 O ffordd eraill yn cael gweithio'n edrychwch ganhaeddiol a fy hollwch gweithio'n gweithio ac i'n rhan o'r ffrwyng i thysg iunintelligible oed, ym blwysig ar y deutii clubs o'r cyd-gweithio'n gweithio, mae'n gweithio'n gweithio'n gweithio bobl yfeddoli, teimlo oherwydd rydyn ni'n ni'n rhaid o hollwch chi oedd yn gweithio'n gweithio'n gweithio'n gweithio'n meddwn ni'n arbyrddol o'r hyn. so, so, my tawt will be largely review, although I'll also show one or two new results, and raise a few of the points that I hope we can discuss in the discussion sections. I'm going to talk about attribution of decadal change, and fundamentally why it's so difficult. It's not, however, going to be solely a sort of a discussion of how terribly difficult it is. I'm also going to be talking about some of the opportunities and some of those opportunities relate to the rapidly growing field of decadal prediction that of course many people in the room are involved in. So the links between attribution and prediction is one of my themes. How do I change the slides then? Well this thing does, does it? Okay, I've done the wrong thing already. There we go. Okay, so as everybody knows and we've already heard this morning I think, the evolution of climate on decadal time scales depends on two fundamental types of processes. One is the internal variability that Mojib talked about and the other is the response to a range of forcings and we can think both about natural forcings like volcanic or solar forcings and progenic forcings. And decadal prediction of course, the basic challenge is to work out well what's that combination of internal variability and responses to forcings that is critical for the next decade. That's the decadal prediction problem. Whereas the attribution problem looks back to events in the past and says well actually what was that particular combination for particular specific observed events. And that includes to what extent the events were predictable because of course some aspects of climate are more predictable than others. So I think it's a fairly obvious but important statement that if we're going to have any confidence in decadal predictions we've got to demonstrate that actually we're really good at the attribution problem. That we can look at all these past events that we've seen in the talks from Jochen and Mojib and say we understand why that happened and we know to what extent it was predictable. Now clearly we're a long way from that situation now but that is what we need to be able to do in order to really have confidence in decadal predictions. So just to say a little bit more about attribution, there are typically two levels of attribution. There's what you might call proximal attribution and ultimate attribution. So proximal attribution is all about what were the processes that were important to some particular event that we saw. Maybe it was some shift in the storm track or maybe we have some evidence that changes in sea surface temperatures really mattered for this event. And then the ultimate attribution is about this distinction between internal variability and forced responses. And both types of questions are important and of course to have a complete understanding we need to be able to do both the proximal attribution and the ultimate attribution for any particular problem of interest. And of course when we're looking at the ultimate attribution that includes questions about well is the forced response just a linear addition or is there actually some interesting interaction here. Perhaps the forced response is manifest in terms of changes in internal variability as we've heard about and that's something that we need to unpick and that can often be quite difficult to unpick. We should also recognise that when we look back at observational records we can see interesting things going on and sometimes there's nothing to do with climate processes there to do with observational issues. So we have to bear that in mind. So we've heard some examples of attribution problems and I think there's no better illustration of how difficult attribution is than looking at all the debate that has been about the so-called hiatus in global mean surface temperature over the last decade. In broad terms we understand that this is some combination of internal variability and forced responses but of course quantitatively the uncertainties remain large and that's because it's a tough problem and I'll say a bit more about that. Some other examples so Atlantic multi-decadal variability mode you've introduced and I'll say a little bit more about that and the Sahel rainfall problem that Yocanon talked about is another good example of really significant decadal time scale changes in the climate system that we need to be able to unpick and get to the bottom of. So why is it so difficult the attribution problem and the prediction problem? Well it comes down to three areas of ignorance as indicated there. So we have ignorance of the characteristics of internal variability in the system. We have ignorance of aspects of the forcing. For decadal time scales the issues around aerosol forcing and volcanic forcing are particularly important. Greenhouse gas forcing is better quantified on the whole and we have ignorance of the responses to the forcings. So we've got these three areas of ignorance. This slide shows the relative importance of these different areas of ignorance when we think about predictions or projections here of global mean temperature and this shows the total uncertainty and then this shows normalised as a function of lead time normalised by the total uncertainty area. And if you look a few decades ahead you see how this is the internal variability and this is the response uncertainty. So those are the most important sources of uncertainty typically a few decades ahead. Further ahead uncertainty about the forcing in this case this includes greenhouse gas forcing becomes increasingly important. This is for global mean temperature. As we come down in spatial scale and time scale then the internal variability becomes increasingly a big issue. So ignorance of internal variability this is a very nice slide that Ed Hawkins put together which illustrates the magnitude of the problem. So these are a set of current climate models and these are simulations of internal variability in global mean temperature from control simulations with these climate models. And what you see is a huge range of behaviour variations in the magnitude of variability and the spectrum of that variability. And these are state-of-the-art models and actually we don't know which is the best model here. We can have a go at saying which might be implausible but actually it's quite a hard issue to constrain because for the real world we can't cleanly separate internal variability from forced responses. So what is realistic decadal variability is a tough problem. Ed did a nice analysis here where he just looked at could we come up with some simple constraints from observations. So in particular what he looked at was just taking the observations and then supposing that the forced response was either a linear trend or some kind of high-order polynomial fit. And then if you look at the residuals and say well we'll call that the decadal variability then you can come up with some constraints that look like this. So on the horizontal axis here we've got the standard deviation of global mean temperature. On the right-hand axis we've got the lag one autocorrelation so how red is the variability. And so this area in the middle might be the set of models that are if you like not obviously inconsistent with observations. And you can see from this basic analysis that actually quite a lot of models do seem to be inconsistent with observations even in this basic parameter of internal variability. So that's important and sobering. And that's just global mean temperature. So then if we look on smaller scales then actually the issue in many ways is even more acute. So this was an analysis that Ed and I did a few years ago and it shows if we just focus on the top panels here. It shows the magnitude of internal variability in surface air temperature obviously regionally here. And the spread amongst models this is amongst the same in five models. And what you find if you look in certain regions that there can be easily be a factor of three in the magnitude of the standard deviation of annual mean surface air temperature between different models. So that's a huge range. And again it reflects how difficult this problem is and our larger uncertainty. So we need to make progress and obviously a key issue here is better understanding the mechanisms of internal variability. And just linking with something that Mojib talked about there how the mechanisms can be dependent on the mean state of a climate model. This is a nice just one figure from a nice paper by Matt Menry who works at the UK Met Office and recently completed a PhD with us in Reading. And this is an analysis of the CMIT 5 model simulations of North Atlantic decadal variability. And the interesting finding here is the high correlation between the bias in these models in the North Atlantic in fact in the Labrador Sea here. So this is the bias in temperature and salinity. And on the vertical axis is a measure of the relative importance of temperature and salinity anomalies for controlling density anomalies. And those density anomalies are critical for affecting the MOC for example, the overturning. So what this analysis shows is that the mean state of the model is crucial to get right if you want to have a reliable simulation of North Atlantic decadal variability. So that's a tough problem. You can't just say well we'll just look at anomalies and not worry too much about the mean state. You've actually got to get the mean state right if you're going to get North Atlantic decadal variability right. Coming on then to response uncertainty. So this is the question of what is the response to particular forcings. It could be greenhouse gas forcings, it could be other forcings. Clearly there's lots of discussion about climate sensitivity, the transient climate response for example. Everyone knows that models show a large range in climate sensitivity. But when we're looking on regional scales there are many other dimensions to response uncertainty. So in particular the issue of circulation change in the atmosphere and ocean is a crucial issue for understanding the decadal variability and change on regional scales. A lot of uncertainty about circulation change as a nice review by Ted Shepard in Nature Geoscience focused on atmospheric circulation in particular. And we need to be very clear that so often we estimate this response uncertainty as in this analysis here, sort of the blue part here. By just looking at the spread amongst different models we say this is some kind of measure of our uncertainty. We need to be very clear that that's a very crude measure of our uncertainty. There are processes missing from all the models and so our true uncertainty is almost certainly larger. The real world may well do things that aren't in any of the models and that in some sense is part of the response uncertainty and we need to understand that. So the big issue here is what I call model adequacy. Are the models a good enough simulation of the real system in terms of their internal variability, in terms of the forcings and in terms of the responses? This is the big issue that's very tough. So let me come to a couple of examples. Well, particularly I'm going to focus on the Atlantic. So the Atlantic is very interesting. This is obviously a smooth record of its sea surface temperatures. But the interesting thing about the North Atlantic is that the decadal variability is large by comparison with both inter-annual variability and actually comparable to the long-term trend in the North Atlantic. That's not true of other regions of the world. So the question obviously arises why is it there? Is it purely internal variability? Is it some response to different forcings? And people have made arguments about the importance of aerosols, greenhouse gases, volcanoes, solar variations. There's evidence about the importance of the overturning circulation. That could be purely internal variability or it could be influenced by forcings. So this is a really interesting problem and it's been the subject of much debate, as many of you will know. But the state of play is there is no current consensus. So let me hope that someone in this room will solve this problem in the next decade or before. Just to go into a little bit more detail, I'm just going to talk briefly about a couple of studies here. One a couple of years ago and one more recently. So one of the ideas about North Atlantic variability, if you like the simplest idea, is that... So if we're trying to explain the variability in sea surface temperatures, a simple idea is that that might just reflect the variability in surface fluxes. And two papers that have essentially argued different versions of that theory are shown here. So the paper by Ben Booth et al argued that changes in anthropogenic aerosols have been the cause of this multi-decadal variability in sea surface temperatures. And the basic idea there was that the aerosols reflect in a more short way, either directly or through their effects on clouds. And that modulated surface fluxes and modulated sea surface temperatures. More recently there's this paper by Clemens et al, which argues that all there is to the AMO is essentially an ocean mix layer response to a modulation in this case of the turbulent fluxes more than the radiative fluxes. So these are interesting ideas, but we should be sceptical. So we need to ask the question, are these models adequate in their representations of internal variability forcings and responses? And related to that question is how are the models compared to the real world? This is a key question. So with regard to the Booth et al paper, so Rong Zhang and others in the room wrote a paper which pointed out that one needs to look at a range of metrics. You can't just focus on sea surface temperature in order to unpick all the processes involved. You've got to look at all the relevant variables. And this paper points out that if you look at a set of relevant variables, for example, north Atlantic heat content, salinity and so forth, then actually there are many discrepancies between those model simulations and the real world. So that we shouldn't jump to the conclusion that just because the sea surface temperatures have evolved in a way that looks similar to the real world, that means that the model is capturing the same processes as dominating in the real world. Another study which I was involved in looked at the observed evolution during this cooling period and found that there are interesting changes in atmospheric circulation during that cooling period, which indeed seemed very likely that they modulated the turbulent fluxes and that that certainly played a role in the evolution of the cooling. There are many other things going on in fact in the north Atlantic that I don't have time for, but I think the basic message is that it's actually quite complicated. There are many processes going on and an overly simple story of Atlantic multi-decadal variability is unlikely to be correct. The other line of evidence that I want to just talk about is the issue of ocean heat transport. So if you have the paradigm that you just need an ocean mix layer in order to understand Atlantic multi-decadal variability, the implication is that ocean heat transport plays no role. Well actually there's a huge body of evidence that ocean heat transport does play a key role and I'll just touch on some of that evidence. So this is from a paper by John Robson who's over there which demonstrated that the evolution of heat content in the north Atlantic in the 1990s, so these are analyses on the left and then model simulations on the right and you can see rather a good agreement between the model simulations and the analyses in terms of the evolution of ocean heat content. And of course what's nice about the models is you can do a closed heat budget, so you can actually work out what were the processes that really mattered. And you can demonstrate that changing ocean heat transport were first order important for this very interesting rapid warming of the north Atlantic. These are heat content records. This is the SST which occurred in the 1990s. And this has been independently supported by a more recent study from NCAR, so it's not just one model. And interestingly it's also backed up by a number of independent studies looking at the predictability of this event. So if you initialise the ocean state you can actually predict this very interesting warming event. And in these three different model systems it's essentially the same mechanism. It's the predictability of an increase in the north wood heat transport that's fundamental to capturing this warming event. So there's lots of evidence that ocean heat transport is fundamental to Atlantic multi-decadal variability. Interestingly there also is some evidence and this is something that needs looking into with other systems that the cooling in the 1960s. Again the change in ocean heat transport was first order important for that cooling. And there is evidence that actually we might just now be going into a state, not unlike the 1960s, where the north Atlantic is starting to cool again perhaps quite rapidly. These are the recent trends in ocean heat content compared to the previous trends and John Robson is going to talk about this further tomorrow. Very interesting times in the north Atlantic. How many, two minutes? Okay, so how rainfall? Very briefly then, so this was just another example. So we, this was the study that Sir Boo and Dong and I did looking at the recovery and cell to hell rainfall that's occurred since the peak of the severe drought. And we were interested to understand the mechanisms that may be going on here. And what we found in these, so these were atmosphere model simulations and we found that across a range of metrics and you can see what they are here. So not just the precipitation but for example the surface air temperature, sea level pressure, aspects of the African easterly jet and so forth. These simulations are able to capture the observed changes quite well which gives us some confidence that the same mechanisms are going on in the model as in the observations. And then if we unpick the relative importance of different forcings the interesting and surprising finding in this particular model was that in fact it's the direct, so these are, this is the impact of changes in sea surface temperatures changes in greenhouse gas and aerosols and changes in greenhouse gases. And actually in this model it's the direct impact of the changes in relative forcing that seem to be particularly important for driving the recovery of the hell rainfall. Now what about model adequacy in this case? Well it's a very fair question and absolutely this study needs to be repeated with other models but it's an interesting idea that actually the direct impact of increases in greenhouse gases has perhaps been very important for the recovery of the hell rainfall. I probably don't have time for this last slide although it's possibly my favourite. So have I got time for this, Jocelyn, or shall I just wrap up? No time but go ahead, I like that. Okay, so it's a little bit of a digression but we tend to think of attribution as being reliant on general circulation models and of course to a large extent it is but the reason I like this analysis is that it isn't reliant on general circulation models and it's just a very simple analysis that I think tells us something important. So what it shows is simply a regression of decadal mean surface air temperature in individual locations against the global mean decadal mean surface air temperature over the instrumental record and these are the maps of the regression coefficients and this is the fraction of variance explained and what you see is that global mean temperature explains 60% or more of local variations in decadal mean surface air temperature over most of the planet over the instrumental record. Now what's interesting about that is that if we had a system that was purely internal variability and here's an example from one model but other models have differences in detail but the fractions of variance are always similar then not surprisingly but importantly the fraction of variance was very low if we had a system that was dominated by internal variability by contrast if we had a system that's dominated entirely by false responses then in principle but if the false responses are entirely linear then in fact we get 100% of the variance explained. There are some interesting regions where perhaps the false responses are not linear but the fact that these fractions of variance are as I say 60% or more is I would say very important basic evidence that we're actually looking at a system dominated by false responses over the last 100 years or so and that as I say that's the result that's not dependent on specific GCMs. The North Atlantic is interesting. So some conclusions there. So there's a close link between attribution and prediction. I didn't emphasise but I did illustrate that I think initialised predictions are a really nice tool for attribution and we should think about them in that way. In order to have any confidence in the predictions we've got to get a lot better at attribution so that's a challenge. Model adequacy is something that we should continually debate. If people show you a model result you should ask yourself have they shown you enough evidence that this model is good enough for the aspects of climate that they're talking about to convince you that this might also be true of the real world? That's a fundamental issue. We can get some confidence in those issues by looking at multivariate fingerprints but there's no simple recipe for how to get confidence. You have to dig into the details. The last point there about AMV I think it's more complicated than a mixed law response to what the fluxes might be doing. I'll stop there.