 will be Hemi Kim from SUNY in Stony Brook. Hemi, thanks again for your talk during the colloquium as well whenever you're ready. Sure. Can you see my screen? Yeah. Okay. Thanks very much for the invitation and giving me the chance to present my current work. So let me turn off the video for smoother connection. So here in this talk, I will talk about the maritime continent prediction barrier. And I'm glad that I chose the same animation that Eric showed. So this shows the nice MGO related precipitation. So you see that when we separate it into the MGO into eight phases, in the all your phases from one to one to three, it represents the MGO propagation. It represents the MGO in the Indian Ocean. And then the observation propagates through the maritime continent. But actually sometimes, and like many of the MGO events also have this is seized by the maritime continent. And there is a lot of interaction, diurnal cycle, lengthy contrast, orography impact, and so on. That makes the MGO not propagate through the maritime continent in observation. And this maritime continent barrier is exaggerated in the prediction, S2S prediction model. So today my talk is about the maritime continent prediction barrier, which means the exaggerated MC barrier in the S2S models. So just to make sure we are on the same page, this is the MGO prediction skill using the RMM, the real time multivariate MGO index in the S2S and sub X models is the correlation coefficient as a function of focus lead time. And here you can see that the best model is ECMWF model, which is about taking the 0.5 as a stress word. It's about 32 days. And the rest of models are mainly distributed between three to four weeks of the lead time. And similar. So if you remember Frederick's talk in this morning, he showed the amplitude bias and most of the models have weaker amplitude. And here this shows the actually the same plot, but separating it into the initial MGO phases. So Y axis shows the initial phase and X axis is the focus lead time. So if you just focus on the ECMWF model, it shows that the brown shading is weaker amplitude. So you can see that when the MGO is all your phase from phase one, two, three, four, when the MGO starts, when MGO is initially in the Indian Ocean, then it reduce it gets weaker faster than the other phases. So there's a weaker amplitude that the all your phases, though that means there is a quick decay of MGO signal when MGO starts in the Indian Ocean and propagates through the maritime continent. And this is quite consistent among all models like here and Cassius and one has also the same issue. And most of models have this quick decay of MGO signal. And here's another view of this quick decay signal in the all your MGO phases. So here this shows the phase diagram, which represents the amplitude and the phase of the MGO. Here the black line is the observation and the blue line is the multimodal ensemble mean of eight sub-x and S2S models. So what you see here is that in all your phases when the MGO starts in initial phase one, two, three, then you can see at the beginning it's already the amplitude of the MGO forecast. The initial amplitude is already decayed, which is consistent with Frederick Schott. So the talk today is about this maritime continent prediction barrier. The first I will briefly introduce the impact of the mean state and the mean state bias. And then second is to use some deep learning technique to make this MGO systematic bias correction. And then the third is the loss of prediction scale by the estimate the loss of prediction scale by the maritime continent using the NCAR CSM2 aqua planet simulation. So first let's start with the mean state bias. So here this shows the propagation of the MGO in the observation and the ECMWF model. So shading is the oral anomaly. And when it starts from the Indian Ocean, Y axis is focused lead time. So when it starts in the Indian Ocean in the observation it propagates nicely through the maritime continent and reach the data line. In the model, not only is it in the graph about all this stress and sub-X models that are analyzed, they show the damping of the MGO signal. So we try to understand how the mean state bias impacts this weak MGO propagation. And here this shows the observed winter mean moisture distribution. So it's Q850, the specific humidity 850 at the PASCAR. And the reason why I choose this specific level is because the sub-X models only provide the 850 at the PASCAR moisture. So here this shows the ERA interim, the distribution. And here we try to understand this mean bias with the MGO propagation processes using the moisture mode framework, which she don't explain two weeks ago in the colloquium. So I'm not going into the detail, but the key point is that for the MGO propagation, this term, the horizontal moisture advection term is important for MGO propagation. So here, for example, in the initial condition, like say in day one in observation when we have MGO in the Indian ocean and the suppressed phase in the western Pacific, this active MGO convective or normally induced the Kelvin wave response and the suppressed MGO reduced the grossly wave response. And this together, which is the V prime, the MGO related wind, would vex this mean moisture in between these two. And then there's a moisture advection. We calculate this from ERA interim, then you see that there's a moisture advection. And this makes after 10 days, the MGO propagates where the moisture tendency is positive. So then we can compare this moisture advection term in the S2S models. And here this shows the moisture advection term, this term as a function of focused lead time. So you see over this area, the boxed area, and you see in observation, it decreases as MGO propagates through the western Pacific. But in the multimodal, you can see the moisture advection is much weaker than the observed. And then we try to understand how this mean moisture distribution impacts this advection and propagation. And here this shows, again, the observed ERA interim Q850. And the lower panel shows the biases in the eight models. Here this is the brown shading is trial means the dry bias in dry bias. So it shows that in the all models, it has in the paper, we have the distribution from all models. And it shows that all estres and sublux models have dry bias in the lower troposphere. And then the dry by mean bias can impact this, the gradient gets weaker. And then it can impact this horizontal moisture advection. And because of the weaker moisture advection and the weaker the gradient of the moisture that can impact the MGO propagation scale. Here this shows the bias in ECMWF model and the stress models provide the true for the whole level. And then you see that this is the bias as a function of focused lead time over the Indo-Pacific region. And you see that this dry bias, of course, in the first few days, and then it gets amplified. So and we discussed this in the paper that that may be related with the entrainment rate. So in the paper, we showed the convection, of course, too frequently in the model. And it's not sufficiently inhibited when the tropospec moisture is low, and which is likely related to the representation of the entrainment. The more details can be found in this paper. And then so for to understand the maritime content prediction barrier, we try to understand to link it with the mean state bias. And then if so we have the mean state bias, and also we may have some systematic MGO bias. So here, for example, MGO amplitude gets systematically weaker in the specific phases of MGO. And here we try to correct this to make a bias correction using some deep learning techniques. So I'm not going to explain too much detail, but this is the some deep learning method, which is the long short term memory cell. And here, what we do here is we put the S2S models RMM1 and 2 index as an input, and then the output, it will be the observed RMM1 and 2. And during the training period, we train this model. And then in the real forecast period, we input the model forecasted RMM index and get the corrected RMM1 and 2. And we build the model separately for each MGO phases because each MGO has have different systematic biases and differently by separately by phase, model and forecast lead time. So in here just shows the results. So this here again, the black is the observed and blue is the S2S. And the red line is the using the deep learning bias correction method. So here you can see that at the beginning, this amplitude is nicely corrected. And then the following days, the following forecast are also at least better than better than the S2S forecast. And here on the right hand side, shows the MGO propagation. This is the overall normally in window normally here. This shows the ECM.ref model, which has quick damping of the MGO propagation signal. And then after using the deep learning method, the corrected RMM indexes, we can get some more closer, better MGO propagation. And then we calculate the MGO forecast error. This is the four weeks average of forecast error. And here it shows the each model that we use. So here the blue shade, blue bar is the S2S model and the red bar is the after we apply the deep learning. So this is a forecast error. The light color is the amplitude error and dark color is the phase error. And for the multimodal mean, on average, it is about the error is about 80% reduced. And we also compare this bias correction with linear regression model, but the deep learning model has better performance. And because the deep learning model has nonlinear activation function, so that allows the nonlinearity and allows the model to be complicated. So we use the bias correction model using the deep learning. And then the last part is the some new study, which is in collaboration with Jim Benedict. And here we try to understand how much prediction skill is lost by the maritime continent. So here we use the NCAR CSM2 aqua planet. On the right hand side, this is the SST distribution. So aqua planet is fully covered by water only. So there's no land or no sea ice. And it's only using this prescribed SST. So first control simulation will be the one pool SST, which is the zonal asymmetric SST. And then we did the MC barrier experiment. So this is the SST. But here, if you see the maritime continent, the SST, here over the maritime continent only we use the SST decrease with elevation using the lapse rate. So that acts as an aqua mountain. So here, this is the difference of SST between MC and one pool. So you can see in the maritime continent barrier experiment, it's much colder. And then the right hand side is the precipitable water. So you can see the dry it's more drier. So that and there is no, it's very the convection is suppressed because of the cold water. And here this shows the total precipitation as a function. This is average of over five day south and north. And it's an arbitrary just from one 500 days simulation. So in the one pool, you can see this red box represents the maritime continent area. And in the one pool simulation, it has some eastward propagating tropical waves. But in the maritime continent simulation, there is no conveyor, it's very limited to convection activities, not very strong. And then we compute the black correlation and it shows that in one pool simulation, the MJU like simulation is well propagating through the maritime continent while the MC when we have this cold aqua mountain, it does not. So two slides more. One moment. Okay. So and then we use this MC barrier SST and use the because we want to understand how much MJU prediction skill is reduced by adding this maritime content barrier. So we did the perfect model forecast experiment. So here step one is to the perpetual run. So here we use the one pool SST and similarly 10 years. So and then this is considered as truth. And then we save the restart file for every 10 days. And then in the step two, we do a perfect model forecast experiment. And here we use this restart file to initialize the model and then make 10 ensemble 45 day forecast over the 10 year period. So in this case, we use as a control simulation one pool SST. So if you compare the ensembles with the truth that is that give us an estimate of potential predictability regarding that it's a perfect model experiment and only the error comes from the initial condition. And then we simulate the maritime continent barrier. And we prescribe the MC barrier. So that means when we do this using the same initial condition, but then we have the maritime continent barrier in the barrier effect. So that mimics the current S2S forecast maritime continent barrier. So the question is how much focused skill is reduced by the maritime continent barrier. So this is the result of my last slide. So here is the RMM prediction skill by the one pool and maritime continent barrier. So here using the one pool, which is the perfect model experiment is the what is the upper limit of the MGO prediction skill. It's approximately six weeks if the model is perfect, although there's a lot of question about this perfect model experiment. But it seems that using 10 ensemble members, it has about six weeks predictability while the current model has three to four weeks. And then when we compare the MC barrier, you can see how much skill you reduce by maritime continent barrier and we can estimate and it is approximately five days reduction of MGO skill by the maritime continent barrier. Okay. And I will leave this and stop my presentation. Thank you very much. Great. Thanks a lot, Amy. A lot of results were covered. This is great. Any questions for Hemi? Highland has a question. Hemi, it's a very nice talk. Actually, I have two questions. First one is simple. It's how much improvement for the bivariate MGO correlation skill after the machine learning correction. Go ahead. Yeah, maybe. Yeah. So the RMM prediction, the bivariate correlation is about two to three days improved on average. Just like the big improvement. Okay. Okay. The second question is it's quite interesting that almost all the models have very similar feature is that it seems all like underestimated the kinematological like low level moisture. So what is the main reason? Is that they all produce too much precipitation? They just drop out of the moisture? Or is there other reasons? Something common? Yeah. So it seems that here shows the precipitation distribution in the S2S models and shows that here the black line is the GPCP precipitation. And here it shows that the model produce more frequent drizzle like precipitation. And so there's more drizzle precipitation and here there's just a moisture precipitation relationship. So here if you focus this black line is GPCP and all models you can see here they start the precipitation too early in the low moisture regions. So that maybe the deep convection is not sufficiently inhibited when this moisture is low. And that makes a moisture depleted atmosphere which can induce the dry bias. Okay. So it's a matter of improvement for the convection scheme? Yeah, I think so. And then we try to actually increase the internment rate in the CSM2 aqua planet. So here the red line shows the observed ERI interim and here the 1.0 is the default CAM6 version. And then if we increase the internment rate like 10 times more mixing like the internment rate then it gets closer to the observed line. And actually the MJO is also improved. That's very interesting. Also that's so we did some experiments with our model. It's also our funding. That's the internment is quite okay. Great. Thanks. So there are three other questions Hemi, but if it's okay with you would you respond to them on the chat Zayn? Okay, sure. Yeah. Have a question. Yaakal and could you post your question on the chat as well? We'll move to our next talk. Thanks again, Hemi. That's a great talk. Thank you.