6 and then run the model into the future and again the CNWF will only run up to 2014. So coming back we have some runs completed so we have completed five members of the AMIP at TCO 199, three members at T399 actually we have completed two and one is in the machine and we've also completed a couple spin up which is a single run both with low resolution and high resolution low resolution is one degree and high resolution is a quarter of a degree. Now I mentioned before that we were expecting to see a reduction of the bias in the North Atlantic due to the introduction of the quarter degree model there was some evidence in the seasonal focus but perhaps there is even more evidence here in the long runs so these are maps of the SST biases from this first basically 50 years of spin up of the couple system and of course the spin up is a bit of artificial period because you know it has forcing fixed at 1950 so you don't have an exact realization of that in the real atmosphere but basically here the data are compared with SST in the 1950 decade from the N4 data set and if you look here at this big blue blob this is the low resolution run you see that there's quite a substantial bias while the North Atlantic in the low resolution model and if you look here at the high resolution model you see that this has been substantially reduced so when you let the model adjust you see perhaps more clearly the positive impact of the high resolution in the ocean now another advantage which is related to this is the fact that the high resolution is much better able to maintain the strength of the meridional overturning circulation so you see here the intensity of the overturning stream function from the low resolution run on the left and the high resolution on the right you can see from the number of isolines that the high resolution has a higher intensity and if we actually plot the strength of the transport in Sverdrup at 26 degrees north we can see that basically in the low resolution run there is a pretty long adjustment with the decline of the intensity of the MOC about two thirds of what we expect from observations when we do the high resolution run there's some adjustment in the first say 15 years but then the model stabilizes and it stabilizes with an average transport of about 16 Sverdrup which is quite good in pretty good agreement with observational estimates another advantage of running a coupled system with respect to running an AMEAP system is that you can actually but also the difficulty is that in an AMEAP system you prescribe SST so the warming of the SST is actually defined by the SST you prescribe in the atmospheric model the warming will come from the combination of the effect of this SST and also of the forcing the radiative forcing that prescribed these two things can actually be slightly different so you may have a slightly different trends in temperature in the prescribed SST and in the atmosphere above the sea and this actually is reflected in the transfer the so-called non-solar heat flux so the sum of late and sensible heat and that long wave radiation at the surface of the ocean actually takes some time to actually adjust to a fixed value and this is because you need just as likely different trends in the prescribed SST and the atmospheric flow to actually change this balance by a few watts per square meter now if you instead you actually look at what you get from the coupled run then you can actually see that the again after about 10-15 years the curve of the non-solar heat flux is essentially doesn't have any trend of course it has decadent oscillation because there is internal decadent variability ideally the red and the blue curve should overlap to have an exact balance but actually if you you cannot read the scale but the difference is about one watt per square meter now half of it is actually a sort of post-processing deficiency to actually compute exactly the balance you should actually add to these three terms that I mentioned they basically melting of snow over the ocean because that is an additional contribution to the surface heat balance and that accounts for about it less than half a watt per square meter so if I added that then the difference would only be half of watt per square meter and that is in fact a residual that is within the observational estimates because the earth is warming so if we had an exact balance between the heat that the ocean receives and what it loses then the ocean would not warm but in practice the ocean has warm because we know global warming and in fact the estimate is that on average you have an imbalance at the surface of about half a watt per square meter so when we this particular difference when we take these two effects into account is something that we would sort of expect you can also see here the variation in two meter temperature over land this is the variation of the SST produced by the model the green line that is almost overlapping you can see is basically a two meter temperature of air over the sea this cracker is the measure of the fraction of sea ice cover over the northern extra tropics so as you expect you would have a sea ice decline and I think one interesting thing that I would like you to notice in this graph you know this is a simulation of coupled internal variability with no variation in external forcing and when you have such if you remove this external variation then the critical variability of temperature over land is almost exactly in phase with the variation of temperature over the sea and this is something that a number of authors have actually argued that you can very well model the warming over land if you know what the warming over sea is and you know for example Prashant Sardeshmok has argued this for quite some time but this has also been revived by the recent debate about the hiatus so if you know what your ocean does and how your ocean transfers heat to the atmosphere then you are able to a very good extent to predict what the warming over land will be and not only it's long term trend but also it's the critical variability now let's move to more regional indices and of course you have a coupled system you want to know whether you get a good El Nino or not and so at the top here you see the plot of the El Nino 3.4 SST anomaly from our REMI plan so in this case it's not a modern simulation you know this is actually prescribed from the other that I SST 2 data sets so you see the peaks in the major El Nino events actually it turns out that if you look to the El Nino 3.4 and even more so if you look at El Nino 4 some of the later El Ninos actually have a stronger signal this is because in recent years we have a more stronger prevalence of SST anomalies in the central part of the Pacific some events that people call El Nino modoki rather than El Nino events that were occurring earlier so if you actually look more towards the central part of the Pacific and more recent events become more prominent with respect for example to the 92-83 now what is the coupled model doing again this is so this is unconstrained a free run of the coupled system and the bottom graph shows you the 3.4 SST anomaly from the spin up run I skipped the first 10 years again I mentioned that yeah probably for another 10 years the model is adjusting basically after sort of 15 years into the run then you have a rather realistic variability of El Nino with actually peaks of about 2.5 degrees which are comparable with the strong El Nino cases like the 2015-16 and actually in this integration we actually happen to have three pretty strong El Ninos separated by about 5 to 6 years and then periods of more moderate events so it will be interesting to see what happens when we continue these integrations also the patterns we have looked at the connection between the SST in these regions and the precipitation in these regions and so these are maps that basically show the covariance between SST in the Nino 4 region and SST everywhere or for the bottom panel the covariance of SST here with precipitation everywhere else so here you see the result from the EMI run so in this case again this is just the result of the prescribed SST on the right hand side I put panels from the era interim period and here the difference comes from the fact that era interim used different SST but also that era interim focuses on it's only available from 79 onwards and in this period we actually had three pretty strong El Nino events so the amplitude of El Nino computed from just the last three decades is actually a bit larger than the amplitude that you get if you extend back to 1951 so if you look to the simulation of the precipitation anomalies which are induced by this SST anomaly you see that the model is doing quite a good job we were somehow expecting this because we are not expecting dramatic differences in our EMI plans with respect to our season forecast but the good news is that even when you do this exercise with the couple system you get some very realistic teleconnections and so now even the SST map is produced by the model so basically this tells you what is the correlation between the Nino for SST and SST everywhere else and you can see that the structure is actually quite realistic and also the precipitation associated with this anomalies is actually quite good okay let me skip the Indian Ocean in the interest of time let's move to the northern extra tropics one characteristic of recent versions of the CMWF model is a very good climate for the northern extra tropics and you can see here map of mean geopotential height at 500 hectopascal on the left hand side is the climate of the EMI plan so we have at the moment two members run up 1950 to 2014 so we have about 130 years of simulation and these are compared in this case I wanted to have a map which is actually for those particular years so I actually took the data not from IRA Interim but from our latest 20th century re-analysis with the coupled IRA 20C and that gives a more comparable map and you can see at the bottom right the difference between the two the contour interval is about 10 meters so you have a maximum biases the maximum features are of the other plus or minus 30 meters which is quite a good result but perhaps we were even more positively surprised by the fact that even the coupled system has a bias which does not exceed 30 to 40 meters which again is a sign that the climate of the system is reasonably stable in this case as I mentioned there is no constant 1950 realization in the atmosphere so if you actually take the fields without doing any correction the whole map of the difference will turn out to be blue because a 1950 atmosphere is colder than a current atmosphere and therefore the geopotential height will be a bit lower so what I did is that I actually applied the correction by basically re-centering this map at the center of the periods covered by these two re-analysis by using the mean trend in geopotential height estimated from the IRA 20C and if you do that then you see structures like the one you are seeing you can actually notice that the bias is actually smaller if I compare the system with the IRA interim period and that we think is because the model actually had three strong and linear during this first spin up period and so because of the answer to the connections the geopotential height looks actually more comparable to the geopotential height in a period when we had three strong and linear in fact if you actually look at this particular map of bias this looks very much like a linear connection so we actually think that maybe some of these features may even be reduced in a longer run of the system if a more balanced alternation of strong and weak and linear will come out we have looked at individual modes of variability like the North Atlantic Oscillation doing your F analysis and again covariance analysis and again the result the good news is that the results of the couple system are actually comparable to results of the AMIP and again you know pretty good agreement with observations I think I will just in the interest of time get to the summary so we have seen basically three sets of big experimentations one devoted to the improvement of our Season of Focus system the other two are actually research experiments so the new Season of Focus system that we are going to introduce in autumn shows reduced bias in tropical SST improves killiness of prediction with respect to the current system performance in the northern extra tropics is I would say comparable some aspects are improved some aspects are not improved but definitely if you are interested in the tropical couple system is definitely a better performing system the Mantis experiment have just been finished we are going to assess the results in the next few months so we will basically be able to see the benefits of using high resolution as we use in the medium range ensemble but when we extend this to the sub-seasonal and seasonal range and then from the Primavera project we will be able to assess the climatology of our system basically once it reaches its asymptotic or near equilibrium state of course the deep ocean may still drift but at least the upper ocean will be in a reasonable balance and so at the moment we have just run a couple of spin-ups but as I showed the preliminary assessment of this run with constant 1950 forcing shows encouraging results in the sense that the model climate stabilizes after about 20 years and the biases are comparable with those of the impact experiment so we will be able to use these runs actually to assess really well the climate of our couple system for long-time scales and see what kind of errors the model simulates once it is basically allowed to adjust to its own climate and that's it for me