 Okay, okay, so So my name is Alicia Karsbeck. I'm at NCAR and so I prepared actually a very broad overview talk about thinking about some of the the issues around Ocean data assimilation in initialized to cattle prediction. So this is again very broad. There's some results in here, but It's gonna feel like a bit of an overview for a lot of you So I wanted just to sort of step back a little bit and Over you where do we get this idea that the ocean visualization matters for to cattle prediction? So there was a whole set of idealized studies that happened years ago that kind of created the underpinnings for this this community So I'm going to just trot out this This diagram that I'm sure many of you have seen this is Adapted from a mule at all paper, but it's been in a number of papers. I think it's in an IPCC report That shows that in terms of decadal prediction. We sort of sit in this muddy region here between This the daily to the seasonal out to the long-term projections where we think that the influence of the initial conditions Is going to matter in terms of the skill of our decadal predictions And there's a whole series of papers that have been in idealized studies that have Try to illuminate that there is some skill that you can get from initializing The ocean and this is a result from more recent paper by brand Statter and Tang 2012 that I really like and what they're doing is they're looking at this is again in an idealized framework Looking at skill in terms of this measure called relative entropy So relative entropy for those of you are not familiar with this skill It measures both shifts in the mean of your predicted distribution and also in the width of the ensemble So the variance and what they're trying to show here is that as you move through time You actually lose skill due to this initialization, but you gain skill due to the external forcing, okay? So this is sort of this idea and what they showed here is this there's a little crossover point that's here around in the decadal regime here where there's an equal Influence of both the initial conditions and the external forcing and one thing that we don't often talk about in terms of this But I find very relevant is if you were just to visually add up these signals, which is the total skill What you'll notice is that actually we sit here in this minimum, right? So let's be easy on ourself and recognize that even from a theoretical perspective, right? We actually are in this area where the total skill is thought to be a minimum Okay, so what about the idea of digging in there? Where is it in the ocean that we think is it important to do the initialization? And there's actually a relatively rich literature about this in the seasonal prediction community Not as rich in the decadal prediction And this is a again a relatively recent paper by Dunstone and Smith where they try to actually understand Where is it in the depth of the ocean that the initialization matters again? This isn't an idealized framework So this isn't in there's a four different regions global ocean Pacific Northland again south southern ocean and This in pink up here is the skill you would get at different lead times if you only initialized with surface information So SST and atmospheric forcing information Whereas you get this reduction in error. So this is an RMS. So lower is better When you initialize the subsurface of the ocean all the way down to depth, okay? So this is a very general result That also gives us idea that when we think about it what think about ocean initialization We're really talking about initializing the entire subsurface ocean Okay, but what about what do we know now about the role that ocean initialization plays in determining real prediction skills? So everything up till now has been in idealized studies Okay, so we've already gotten some results from brim. Thank you for from the doblis rays nice paper where they Where they analyze the skill from the whole set of seam of five archive? And this is just an overview plot of this So what you see in the top here is at two different lead times This is the total amount of skill read here is high high skill now And then down here is the portion of that skill that's due to the ocean initialization And you can see that the really the only region that's left with significant skill and significance here is in Stippled is here in the North Atlantic, but for the vast majority of the ocean you it's really difficult to see that there's any influence of Initialization on the skill So this is that was a multi-model ensemble This is very consistent with the results that we see with the NCAR CCSM for a model where I'm only showing here now things that are statistically significant There's broad regions of high skill But if you ask the question, where do we see skill that we can attribute to the initialization? There's actually only one place that we have left and that's the North Atlantic So this is really seems to be a key region in terms of where there's actually practical skill in our models do the initialization So here's a Overview now of what are these big challenges that we've been talking about and there's been a lot of talk sort of Mentioning these big challenge in terms of decadal prediction. We have model quality There hasn't been a lot of talk about detectability of skill But it seems like there's always an undercurrent of talking about statistical significance and how we can detect skill And this is really I'm trying to focus here for the rest this talk here on this other piece Which is initialization difficulties or complexities and in specific these two pieces of it Which are what are the best strategies that we know about for initializing our models? Turns out those are actually unclear at this point and that there doesn't seem really to be a strong consensus on What is the state of the ocean now or in the past? Okay, so for those of you again who are students and this is just an overview of what we mean when we talk about an Initialization system for a decadal prediction. So you can imagine you have your GCM. It has external forcing Hopefully you have some users on the other side. I'm not sure who they are and Then you have some observations of the climate system And I'm using this word very vaguely right because this can be a whole set of different observations It could be observations of the atmosphere observations of SST observations of the subsurface lots of different things and in fact people kind of exploit the fact that a Lot of this information may be redundant And there are many many different ways that groups choose to initialize their their models and here is just a set here of Actual ways that people initialize their models Of course we think about the most we think about data assimilation systems as being the obvious way But in fact many groups do other things they do an atmospheric forest ocean hind cast as a way to initialize They do nudging of an ocean or coupled model to just an SST product Nudging a coupled model to a subsurface data from another product Interpolation of a foreign product to your model grid you can choose to do anomaly Initialization where you're actually using the climatology from your model or full field So there's really a whole variety of ways that you can think about doing getting your ocean into a state that could launch a forecast Okay Okay, so this is So what do we now know about what are the best strategies that are useful for initializing decadal predictions? And the the fact is that there's actually at this point in time No clear way to tell the difference between these different strategies. That's the state of affairs right now I've listed six papers here In terms of the question of whether it's better to do anomaly or for full field initialization Here's three examples of papers that basically say that there's no way to tell the difference at this point between these two strategies In terms of the question of whether or not it's better to initialize from an atmospheric force time-cast or to use data Assimilation here's three papers right here who say that you first order you can't really tell the difference between these two strategies So this is really the state of affairs right now. We don't know what the best way is to initialize our model Here's an Okay So here's in a quick example from the end our experience with the NCAR model Okay So this is an a sort of an assessment of skill in the Atlantic some polar gyre using The same model but initialized with data assimilation Product and initialized with just an ocean hind cast and let me just cut to the chase here These distributions right here are distributions of skill in the green. You said there's marginal Improvement actually when you use don't use data assimilation. So the data assimilation actually Reduces the skill to a marginal extent in this area But then again, if you look in the equatorial Pacific you see that if you initialize with this particular oceanized Hind cast you actually get a huge degradation in terms of the ENSO skill and you get modest skill if you use data assimilation So there's really no clear answer here In terms of that for us Okay, and so what about this idea that there's we don't really know what the state of the ocean is So this is a plot that many of you have seen a lot and it's recently been updated in a new Special issue. This is from Magdalena at all looking at ocean global ocean heat content In a whole variety of different ocean analysis products And we see that if you look on the upper 300 meters for example, you might say that's a pretty good agreement But once you as you get to larger and larger depths You start to see that there is really quite a bit of uncertainty in terms of the way we understand what the ocean heat content is Okay, and if you think about Elements of the ocean state that are actually more interesting in terms of decadal prediction You see that there's actually a wide range of different Variability and in trend to cattle trends in terms of what you get from different ocean reanalysis products So this is a set of six different ocean reanalysis products That I analyzed here at 45 North in terms of the amok so the amok is a good proxy for the northward heat transport into the Subpolar gyre and one of the interesting outcomes of this paper was that it does by I you would say yes There's not a lot of agreement here and interestingly if you don't do assimilation with those same models They tend to be in greater agreement. So to some extent the the act of assimilating is actually drawing these Solutions away from one another Okay, and so and this is something that we found in a lot of different papers. This is Munoz at all has shown this similar thing where in terms of variables like sea surface temperature Which are very well observed you have a lot of agreement with ocean products But as soon as you look at emergent properties like the northward heat transport, which are very important for decadal prediction You don't see a lot of agreement anymore Okay, so Okay, I'm gonna skip this except to say you can I hope you read my paper 2015 We're actually map out where where in the Atlantic you have similarities in terms of the hydrography between different ocean reanalysis products Okay, so but let's not to be too harsh on people who date do data assimilation So I did a talk a while back on that data assimilation basically has all the ingredients for a very challenging problem We have to deal with the fact that our models are imperfect and biased We have all sorts of sub-optimalities that are due to the way that we do our data assimilation And plus we are dealing with an observational network that sparse in homogeneous and has a changing Expression in time as well So it's a very very challenging problem to do ocean data assimilation and unlike atmospheric data assimilation where a Miss you can actually lose the memory of the mistakes you make doing data assimilation in the ocean You retain a memory of every mistake that you make when you're doing the data assimilation Okay, so I'll skip this that's our goal. That's a changing observing network. It's very very obvious and say okay So let me this is just the conclusion So I spent some time thinking about some of these issues and what are the path paths forward? Okay, so I sort of touched on two main issues one about that There's no consensus on the historical state of the global ocean and the assertions that I have to make about this Is that really there's nothing we can do about our lack of information in the past or really in the near-term future in terms of the Observing network that's something we have to live with at this point There's nothing inherently wrong with diversity of models or diversity of DA methods And those will naturally result in a whole diversity of different different solutions And the other thing is that like model development DA R&D is a long-term endeavor Okay, and there's a long-term payoff associated with that and so in terms of the path forward We can think about things like embracing the uncertainty in our ocean Understanding of the state of the ocean and that involves forming prediction ensembles that use multiple ocean reanalysis or multi-model archives Do that very naturally? The other thing we can do is focus on observing networks of the future Enter and think of those from the perspective of what we need to constrain the global the global ocean My perspective is they should be cheap noisy ubiquitous and sustainable So I don't mind if they have a lot of errors as long as they have a lot of coverage The other thing is to think about investing in ocean DA systems really with the same seriousness that we invest in models in Terms of the issue about best strategies for initializing models being unclear at this point I would say that it is sort of the nature of the decadal prediction problem that our effective sample size is too No too low at this point to really detect small modifications in skill Okay, so that's just sort of the fact that we have to live with now and The benefit of using advanced ocean DA system may become more evident as the sample size increases and as models improve The other piece is that the performance of real-time forecasts is really the only true measure of progress And so the path forward based on those assertions is that there's really should be a big importance placed on long-term archives A real-time forecasting and there's some examples of some that are happening now And that those archives may need to contain some information about how those forecasts were initialized so that we can have records of that moving forward And that even though we're not seeing this substantial benefit of DA at this point It's worth keeping that in the mix of possible initialization methods That's all