 Our next speaker is Antje Weishheimer. And Antje kindly also gave a lecture last week. She's at EASMWF and the University of Oxford. And in her free time, Antje likes to ramble the English countryside. Thank you, Judith. Nice memories from years back when you were in England. Thanks for all the possibility opportunity to speak. Let me share my screen. Can you see that? Okay, Judith. Yeah, perfect. Okay, so I'm gonna talk a little bit about some recent work I've been doing with my colleagues from EASMWF and also from the Bureau of Meteorology in Australia on multi-decadal variability in long range, only new predictions during the 20th century. And so the motivation for this really is, I mean, as I'm sure you all know, that any new ENSO is arguably the most predictable climate mode at seasonal time scales. And it provides the scientific basis for global seasonal climate predictions. And over the last years and decades, significant progress has been made in our understanding of ENSO and the complexity of ENSO, but also in the development of the observing system, systems, the model development of couple GCMs of data assimilation techniques to improve the initialization of forecasts. And it's probably fair to say that current forecast models can provide effective predictions of an Indian woman cold events approximately a year had six to 12 months ahead. Estimates of this type of predictability and so skill are commonly based on retrospective forecast, most commonly over the last two or three decades, which results in an overall quite small sample size of ENSO events. They come with a substantial uncertainties in the skill estimates. On top of this problem is this small sample size, decadal scale changes in ENSO background states can make certain decades less predictable than others. And these background fluctuations can be seen as some spontaneously generated multi-decadal variations in ENSO diversity. So for example, I mean, I listed a few papers, all the papers on this topic here, changes in a tropical Pacific mean state related to the equatorial thermocline, they have been suggested as a source of inter-decadal modulation of the amplitude of inter-annual variability and survivability with an intensification of the ENSO signal observed in the second half of the 20th century. Especially the papers by Kurtman and Chopf here and others looked a bit more in detail of how changes in the ENSO amplitude can affect ENSO predictability. And they proposed a relationship between the amplitude and the variability and based on some sort of simple models at a time which sort of could be explained with the classical theory of the delayed oscillator mechanism which robustly maintains self-sustained oscillation that drive the SSTs anomalies and so on. But also stochastic noise from the atmosphere, for instance, through westerly wind burst can influence SST anomalies. And it was shown that these effect them stronger during epochs and the delayed oscillator is perhaps damped by colder SSTs leading into overall reduced predictability. So an in-person attempt to re-forecast historical ENSO events was done in the sort of seminal paper by Chen et al, 2004, who run the Zebjörg and Cain model by initializing with some reconstructed SST data. So the Zebjörg and Cain model, sort of model of the coupled pacifical tropical ocean atmosphere. And they showed in that network that the ENSO predictability depends on the time period over which it's estimated and that periods with high skill were dominated by strong ENSO events. And yeah, I mean, basically motivated by these earlier studies, we here try to use a state of the art modern now with our standards nowadays, modern dynamical forecasting system, the ECMWF forecasting system and try to revise where we are in terms of predictability based on these earlier works with the more idealistic model setups. So the purpose here of this work is two-fold in a sense that we'd like to explore long range retrospective ENSO predictions in the 20th century. And when I say long range, I mean up to two years with the state of the art dynamical seasonal forecasting system. But we also wanted to test sensitivities to the ocean observing system and to atmospheric force things here. So what we did is we created a set of hindcasts we forecast of the 20th century. We call them CS5, 20C, CS5 because they were done with the ECMWFs. It was a version of ECMWFs operational seasonal forecasting system, CS5. A version meaning we're using the same model as such but we run it in a lower resolution because we couldn't afford the higher resolution of the operational forecast. So it's the resolution of TCO 199, 91 vertical levels. That's roughly how many kilometers? I think it was the order of 60 kilometers. And the ocean version, the ocean model has a resolution of one degree with 42 levels. We run forecasts during the 20th century here over the period from 1901 to 2010 for two start dates per year. We initialize on the 1st of May and the 1st of November over these 110 years and let the model run for 24 months. So we have these biennial forecasts here. There's an ensemble size of 10 members because the focus here was really on the new. This was possible due to the availability of the CRR-20C coupled reanalysis from ESMWF which covered exactly this periods which we use to initialize these coupled predictions re-forecast just a few words on the CRR-20C reanalysis. So it's a coupled reanalysis that comprises the atmosphere, land but also wave, ocean and sea ice. It assimilates only surface pressure and marine winds in the atmosphere and for the ocean it assimilates subsurface temperatures and salinity profiles. So with this data, we created our re-forecast experiment our control experiments as labeled CS5-20C. This table here gives an overview of the initial conditions. So in this control experiment the atmospheric land initial conditions as well as the ocean and sea ice initial conditions come from this coupled reanalysis coupled CRR-20C that uses data assimilation atmospheric forcing it's a coupled setup but we then run some sensitivity experiments with a similar setup where we tested the influence of the ocean observations. So we have experiments without data assimilation node DA in the ocean which use the same atmosphere and land initial conditions from CRR-20C and the atmospheric forcing. So this is not a coupled, there wasn't a coupled setup in the initialization came from also from CRR-20C. This is the orange, the second line here. And then we have two experiments where we force the ocean initial conditions with a different atmospheric forcing and instead of coming from CRR-20C it comes from IRR-20C which is the previous generation of atmospheric reanalysis of the 20th century, which was atmosphere only. So they have the forcing that the ocean initial conditions saw is different and we have two experiments here one with ocean data assimilation and one without. So we're interested in the skill and here is perhaps an overview plot of how the NENU 3.4 SST Ensemble Mean Anomaly Correlation skill has changed over the 20th century over our hind cast period based on the November 1st initialization here and let me talk you through this plot. So we see on the x-axis the time of the re-forecast period it goes from 1901 to 2010 and we compute the skill over 30 year moving windows, hind cast windows. So each point in this plot is estimated over a 30 year window and plotted at the center in the middle of that 30 year window. That's why you have these, you know the wide ends at the beginning and the end of the period. And then the y-axis shows the lead time of the forecast our two years and it goes from bottom to top. So we start in November so the first season here is DJF and this is a three months moving window across the lead time here in the vertical axis. The colors indicate the level of skill, the correlation skill and the dashed areas are periods in terms of hind cast periods and lead time where the skill is not significant according to a symbol T test. So what do we see? We see lots of interesting things. You can look at this plot from various ways and let me just start with one perhaps. So if we do cross sections here for different epochs in the hind cast period, you can see how the skill behavior over lead time changes. And this is shown in a plot below. So if we take the last data point here plotted at around 95 which includes the data from the last 30 years in our hind cast period and we do a cross section across lead time. We end up with the red curve in the plot below here. So we see the solid lines is from our simulations and the dotted line, the red dotted line is a simple persistence forecast persisting the initial conditions, the SSTs. And we see that we have to, you know, after the first end so peak season, we reach the predictability sort of barrier, spring barrier here. But after that, we have quite a long, like almost a year long period of a plateau of very high levels of skill, highly significant here. And then we reach the second spring barrier and then we lose the skill after that for the latest forecast period. If we go to say a period here in the middle where we see that we lose lots of skill after the first spring barrier, here's lots of non-significant skill lead times here. This is the blue line in the plot below. And we see already right from the beginning the behavior is quite different. The skill drops much quicker and we observe also like a plateau of skill but is hardly significant here. And it shows like quite reduced levels of skill after this spring barrier out on seasonal, longer seasonal timescales. And then at the very beginning of the century, interestingly, we see in the yellow line here, the first on the seasonal timescales, the first months behave relatively similar to our most recent periods, high levels of skill. We reach a plateau after the spring barrier which is a bit lower than the one for the most recent period but still significant in terms of levels of skill. And then again, and this sort of after the second spring barrier we lose that skill here as well. I mentioned that we did... Oh yeah, this is another way of looking at it. Of course, cross sections, not across lead time but across hind cast period. So this is now looking at the first winter season, peak season of El Nino lead time to four months, cross section here as indicated by the black line. Just look at the red line for now which is our control experiment. And you see similar to findings the people from CPC found. We have a long period from the most recent decades to the roughly 1960s where the level of El Nino skill forecast here is constantly very high. You see that is 0.9 something here for a long period of time. And then around the 1960s, the skill drops a bit and there's a prolonged period of a few decades where the skill is lower to roughly the 1930s and then the skill goes up again and the skill at the beginning of the century is very high. It's almost as high as at the current for the latest periods in that century. The gray ban show gives an idea of the uncertainty. Of course, with the correlation skill we have an upper limit of the uncertainty there at one but you see that also the uncertainty has much increased during that period when we have a bit of a drop of skill. I should say, because this is El Nino when I speak about a drop of skill this is still high levels of skill in terms of other parts of the world. We're still talking about 0.99 correlation skill in the periods where it's lower. You have about three minutes or so to wrap up. Okay, thank you, Judith. If you look at the second DJF season so this is then lead time 14 to 16 months we can do a similar cross section here. Again, just look at the red line and we see some interesting variability as well and during that period in the 1930s, 40s the skill then becomes non-significant here where it is, it still is at the beginning of the century and at the end of the century here. Just to mention we did these experiments to test the sensitivity to the ocean observing system the data assimilation system and we see on the left-hand side at the top the experiment I just showed you the control one the same plot and below there is the experiment where we don't assimilate ocean observations for the initial conditions and you see if you compare these two you see the impact that the ocean observations have for skill especially in the second half of the 20th century when we have relatively good observations in the ocean we see how beneficial the ocean observing system is on these lead times after the first spring barrier in particular but we also see that the levels of skill at the beginning of the century are not much affected by the changes in the observing system or the assimilation of these data here. Just briefly on the right-hand side you see the sensitivity to the atmospheric forcing so using a different atmosphere to force the ocean initial data and there we see a clear sensitivity at the beginning of the century where the sensitivity is most pronounced here and again you can do all kinds of games of doing cross-section and we see for these sensitivities we see a clear separation perhaps it's clear as to see on the bottom right plot that the two lines that have the highest skill are the ones that use data assimilation in the last decades and the last part of our hind cusp periods whereas the two simulations without data assimilation really have lower skill here. This is very different at the beginning of the century at the beginning of the century it's the simulation that used the forcing from the zero 20C atmosphere at a higher skill and the ones that used the forcing from the other atmospheric data set have a lower skill so there's some interesting sensitivity behavior here over time. Also interesting but I won't talk about it now is the behavior for different initialization time of the year in the May start days as we all know and so it has strong sensitivity to a seasonal cycle and it is clearly visible in these other start dates but just to mention as a possible explanation for some of this behavior and I mentioned it in the introduction perhaps we know that the ANSO amplitude the ANSO inter-annual variability has undergone a strong increasing trend from the middle of the century to the end of the century this is here shown in the black line again 30 year moving window and this was hypothesized that this increase in the amplitude the signal strengths is one of the reason why predictability why these skill measures went up for these periods as well and well I won't have time to talk about loads of this I just want to show you if you look at the red line this is the similar estimate from the model that forecast with a lead time two to four months and it's good to see the model are able to simulate that increase in ANSO variability in the second half of the century as well but you also see that at the beginning of the century when the amplitude was also relatively high that is not so well predicted on seasonal time scales with our models here and if we look at the other colors these are the longer lead times we lose that behavior and the amplitude completely but let me briefly summarize so I was trying to show you some results from an interesting data set that we recently produced using the ESMWF forecasting system over a long period hind cast period from 1901 onwards we looked at the multi-decadal variability of ANSO skill and found constant and quite high levels of skill from at least the 1960s onwards up to the first spring barrier and we also noticed a clear benefit of the recent ocean observations with a skill extending up to the second spring barrier here 18 months but also saw a pronounced intermediate drop in skill between the 1930s and 1960s and then before that the earlier decades of the 20th century high skill despite the poor observational coverage which is interesting we saw some sensitivity to the atmospheric forcing especially for the first part of the century and the importance of the spring barrier especially for the May star dates I didn't talk much about it but that was some interesting motivation of maybe looking at longer term predictions skillful predictions also for like perhaps the future operational systems here and there are several hypotheses for the courses of these non-stationarity in the skill to do with the amplitude the interannual variability and we saw the last slide some of these behavior but also with changes in the intrinsic persistence characteristics which I didn't show here but we're looking at these and variability of the discharge and recharge strengths in the endso-cycle and with this I thank you very much Thank you very much We have time for a question or so please oh yeah Zane has one please go ahead Hi Anci I really like that talk I was wondering if you looked at the thoughts some of the talks yesterday when you think about this potential predictability in these runs you know can you have a bunch of ensemble members so have you have you thought about that at all and whether that you maybe wouldn't expect it to change over time but then again like I don't know what to expect so have you looked at that or thought about that Good question Zane I think I did just too many data to remember at the moment I didn't notice anything spectacular but I should have updated myself a little bit on this In terms of non-similarity Nothing spectacular you mean it was just it was just flat pretty much over time there wasn't much change I don't think it was as dramatic as what we saw in the actual predictability estimates Yeah you might expect that Yeah thanks Thank you I actually will ask a related question and this is have you looked at that like some really long simulations I mean with CSM we have like a thousand years and so it would just give you more data to look at that modulation of the NAO and then so NAO is the other talk I know people have done this I mean Richard Greatbatch group they have been looking at some pre-industrial CMAP simulation and so for the teleconnection into the exotropics Unfortunately for the ECMWF model we don't have these long simulations Right, no, no, yeah, no I was doing things No, I'm more coming from the prediction you know the operational aspect of the I fully agree that would be interesting to do and I'm sure people have done and found I mean these earlier earlier works I guess when the initialization was a bit more of a problem people looked at free-running models to check variability there and there's quite a lot of literature out there I haven't looked at it myself Thanks so much Anna, Jacqueline has a question and we'll make this the last in-person question Hi, thank you for the talk I was wondering if you have looked at physical processes you showed that the model has a high score for the leading time so from two to four months and that it's an indication that the model is picking up the ANSO the El Niñon set, right there are physical processes already in the Central Pacific perhaps higher or warmer SSD anomalies or like weaker trade winds I was just wondering if you kind of went beyond just the correlation analysis to look at like what insights do we have from the models itself Yeah, it's very much ongoing work to look at this data and so we have looked at a bit of like trying to see mechanisms with the ocean heat content and the discharge recharge mechanism and especially with the view of trying to explain the different behavior and the different epochs there and there's some indication from the sort of persistence auto correlation behavior that you can see there the atmospheric wind forcing things I haven't had a time yet to look at I mean one of the purpose of this was really also I should have said it's clear that we very much invite people to look at the data it's a huge amount of data available a resource for several questions and we have collaborations with the Australians for instance who look at certain aspects and I don't have enough time to look at all these things but if people are interested we would be very happy to share this for my for more explorations really All right, thank you, thank you so much