 It's my pleasure to welcome Professor John Methuen. Some of you also know John through the hydro tutorial. John and a lot of work along with Andy to organize the hydro tutorial for this summer colloquium. John is a professor of atmospheric dynamics at the University of Reading. John's research interests cover two distinct areas, one of which is the dynamics of weather systems with an emphasis on the role of Rossby waves. And the second one is airborne transport of atmospheric constituents, such as water vapor and pollution with an emphasis on chemical transformation during intercontinental transport. John is a co-chief editor for the QJRMS, Quarterly Journal of Royal Meteorological Society, and a co-chair of the WMO Working Group on Predictability, Dynamics, and Ensemble Forecasting. John's also the department director of internal partnerships at the University of Reading. Thanks, John, looking forward to your talk today on relating on some of the issues. Thank you very much, Anish, and thank you for inviting me to speak and to take part in the whole summer school. So the topic I'm going to consider today is actually related to the monsoons that we've been hearing about earlier today in that I've been talking about predictability and predictive signals in Southeast Asia in particular. And then I'll also talk a bit about relating ensembles to high-impact weather scenarios with examples from the UK, actually, forecast in the UK, which I don't see really influenced by the monsoon. So the motivation here is quite a broad one. We would like to predict high-impact weather and, of course, the impact. So floods, wind damage, anything that we have weather-related damage, and users want that information as far ahead as we can give it to them. But, of course, it needs to be useful information. And we know that the skillful range of prediction falls fast and is highly scale-dependent. So we have kind of naive prediction just based on climatology from the seasonal cycle, but also in modifications to the seasonal cycle. We have longer-range prediction in large-scale modes. And we've heard a lot about that during the summer school. Obviously, synoptic weather systems. But we need to go even further downscale for the high-impact weather variables themselves, like extreme precipitation or damaging surface winds. And so in this prediction, we need to consider the handover from the most predictable scales to the less predictable things that we're actually interested in. And in particular, in this talk, I really want to address how we harness the more predictable phenomena to obtain useful forecasts of weather risk, or risk associated with a high-impact weather. And sometimes people call this a kind of raging conditioned approach. So I'll explain that as we go along in several contexts. So first of all, we need to use ensemble forecasts. I think that goes without saying. And in this approach, we want to characterize the risk associated with a predictable phenomena. So you say a phenomenon is in a particular phase, or it has large amplitude that might change the risk of some high-impact weather, for example, daily precipitation above some high percentile. And then we use the ensemble. This is the dynamical ensemble to predict the predictable parts, not the precipitation directly. And then we obtain a raging conditioned forecast by using the combination of that statistical information with the dynamical forecast. And one complication, which comes up most obviously in tropical cyclones, but in many other contexts, is the ensemble may split into really distinct clusters. So then forecasters need to identify this behavior and then use it in their communications. But it's not immediately obvious how to see that. So here's a good example. This is actually the forecast for today in the UK. And you notice it's actually got an amber warning at the moment for extreme heats in the Southwest. So it's pretty hot in Reading actually at the moment. We're kind of on the edge of that. Let me put the laser point. So Reading's about here. I don't know where it is. So in the UK and many other countries, they tend to accompany their warnings with a risk matrix. So it basically conflates likelihood or you might say probability of a hazard with a potential impact of that hazard. And so today just makes it into the amber because they say it's medium likelihood and medium impact. So it's here. It's quite hard to get into the red category. But it does occur, of course. And so the question is how you use ensemble information to determine where you are on that matrix. And in fact, operational forecasters do that through a combination of quantitative and subjective means. So in this talk, I'm going to be using a couple of things as illustrations. So I'm going to talk first about Southeast Asia in particular because it's a bit of a frontier in prediction. It's predictive skill is much lower for that region than it is in middle attitudes. In a particular sense, there is long range prediction, as they say, within so and so on. But weather prediction is generally less skillful. So I'm going to talk about direct prediction of precipitation. So that's using precipitation output directly from an ensemble to make probabilistic projections. And then I'm going to talk about these hybrid forecasts where we take the predictable components and then use those with a statistical association with the high impact weather to do this kind of hybrid statistical dynamical forecasting. So I'll illustrate that first with equatorial waves, which are thought to be quite a predictable component of the tropical atmosphere. And then I'll talk about weather patterns. So the different series, of course, these waves propagate and they propagate in a predictable fashion, or a typical approach to weather regimes or weather patterns is to identify static patterns through statistical analysis of the data. So they're actually describing very different things. And so the approach here is going to be use large scale and synoptic scale weather patterns. But in both cases, then we see the regime position forecast using S2S ensembles. And then finally, if I have time, I'm not sure. I talk about how we obtain multiple scenarios from the ensembles when there's flow dependent predictability. But the illustration for that is for the EK. So this work in Southeast Asia is really driven by quite a large project with a mess office, or actually a collection of projects. So the mess office actually ran their NWP system downscaling in a kind of nested sense in a semi-operational way for six months. So this is a very expensive operation. So basically, from the global model, they had an 8.8 kilometer kind of large regional ensemble. These are all ensembles. So each member is nested in the global member and downwards. So there are 17 members. Then there is a 4.4 kilometer model domain, which is obviously quite a bit smaller, and then the 2.2 kilometer resolution models. And they're basically, they're in these three locations because they're centered on Kuala Lumpur, Jakarta and Manila, who were the partners in this. So first of all, you have to look at whether the model can actually stimulate the thing that you're interested in. So we were interested in precipitation. This is estimates of daily accumulations of precipitation, so climatology, where the three countries involved. And then this is day one into the forecast, so climatology of the forecast. So it looks reasonable at first sight, but obviously there are some pretty big deficiencies. Like for example, the model likes to put all the rain along these mountains on the west side of Sumatra and the observations, a lot of the rain is actually over the ocean near the coast. But part of that is that there's actually a chain of islands here just offshore, which are not represented in the model, even the high resolution model. So a lot of this rain is actually over these islands. So this is kind of one problem you have to deal with. You can see a similar thing going on in Java. You know the model likes to rain on the mountains essentially in the, in GM we see a lot of rain over the oceans. Although also there are a lot of deficiencies with the rainfall estimate from GPM as well. So, you know whether this is actually real is also quite hard to ascertain. So first of all, we have to think about how to evaluate these high resolution forecasts. So this is a typical approach using fraction scale score. So the forecast has, you know, the gray cells of where precipice occurring. So this might be like set of convection example. We would say it was successful if over the scale of interest, which is this neighborhood here, we've got just as many cells that are precipitating and as we see in some form of observations, so this might be radar or the GPM satellite estimate. So yeah, this is good because the fraction is the same, even though in the particular grid cell in the middle, this one is dry and that one's rain. So this just accepts that on the convective scale, we're never going to predict the location of individual updrafts. And then fraction scale score is just the pain from this comparison, but I won't go into the details of that. But the important thing about fraction scale score is that you vary the size of that neighborhood lane and then you get the measure of fraction scale score that basically increases, it asymptates to a high limit. So it's only asymptates to one if there's no bias in the forecast, otherwise you can predict what it asymptates to. And it starts at a fairly, well, I guess it really comes from zero on the grid scale. And then there's a kind of range of skillful and useful scales. And typically we define the smallest skillful scale being when the fraction scale score is a half. And that's what we use here. So actually then this is for these forecasts. So six months of forecast from Malaysia, Indonesia and the Philippines, and just focus on the red line. So the red line is where the fraction scale score equals a half. And then we've got spatial scale going along the x-axis and lead time going on the y-axis. So basically the skillful spatial scale increases and lead time indicating that we are less and less certain as the forecast go through. There's obviously a strong diurnal cycle which is very typical in this region associated with the convection. So basically skill is larger in the daytime actually from forecast from about local midday and much smaller at night. And that's partly because the precipitation is over the ocean during the night and the model is not very good at that bit. Philippines doesn't such a strong diurnal variation because all that's associated with in the tropical cycle. John, sorry to interrupt you. Okay, actually there is scale. I think your audio was breaking up, maybe for bandwidth three since you could switch off the video. Oh yeah, so I switched it off, yeah. Okay, thanks. Okay, is that better? Yeah, it's better. All right, thanks. So the scale also varies with location. So actually this is basically day one forecast because the gray shading above within the red lines is basically where there's most scale. So you can see there's more scale over the land actually and actually in this region of the sea there's less scale in other regions. So it's interesting to catch is that these are for longer lead times, very long as here don't quite make sense but this is basically the next day in the day. So that's attempting to forecast the high impact weather directly using the dynamical model. So now we're going to try and do it without using the precipitation from the model. So first of all, look at equatorial waves. So there, and when we show them from data that they're associated with high impact weather. In fact, with Kelvin waves for example, extreme rainfall. So you can actually look at higher percentiles in this. It is up to four times more likely when you're in the right phase of the Kelvin one. So if you can predict the Kelvin wave you're predicting the risk of precipitation that's the general idea. And they have very distinctive structures and they have distinctive phase speeds. And Falko Jutes and a number of other people have argued that potential predictability in the tropics is very high due to the presence of these waves. Although as I'll show later the actual predictive skill in current systems isn't that good. So there's a lot of potential predictability which is unrealized. So the basic method is that we take global analyses we identify wave amplitude and phase and then we find the statistic of association between high impact weather at every grid point or every neighborhood with wave amplitude and phase. So we get conditional probability in wave phase space. You could call it that. And then we take global NWP forecast. I've just shown you could use model precipitation directly although actually the precipitation from the global models is really poor. You need convection permitting models really. But anyway, from the global NWP you can forecast wave amplitude of phase. You combine that with the statistical information you get a forecast probability of that weather. So it's bypassing the use of the model precipitation. This is what the waves look like. So actually our identification method is a spatial projection method onto these structures. So these structures come from theory. There's a few parameters that fit but it's fairly straightforward. So we're particularly interested in Kelvin waves here and these westward mixed Rossby gravity waves. So these have meridional wind at the equator. Kelvin waves have zonal wind at the equator. So here's an example of that statistical relationship. So the color field is the GPM actually is trim. It's a bit older. It's trim precipitation but the satellite is the precipitation. And the consulate is where there's convergence in the Kelvin wave. So when there's the Kelvin wave has land soon here the precipitation is more or less with the convergence not very surprising. But you can see there's a lot of structure in where the precipitation is. And so the statistical association knows that. And then here's another example for a westward mixed Rossby gravity wave. So here the centers of action are off the equator. Although the meridional wind is strongest on the equator. And you can see there are heavier precipitation in these solid planters and there's reduced precipitation or suppressed precipitation on the other side of the equator. So these are the kind of signals that we're looking for. And these propagate. So this one is propagating westwards. The Kelvin wave is propagating eastwards. So that's the predictable signal. So then how do we use that? So this is then, sorry, each of these this is Western Indonesia, Southern Indonesia Northern Indonesia and Eastern Indonesia. So we just split it into sectors and then the wave amplitude is given by the radius into these diagrams. So we just split that into three bins and then phases split into four phases. That's partly because we've only got the waves move quite fast and we've only got six hourly data. So four phases is about the best that we can do. And as you might expect, a fraction of time in any one phase at a given location is roughly equal probability because the same wave is propagating through. So it just depends on the phase of the wave. But when you calculate the conditional probability of heavy precipitation given wave phase of amplitude you can see that our probability is much, much higher when we've got high amplitude and we've got the convergence phase of the Kelvin wave in all of these locations. So this is the degree to which this is much larger than all the other boxes tells you how much potential predictability you have. If this was a flat diagram you'd have no potential predictability you might as well just give up there anything. This indicates that we've got a fighting chance and actually it's quite large. So high impact weather defined by this particular percentile threshold is the probability of high impact weather where and where in this region is 30%. I mean, that's really quite a high probability. So can we predict the waves? And because that's the next question, yes, for some extent. So just focus on the dotted lines here. This is the skill in the Kelvin waves versus lead time. I don't have time to explain all the other curves but basically from the analysis it drops off roughly linearly lead time skills getting a bit marginal by day six. The other waves like the Rossby one here has a bit more skill out to longer lead times. So this is just the state of the art. This is where we are currently and it's pretty similar in other systems. This is in the Met Office ensemble. So now we're gonna jump to other predictable components over the region. Because although our equatorial waves have a lot of potential predictability the models are struggling to forecast them beyond the week. So then if you want to look at lead times beyond the week, I think you have to look for something else. So let's think about middle attitude predictability. So this is actually a diagram produced by Lara Ferranti who spoke last week of course. This is a middle attitude example where as lead time increases we're looking at the projection of the forecast or the split of the ensemble between different Euro-Atlantic patterns of variability. So the red one here is blocking. So we know pretty well that these really large scale patterns have skill in the S2S range. Whereas weather types or things like gross weather lagoon or the very old land weather types in the UK they don't have much more skill than individual cyclones. So there is skill in the five to 10 day range. Individual cyclones maybe three to seven days if you actually want to forecast the amplitude and position of the cyclone. Frontal features are actually quite predictable given that they're so narrow because they're completely slaved to the cyclones. So in somewhere like the UK there's very high predictability for fronts. But the actual high impact weather phenomena there's very low predictability, typically less than two days. This is where the high resolution forecast come in. So we're going to use this thinking to think about Southeast Asia which hasn't been explored in this way before. So I'm going to take a two-tier approach where we define a really large domain. In fact, this is a fifth of the Earth's surface and we look for patterns of variability using the data, basically using a K-means clustering approach. In fact, we precondition the data using EOFs just to truncate the data set and then we do K-means clustering which is a usual approach. And then we take a secondary domain within that which has just spanned Southeast Asia which is actually a quarter of the area. So these kind of ratios are deliberate. So we're trying to separate skills artificially in a clustering approach. Then in the second tier, we limit the data that's going into the clustering to only those days that are in one of the tier one regions. So the tier two is deliberately a subset of tier one. So it's like looking for the weather regions that are important when we're in regime of tier one. And so this is the diagram kind of indicating that. So we've got eight tier one regimes and then we have all these tier two regimes as subsets in phase. And I'll show you what they look like a bit lazy. And we did compare it with a flat approach where we just use the same number of regimes but adjusted the clustering over Southeast Asia straight away rather than the conditional approach. But I won't spend much time describing that here. I should say the cluster variable is wind. It's not precipitation. It's 850 hexapascal wind. That's because it's more predictable variable. So let's look at the tier one clusters first. So this is day of the year. And then this is all the years in the data set from 1979 to 2017. And then the color is just the cluster assignment. So you can see to start with that the tier one clusters primarily pick out something that's related to the seasonal cycle. And that's really obvious. Perhaps what's obvious is that if you then reorder the years by phase of ENSO there's actually quite a strong ENSO signal in there. So in boreal winter, the cluster two is related more to La Nina and cluster one is related to my ENSO conditions. So you can see that in the transition from orange to blue. And also that some of them which are related to the summer monsoon Asian summer monsoon there's clearly dependence on ENSO in the timing of the regime. So basically it's a lot later in our linear than it is in La Nina. So the cluster is picked out this information. We didn't impose any seasonality. We just gave it the full data, see what happens. And if you look at the regimes, they make a lot of sense. So regime one and two are both North East monsoon in DJFM. It's just this one is more active in terms of precipitation than that one. So sorry, the color fill is the precipitation. Whereas regimes four and five are summer monsoon and raging flavors are more intense summer monsoon which is also extended a lot further East than in regime four. And then there are transition season regimes. So then we take that into the next year and there are 51 of these. So I can't show you them all, just pick out a few which are quite interesting. So some of these are related to what are called cold stages in this region. So Grand Central 2B here has a really strong cross-sectoral flow between Borneo and Malay, well Borneo and St. Marcha. This is called a cross-cagrile surge associated with heavy precipitation in Java frequently. So this is a kind of large synoptic scale regime. This one here is related to the floral summer interseasonal oscillation. So actually quite similar to Nina's talk where she was talking about the monsoon interseasonal oscillations is a kind of variation on that. And that's coming out of one of these regimes where some of them are clearly related to synoptic scale events. So these are just tropical cyclones being in preferred regions near the Philippines and they're coming out of the clusters as well. So now the question is whether there's any skill in forecasting those things. So we've taken glossy five which is the Met Office S2S forecast. We've taken a whole load of pinecasts. So it's a fairly recent version of the S2S system for the hindcasts of the cover earlier period. And we just take all the hindcasts that there are. And then we assign each ensemble member to regimes based on the era five centuries. So this is another decision you have to make. You could do all the clustering in the model world but we decided not to do that. We decided to project the forecast into the clusters identified from data. And that's a fairly typical approach. So one thing that we learned straight away is obviously there are lead time dependent biases. So this is the tier one. Well, in fact, it's all of them. You can see all of the clusters here but the colors are the tier one. So you can see that there's a lead time dependent bias lead times running out here to day 36 between these two regimes. So basically in the data, this one's occupied a lot less than the red one but towards the end of the forecast it's the other way around. And this is actually because in the model the Asian someone once soon extends from too far east in the model that's just the systematic bias. And you can see it's lead time dependent. But actually the other ones aren't too bad in their representation. There's perhaps some endso dependency of between the red and the blue or orange and blue. So now we're going to use this information to do the raging conditioned precipitation forecast. So the direct precipitation forecast would just use the output of precipitation from the Gossi five model to forecast the probability of precipitation location X exceeding the 90th percentile. But instead we're going to take the probability of being in a regime and which is forecast by Gossi five and then the data relationship between the probability of precipitation exceeding the 90th percentile conditional on being in the regime. So it's a conditional probability. And then we combine that with the forecast probability of the regime to get the conditional forecast probability of precipitation exceeding the percentile. So it's a typical kind of combination of probabilities problem. This is quite remarkable. I still can't believe the results and we've been checking these quite a lot. So this is now the blue one is the tiered cluster approach. And this is prior skill score in the exceedance of 90th percentile precipitation. So there's a scale on the face of it right out to beyond 20 days. And the flat one is less skillful which is quite interesting. So it's important to separate the scales and the clustering. But then we try aggregation of precipitation up to more grid points. You know, the S to S models have just got these one and a half degree data and well at least in the S to S database we aggregate up to more grid points. It doesn't change the range of skill that much although it changes the value of the skill score but this is the straightforward use of the model precipitation. Basically even with aggregation up to 500 kilometers the skill dives to zero in two days. So somehow we've gone from two days to 20 days by doing ranging condition forecast which is amazing really. I mean, we don't get that gain in the latitudes nothing like it but in the tropics it's really important. So here's a case study just to bring that to life. So this is the, it's in a particular week actually only in December 2019. This was the exceedance about 25 millimeters per day daily accumulation from the observations over that week. So this is the kind of thing we're trying to forecast. And then these are the forecast from the ensemble and they were kind of day zero to six which we're calling week zero here and then week one, two and three. So you can see there's actually quite a strong resemblance to the observed exceedance and this is where the skills coming from. So for example, the south or western end of Borneo there's obviously probability of heavy precipitation there and over Java. Right, so now we're going to change tack in the remaining 10 minutes. John, maybe like if you caught up in three, four minutes so we have time for questions. Yeah, yeah, sorry, yeah, that's fine. Yeah, got confused on my time though. This is very short this bit anyway. So this is now focusing on the UK example. So again, it's applying a custering technique but in a different way. We want to identify whether in an ensemble forecast there are several plausible options. So there are 18 members in the basic ensemble here we want to reduce that to two or three important scenarios and they need to be temporarily consistent. So what we do is where I mentioned fronts are important for the UK. So we get a frontal diagnostic using the gradient of wet hole potential temperature and then we look at the position of features in this diagnostic. So in fact, we threshold that field just to get feature features and then we calculate the distance between though distance between ensemble members using fraction skill score. So basically you can relate distance for the neighborhood where the fraction skill score is large that kind of indicates the scale you have to go to for match. So this is the kind of thing we've got a whole ensemble here and the ones with the color frames that as the match are the ones that are falling into the same cluster. And let's look at this one as lead time increases the ensemble spreading increases but we're looking for the point where the sum of distances between all pairs of members suddenly reduces which indicates that there's clustering behavior. So we kind of start a time window when clustering behavior occurs and then look at what happens. This top picture just shows the associated between members and each cluster. So cluster zero is typically got the most members in it and we're interested in this time window here. So members do skip around between clusters which is difficult but we try and look for the ones that are coherent through time. And what was kind of fascinating by this here we've got start time versus lead time. So things on the diagonal are a common validation time or so in fact on this ensemble spread the top picture there are particular events in the flow where ensemble spread is large no matter how early you start your forecast. So longer forecast would have a rapid spread short forecast have a rapid spread. And we find that clustering so the dark blues here is where there's strong clustering that's just related to where the ensemble spread increases rapidly which is what you might expect. But what's interesting that means that there are coherent things going on in the flow where there really are clustering of ensembles which are not dependent on the forecasting system or the lead time at which you start these are properties of the atmosphere. And here's a movie showing this in this case four clusters work best and the different colours are different ensemble members but just overlaying if they call them the same cluster. This is what we're trying to achieve and at the moment we're trying to investigate whether these are indeed associated with different high impact weather scenarios. And this is an advantage of this approach you can look at any field because each of these members is actually just one member of the ensemble. So you have every variable that you have an ensemble forecast. So I'll just summarise there. So, you know, obviously there is skill in direct ensemble prediction of variables but it can be quite limited for some variables like precipitation, particularly if convection is dominant and even with a convection permitting model that's true because the convection itself is inherently less predictable. So it is useful for long range to use these hybrid methods and there's been lots of talk about similar things over the last two weeks. So here we tried it conditional on equatorial waves and also on these kind of clustering approaches where two tiers indicate different scales. And it's, you know, it's amazing how much further you can get them predictive skill even without time aggregation on the lead time. And then, then, you know, given when we have skill ensemble forecasts we have to think about how we're going to use them. And so I think do need to think about different scenarios that could occur in longer range forecast because ultimately that risk matrix that I presented right at the beginning needs to be different for each potential outcome. You can't just present one risk matrix for one ensemble when there are actually two distinct scenarios. So that's where I end there and then quite the questions. Thank you. Thanks, John. Thanks for a great lecture.