 Welcome to today's lectures to the ASP summer colloquium on S2S science. Today we're going to talk about the Madden-Junior oscillation and teleconnections and predictability coming from the Madden-Junior oscillation on the S2S timescale. And our first speaker is Hemi Kim. Hemi is a professor at Stony Brook University. Her interests lie in low frequency climate variability and predictability, topical-extrtopical interactions, S2S prediction, and extreme events such as atmospheric reverse tropical cyclones and tropical cyclones. And she has also dabbled in machine learning. So Hemi, we're looking very forward to your presentation. You're on mute, Hemi. Okay, thank you very much. Let me turn off my video for better connection. Okay, today I will talk about the predictability and prediction of the Madden-Junior oscillation and just to remind you what we have learned so far, because last week at the beginning, first few days, we learned a lot about MJO and then, so I want to just remind you. So we are talking about this S2S timescale, which is too long for the atmospheric initial condition to keep the memory, but too short for the boundary condition to take an effect. So, but this S2S timescale, the forecast scale has been improved a lot in the recent decade. And one of the source of the predictability is the Madden-Junior oscillation. And the Madden-Junior oscillation is not just in the tropics, but the diabetic heating and the related circulation, it has an impact on the globe. So it has a lot of global weather and climate phenomena modulated by the MJO. And also we learned about how the ALC interaction is important for the MJO and also the stratosphere and troposphere coupling is also important for MJO. So I'm sure that you are now convinced that MJO is an important source for the sub-seasonal predictability. And this talk, we will talk about how well do the modders predict the Madden-Junior oscillation. So first of all, for MJO prediction, we need a some good index that represents the Madden-Junior oscillation, such as a lino3.4 index. And here it shows the 3G structure of the MJO. Here you can see the convection and the low-level 850 millibar zonal wind, low-level convergence and upper-level divergence. So the MJO convection is tightly coupled to the circulation. And to define MJO in the real-time forecast, we need some index. And that index is used with this ORR and the low-level wind, the 850 zonal wind and 200 millibar zonal wind. So using these three indices, three variables, we can have an index. This is called the real-time multivariate MJO index. You have heard about the RMM index, and that was developed by Will and Handel in 2004. So what it is, is basically you have ORR and the wind field, upper-level, low-level wind, and then you do several preprocessing and then average it over the tropics 15 degrees south and north and do the EOF. So this is the eigenvector of the eigenvector first and second mode. And this shows the PC time series. So this PC time series is the RMM 1 and 2. And you can see that over the time, there is some time when the RMM index is large. So that is when the MJO amplitude, we say it's big. And then sometimes it is near zero, the anomalous near zero. So we can assume that the MJO is weak. And we also can put that in this phase diagram that you have seen before. And this you can see using this eigenvector and the RMM index, we can define the MJO location into eight phases. So it's from one to eight. And you see phase two and three is when it is in the Indian Ocean. And using that, we also can make a composite. So it's called a life cycle composite. And then you see this shading is the ORR and the wind field, low-level wind field. So we can define eight phases. And you see that, for example, when MJO is in phase two or three, they are in the Indian Ocean. And then we have the low-level easterly to the east. So, and then how can we use this index for monitoring and forecast? So here, this shows an example of a very active MJO event that occurred in April 25, 2019. And that was defined as phase three. So you can see the large convective area here in the Indian Ocean. And here, if you go to the CPC website, they provide the real-time forecast updated almost every day. And then you see this is the observation. So here, the red line is the observation. And here is the April 25. So you see this here. It is in phase three, which corresponds to this pattern. And then they provide the forecast. So this is NCEP GFS forecast. And it provides ensemble and also ensemble mean. So the talk today is about how well do the models predict the MJO, and especially the RMM indices. Okay, so this shows the RMM index prediction scale. So when we have the hindcast, we can compare the RMM index with the observation. And this shows the scale of the dashed lines, the S2S models, and the solid lines are the sub-X models. So this is a combination of a few papers. And you can see that the correlation coefficient, if you take the 0.5 as a reference, then you see the black line is the ECMWF model, which is about 32 days. And then except this one, most of the models are arranged between three to four weeks and maximum approximately four weeks. So that is the current, let's say, current MJO forecast capability using this RMM index. Okay, so the MJO prediction has not that long history, actually. And that is because we did not have the computational resources to do a lot of hindcast experiment. So here this shows the lead time in weeks. And this is the predictability. So predictability, MJO predictability is the hypothetical prediction scale, assuming that the numerical model is perfect. And studies have shown that if the model is perfect, then we can predict the MJO up to seven weeks. In the early days, we use, when we do not have the numerical, good numerical models, people use the statistical models based on linear regression, or many, based on linear models. And the prediction scale was about three, maximum three weeks, it was between two and three weeks. And with numerical models before 2008, so here the years represent the publication year. So the studies that published before 2008, they showed that with the numerical models, MJO can be predicted about two weeks. And then in the recent years, we show that on average, they can predict the MJO for approximately four weeks. And I'm not going to cover everything here, but if you go to this, if you read this review paper, there are more details. And I may only cover about 20% of this paper. So here this shows the continued one example of how the MJO has been improved. And this shows the ECMWF model. Here the red line is the point six correlate when we take the correlation point six. And every year, the model is upgraded. And then you see this continuous improvement of the MJO RMM prediction scale. And especially here you see between this 2006 and 2008, there was a big jump. And that is when the model changes its convective scheme in a way that the convection is more sensitive to the environmental moisture. So you see this improvement. And today what I will talk is about the consensus and key issues in terms of MJO prediction. So just to summarize the consensus among many studies is that prediction scale is higher with strong MJO in the initial condition. So when there is a stronger signal in the initial condition, then we can predict the MJO better. And skill is higher in boreal winter. And that's because we have more well-organized MJO, stronger MJO. And skill is sensitive to initial MJO phase. So when the MJO is initialized in Indian Ocean versus Western Pacific, the skill is different. There are and etc. And there are many issues. And one thing that I want to emphasize is the quick decay of the propagation when you cross the maritime continent. So first of all, let's talk, let's see how the ensemble plays a role in the MJO prediction. So here this shows the S2S models. And the yellow is the control run. So this is the focus lead time when RMM skill reach 0.6. So here you see all the control values, control simulations are less than the ensemble mean. And in some models, like the ECMWRF, CNRM, or BOM, you can see a big jump, a big improvement when the ensemble mean is taken. And these are the models which has large ensembles. So if you have more ensembles, then usually they have a better skill. But it's not always the case. The HMCR has also large ensemble, but you can see the decrease of the skill. Another consensus is that the model has, the ensemble system is under dispersive. You have heard a lot about this and this shows the, from Nina Eddard, this shows the solid lines at ensemble spread and the dashed line at the ensemble mean root mean square error. So you can see this ensemble spread is smaller than the root mean square error. And these were the earlier versions of, this was the first hind intracesional hind cast simulation, coordinated hind cast simulation launched in 2009. And even in the current S2S and sub-X models, they are all under dispersive. Another thing that I want to emphasize is the QPU-MGA connection that was covered by Yaga Richter last week. And this is because before this QPU-MGA connection, before we find this connection, there was not much studies about how the large scale climate can modulate the MGA prediction skill. But then a few papers show that MGAO is more active in the E-Cubio winter. So here, this is the difference of the MGAO activity between E-Cubio and the climatology. So you can see this, the Easterly-Cubio winter, they are much stronger than the Westerly-Cubio. And in a recent paper by Jane Martin, you can see that we have a review paper on this QPU-MGAO connection. So we can find more details in this paper. So now, if the QPU modulates the MGAO, then the prediction skill will be different as well. So here, this shows the paper by Shukang Wang. Here is the MGAO prediction skill in the S2S models. And he used the OMI index. OMI is another index. It's the ORL-based MGAO index. So it does not include the wind field, but it's only ORL-based. So it's only convection. And then you see that in all models, the blue is bigger than the red. So that means in all models, the Easterly-Cubio have better MGAO prediction skill than the Westerly-Cubio winter. But then one thing to notice is that even the low-tub models. So here, the green box that represents these models here, which is classified as low-tub models. And still, you can see that there is a big difference in the MGAO prediction skill between Cubio phases. Even the model does not have any Cubio or stratospheric variability. So in our current study, we showed that there is a difference in prediction skill, which agrees, which all recent studies show. But maybe the prediction skill difference is not statistically significant. And here are some nice editor highlights by Xidong Zhang. So emerging controversy in Madden-Julian oscillation prediction. So here, this covers our recent paper. And that is, for example, here it shows the sub-X and S2S models. And this is the MGAO prediction skill in E-Cubio and Western Cubio. And the thick lines are where the difference of MGAO skill is statistically significant. And when you have more samples, there are some more significant times when we have more significant difference. But then most of the models do not have that much, that enough samples to get this statistically significant difference of a Cubio. So that is one thing that is that I want to highlight. And then another key issue in terms of MGAO prediction is the MGAO propagation. So here, let me show you this one first. So here what is shown is the amplitude bias. So Y axis is the initial phase. And this is ECMWF model, which is in the S2S database. And here the brown shading is the weaker. So that means when the MGAO prediction starts in all your phases, when it is in the Indian Ocean or over the Maritime continent, then it loses its amplitude very fast compared to other phases. And that is actually in most of the models. You can see in NCAR CSM1, you also lose the amplitude fast. And then in the later phases, it's getting stronger. But in overall, in all models, you can see that in all your phases, when the MGAO is over the Indian Ocean, then it loses its amplitude faster. So there is a quick decay of MGAO signal when the MGAO starts in the Indian Ocean and propagates through the Maritime continent. And this shows the paper by Ritard et al. And it shows the percentage of MGAO events not crossing the Maritime continent. So the MGAO is talking from Indian Ocean and not crossing the Maritime continent. You see in ERA interim, they are only based on their metrics. It's only 10% that do not cross the Maritime continent. But then in the S2S models, you can see that they are almost twice as large as the observation. So that means there are more events that do not cross the Maritime continent. And we call this as MGAO Maritime continent prediction propagation barrier. So if you put that in this phase diagram, so here you see the composite of multimodal mean. So this is a multimodal average of eight subjects and S2S models. And here the black line is the observation. So composite observation and the blue is the multimodal mean and the open circles are the individual models and corresponding observation. So for example, can see that for example in phase two, the model cannot, this is a three-day average running mean. So you can see that the model already have a big discrepancy between the initial, like in the first three days. So they're already weakened. And also in the phase and then as it propagates, the amplitude gets weaker. And then in some phases in the later phases, they are actually a little bit stronger. And then in some phases, there is an amplitude error and also phase error. And especially the Indian Ocean, these ones, we try to understand why they lose their amplitude so fast. So this shows the first day forecast after initial MJO phase two. So here this shows the NOAA and the ERA interim window normalize. And here is the observation. I just called this as an observation. And then you have the MJO and associated low-level easterly wind to the east of the convection. And then this is the ECMWF, the first day forecast. And you already see that the convective anomaly has weaker amplitude than the observation. And then if you put that in the Hofmuller diagram, you see this is Y axis is forecast day. And here you see in observation, the shading is ORL and the contour is the U850 anomaly, taking the average between 15 degrees south and north. And see this in the observation, it propagates through the maritime contour, which is approximately 120 east. And then what in the ECMWF, it does not. And the reason why I showed the ECMWF model is in the paper, we have all models, but this is because it's the best model. So we try to understand how does the model mean state or the mean bias impact the MJO propagation processes. And this is from Monday's Chidong's lecture. And here is the, here, there are many theories, as you have seen before, but we try to understand this maritime continent barrier using the motion mode theory. And long story short, if you just focus on the eastward propagation, it is that the largest scale horizontal and vertical motion advection by the circulation that moistens the trapper sphere and then makes the convection center to move. So here you have to the east, we have this advection and more moisture so that MJO moves to the east. And we show this is the observation, for example, here is the winter mean moisture Q850. And there's a specific humidity at 850. And the reason why I'm showing only one level is because the sub-X models only provide the Q850. So I have to match this. And here, this is the mean state. So here, there are many other terms, but if you just focus on this moisture advection, this is the V prime, that gradient Q bar. So Q bar is the mean moisture and the gradient of mean moisture. And we have the MJO wind, right? So here on the bottom, it shows the day one MJO in the observation. So let's say we have MJO active, MJO active convective convection in the Indian Ocean and related cabin wave response. And then the suppressed MJO induced the rugby wave response. And then there will be a moisture advection because the mean moisture has its maximum near here. So the gradient will make this the moisture advection with the wind. So and then if you compare the observation, the moisture advection, you can see that moisture advection is large in this area. And then after 10 days in observation, they propagate through the maritime continent, MJO propagates through the maritime continent where the moisture advection is large. And of course, there are also other processes, but here I will just focus on this one. And then the question is, then if the modus, the sub-ex and S2S modus, do the modus capture this moisture advection? And maybe not because they do not propagate very well. So here, the bottom is the same with the previous one. And then here this shows the moisture advection term average over this region. So here in observation, in the first 10 days, there's a positive moisture advection. But in the multimodal means, sub-ex and S2S modus do not have this strong moisture advection. So they get much weaker, very fast. And these are the individual modus. In observation, the moisture advection is strong, but the individual modus or modus have weaker advection. So if you think about this term, the MJO wind can also impact this advection term, but studies have shown that the mean moisture distribution is important for the MJO propagation that you showed last week from Charlotte DeMote's presentation. And here this shows, again, the observed Q850. And this is the bias. It's this multimodal mean minus observation. So what you see here is that all modus actually have this very dry bias in the lower troposphere. So this is 850. And then you see this dry bias. So the modus are drier. And if it is dry and more dry near the equator, then it will weaken the gradient of the moisture. So there will be less weaker moisture advection because the gradient is weaker. And then the more ECMWF model provides all level of the specific humidity. So we can make this plot here. And what you see here is the Q bias as a function of lead time. So what you see here is at the beginning, there's already dry atmosphere. And then it continues to get dry and continues to amplify. So yeah, so there is a bias in the mean state. And that may impact this propagation and prediction scale. Of course, there are many other theories and that can be applied. But we use one theory. Another approach is so here recently, we did some applied some deep learning for MGA bias correction. So the hypothesis is, you can see here that the bias, if the bias is systematic in a specific MGA phase, for example, in Indian Ocean, they always get weaker systematically. And in the Western Pacific, they are always stronger. Then it can be corrected by optimal bias correction method. So it hasn't done before. And what we use is to correct this, the bias, we use the LSTM. So this is the long short term memory. There are many different approaches in the deep learning, but we use the very simple one. And here, the input is the SGS model RMM 1 and 2, and output is the observed RMM 1 and 2 during the training period. And then in the real forecast period, we can put the SGS models, and then get the corrected RMM index. So this model is built separately for each phase, modern forecast lead time. And we did the real time forecast and also leave one near our cross validation. And this is the result. So here, again, the black and blue is the blue is S2S, black is the observation, and then the red is after applying the deep learning method for the bias correction. So what is interesting is that the first day is closer to the observation, and then the following predictions is also closer to the observation. So even in the later phases, like phase six or seven, they get closer. The stronger amplitude gets more realistic, and the bias, the error is reduced. And then the NGO propagation is this is observation, this is ECMWF model. We see that the weak amplitude after 14 days, but then after making this deep learning correction, it is almost, we can see this propagation. And that is where the anomalies are significant. And this is the focus error using the deep learning method. So here, this shows the subjects and S2S models. So first of all, this is so we can divide the focus error into amplitude and phase error using some metrics. And then here it shows the multimodal mean. This is four weeks average. So here, the blue shadings are S2S, and then the red or pink shading is after deep learning. And the light shading is the amplitude error, the dark shading is the phase error. And you can see that it's on average about 84% reduced. So let me summarize. Yeah, so we have seen a lot of progress in the last decade. And we still have about three weeks room for improvement. And I like to share some thought in terms of the in terms of the current understanding of MGO prediction. So challenge it and question. So first of all, is the MGO predictable up to four weeks, really? I mean, here it shows the four weeks, but keep in mind that we use the RMM index, which is highly filtered. You take the latitudinal average. And then in this index, we use three, three different variables, but mainly the upper level with the U200 takes the dominant role. So if we use the OMI index, where only index or maybe SUE is a large-scale precipitation tracking method, maybe we don't, we may not see this good prediction skill. Or even the, so these are just the index, right, the RMM-OMIR index, the time series. And can the model actually capture the spatial pattern? And here it shows the ORR pattern correlation skill over the Indo-Pacific. And here the blue, this is the correlation coefficient. And the blue is the ECMWF model. So if you, if we compare the ORR pattern correlation to capture the ORR pattern related with MGO, it only goes up to six days. So it's not that good. So keep in mind that the RMM index is a highly filtered data. And also it has the more impact from the upper level win. So it depends on what you defined as MGO. If you, if you think the convection is the MGO, then the RMM index may provide too optimistic a view on the MGO prediction skill. Another thing that is to how to improve the MGO propagation and the related 3D structure. There have been some studies with changing model physics resolution initialization, et cetera. But as you know, they are mainly based on the case studies because to do this, the sensitivity test of these changes, then you need a huge computational cost to do a long-term hindcass simulation. And also do model simulate diverse MGO types. Again, the presentation by Shui Chen last week, there are diverse MGO types. For example, here, this is a paper by Bin Wang and it shows different four types of MGO. This is standing MGO, jumping MGO, slow MGO, fast MGO. So do the model simulate all these types and also the related physical mechanism. The last one I want to ask is to model capture the largest climate modulation on MGO. For example, not only the Cubio, but maybe also the Indian Ocean dipole may have influence on the MGO. But again, to understand the largest-scale climate modulation, we need even more longer set of the high cast, right? Only like 15 years is not enough to separate it into a linear phases. So that's and I'm happy to take any question. Thank you. Thanks so much, Hemi. It was a very comprehensive overview of MGO.