Of super(duper)models and 'l@s nin@s'

00:00:02: A hoi and welcome to the Gen Z podcast, a podcast by early career marine scientists.

00:00:10: Hi

00:00:11: I'm Leonard And i am currently writing my master thesis To finish up my Master in Environmental Modelling at The University of Oldenburg In Germany!

00:00:20: And hi ,I Am Shalotte !

00:00:22: I have recently graduated from the Masters in Marine & Coastal Systems At The University Of Algarve in Portugal And together we are going to record some episodes for this podcast over the next few weeks.

00:00:34: So, Lennart let's dive right

00:00:53: in!

00:00:58: My title for my master thesis is the El Nino Southern Oscillation, Variability and Predictability in an Interactively Coupled Ensemble.

00:01:08: That for now just sounds very specific.

00:01:12: I think... It's a

00:01:14: lot yes!

00:01:15: Yes it definitely is but i mean to start maybe to explain parts of the title which would be in short always called ENSO.

00:01:31: And I think maybe a lot of people have already heard about it, and the case or may be in the news because normally when there's an El Nino or La Nina There are some yeah Some some headlines about it and Enso itself is a climate phenomenon with a global impact on that also why you would find out in the new It's said to be the largest energetic variability on a yearly scale, so in between years it can differ or make a lot of impact.

00:02:05: And even though it is located in the tropical Pacific, it does have those global impacts.

00:02:11: and as I said before already there are two phases.

00:02:15: There's El Ninoe and La Niña and El Niño would be The Warm Phase while La Niño Would Be Cold Phase Maybe to explain the whole concept a bit more.

00:02:28: and El Nino would be, for example when the walker circulation Would be disrupted And that what leads to anomalous warm sea surface temperatures in the tropical Pacific.

00:02:43: That's why it will be called The Warm Event.

00:02:47: I definitely read articles in El Niño or La Niña phases, and also know that El Niñu is this warm event.

00:02:58: But what's the worker circulation you just mentioned?

00:03:03: The worker circulation would be the atmospheric circulation which is very common for the tropical Pacific.

00:03:12: Basically it says or it's built from the trade winds that are blowing towards the west and kind of pushing the water masses toward the West, which normally leads to a warm water pool around the dateline.

00:03:34: So at one hundred eighty degrees west or east and west form this dateline at the coast of Australia and there would be a lot of warm water leading to convection, leading to rainfall.

00:03:54: And to close this circle then you have branch of atmospheric circulation going back into America's higher troposphere with subsidence or descending air masses at the coast of South America.

00:04:15: And also this is for example one reason why you would say that there's the Atacama Desert, because these air masses are pretty dry so they're not a lot of rain in normal circulation.

00:04:32: I understand

00:04:34: and then during an Inje event it will be kind-of opposite That these trade winds would be strengthened.

00:04:45: And therefore you have, because the winds will transfer warm water from the surface more to the west and that's why it has more cold water in central Pacific.

00:04:58: Those are those anomalous cold events how we call them.

00:05:03: Do I understand correctly?

00:05:04: The wind is always blowing into the same direction or what day they're.

00:05:13: they are originating from a different kind of atmospheric circulation.

00:05:20: So if we look at it on the global scale, we have cells or atmospheric circulation going from the equator more pole-wards and normally you have this uprising off air in the equator because there is a lot of solar radiation.

00:05:41: The trade winds are then basically the air that's coming back from a more southern or northern position towards the equator to balance out this uprising of air due to the Coriolis force.

00:05:57: They're normally facing in the same direction, so when they were coming go westwards or to the left.

00:06:05: If they're coming from the north, They are deflected towards the right so also to the West and that's why we have these trade winds.

00:06:12: That way you could call the east of this in case

00:06:16: I understand yes.

00:06:18: And then these straight winds Are impacting where the warm water Where the warm-water flows?

00:06:24: And this intern has impacts on The weather patterns but also on the sea surface temperatures.

00:06:32: Yeah indeed

00:06:38: During these events, for example fisheries are sometimes completely disrupted because there is not so much cold nutrient-rich water upwelling along the west coast of the Americas.

00:06:50: For example which means less fish available and what other global impacts do they have?

00:07:00: First off I was pretty well explained that it's on point.

00:07:04: So yeah, because of the trade winds you have this warm water being pushed away and then deep water is being upwelled.

00:07:14: It's really nutrient-rich but it has very widespread diverse impacts.

00:07:24: so El Nino or La Nina they can transport their signal via teleconnection like stationary planetary waves that transport via pressure differences, they signal over very large distances.

00:07:43: And for example it's shown that El Nino can have or at the end so in general can have impacts For example onto Australian temperatures Or also precipitation Also for North America but also on the European continent or the African continent.

00:08:06: So if you correlate, or regress the El Nino anomaly temperatures for example with precipitation You can see that there's normally Yeah averaged over different events.

00:08:21: of course There is certain regions which are normally warmer and wetter or warmer, drier during El Nino.

00:08:30: And then often they are colder and wetder or colder and drier depending on... yeah doing the El Nina.

00:08:37: so there's really a large impact on global scale also floodings or droughts.

00:08:47: it can have an impact on agriculture infrastructure as well.

00:08:52: that is why this Pretty important topic worldwide to have some sort of a good prediction.

00:08:59: Of this two be able to somewhat Yeah Know the state of the end so and know what you are expecting?

00:09:09: And then also, I'm the answer can Impact global mean surface temperature by quite significant amount.

00:09:18: So if you would look at average of the global means you like that.

00:09:22: The time series off the surface temperature of the world, You would see that normally El Nino would be peaks of this line while La Nina would be more like negative peaks.

00:09:36: And with these surface temperatures do you mean also terrestrial surface temperatures or only ocean surface temperatures?

00:09:45: No no it's really a global mean.

00:09:47: so really Ocean temperature and also normal air surface temperatures.

00:09:55: We can still see that the negative peaks would be normally related to those lannina phases, while lannia would be more adding on top of the normal warming trend we have.

00:10:08: so it will be part of this internal variability in climate systems.

00:10:18: That's interesting!

00:10:21: how it's important to sort of predict in what phase the globe is and when the next different phases coming, since there has such an opposite effect on some areas.

00:10:37: But because they are so connected inside the Earth system... It sounds difficult to forecast!

00:10:48: It's hard to distinguish all the time if it really isn't now an impact of the ANSO or is that just different modes of variability.

00:10:58: I mean, it's a very connected phenomenon and yeah... ...it's harder to trace back all the times.

00:11:08: And i assume this is why people use supermodels for this huh?

00:11:14: Well we try!

00:11:16: Which is a very nice translation to the next topic, which would be right.

00:11:21: Next

00:11:21: part of my title where I talked about an interactive ensemble and what we also call a supermodel?

00:11:41: It's not really clear how that actually works.

00:11:44: Do you have any idea?

00:11:47: A

00:11:47: Supermodel?!

00:11:48: No!

00:11:48: To me it sounds like something from science fiction.

00:11:57: So first an ensemble is when you have different models or take different model runs and then the interactive part in this supermodelist that we Take different models, and let them interact.

00:12:15: so they are exchanging information of some sort And This Is what makes it the supermodels.

00:12:26: So, do I understand this correctly?

00:12:29: You have either a couple of different models or you use the same model but with different initial conditions and then let all these models sort-of talk to each other.

00:12:43: Exactly so in our case.

00:12:45: we do use different models.

00:12:51: We take models talk on a certain frequency, so we let them interact once month basically where they exchange sea surface temperature.

00:13:06: So for us this exchange is limited to our ocean module of the models.

00:13:13: and The thing is that normally if you have an ensemble these standard approach would be averageize after the models have run.

00:13:28: For the supermodel, it's a bit different because The models can interact while they are running which makes them kind of synchronized.

00:13:42: and this synchronization is supposedly improving the model Because its taking Let's say the best characteristics of each model and combining them to a new model.

00:13:58: And that's why it wouldn't be classified as a statistical approach after the models run, but while the model is running.

00:14:08: Also The next thing for the supermodel we are kind using hybrid approaches also by machine learning because Like, we do not just average the state while models are running.

00:14:28: So it's not like... The model is giving equal weight but somebody before me has done a training where they weighted the models according to how they performed best at a certain calendar month and the certain grid cell.

00:14:54: And this is what makes the supermodel, uh...a bit special.

00:15:04: How was this weighting done?

00:15:06: So were there intermediate results compared to observations?

00:15:09: or how did that colleague before you know which model to give more weight?

00:15:17: Yeah exactly so it was training then using some observational data for period of I think it was something like twenty years or so.

00:15:28: And this machine learning approach then tried to different values for the weights and then looks at which, yeah...for each grid cell in each calendar month Which combination of weights gives us the best result?

00:15:44: That's what determines the weight in end.

00:15:49: This is also the reason why you just would not use this one single model, which has values closest to observations because every month a different model could be closer.

00:16:01: So that's then this combined advantage...

00:16:05: Exactly!

00:16:06: It's like a lot of work goes into training and optimizing these models.

00:16:11: I always think it's very fascinating.

00:16:13: Indeed yes That's also one of the downsides for this supermodel.

00:16:20: It does take computational power and it is quite complicated to do, especially because those models are a bit different a lot of times.

00:16:31: so there some technical problems but I think maybe we'll talk later about that more.

00:16:50: You were working with such a super model And how many models part of your supermodel and how was the rest of setup that you were working with?

00:17:04: Yeah, so I was using a Supermodel which was done at the Nansen Center before me.

00:17:11: And it's using three Earth system models namely the NOR ESM CESM and MPI ESM.

00:17:22: They are then connected as i said by their sea surface temperature which they exchange every month.

00:17:31: And then these pseudo observations as we call them, are re-assimilated into each of the models and basically started again from this point on running for next months.

00:17:43: so that step is repeated and so forth all their whole run time.

00:18:03: So now you know how to set up a supermodel with it And how can such a supermodel help to represent the complexity of the Enzo phenomenon and also improve predictability?

00:18:19: Yeah,

00:18:24: so that's an interesting question actually.

00:18:26: So the individual models as we've seen them if would just run on their own they are.

00:18:36: They are somewhat similar, but they're slightly different in their treatment of different processes.

00:18:43: So for example some of them have different grid resolutions or also a different kind of grids than there's also parametrizations.

00:18:52: so that process cannot be represented on the models because it is too small scale and those are mostly empirical done.

00:19:04: So there is still some sort of uncertainty in them.

00:19:08: And also the numerical solving can introduce slightly different results between the models, and also to boundary conditions or just yeah... There's just some technical differences within the model.

00:19:24: so they have slightly different physics even though based on same physical laws.

00:19:33: So the results are just differing a tiny bit.

00:19:35: and if you then combine them That's how we try.

00:19:39: what?

00:19:39: We're trying to do um, Then we're trying too.

00:19:43: Yeah take the best aspects of each model And create A new model with it.

00:19:50: basically

00:19:53: I see, yeah.

00:19:54: That also reminds me a little bit of what we talked last week about the sea ice models and how they all rely on the same thermodynamics.

00:20:01: but then one of the models is much better in using these laws to calculate light flux whereas other model has some aspects which help simulate better cracks or openings.

00:20:18: Yeah, and I guess by taking the best aspect of each model you're also excluding all the imperfections or a lot of the imperfection that models still have.

00:20:30: Of course

00:20:32: yeah exactly!

00:20:33: You can definitely see it like this.

00:20:35: but so... The supermodel can take these advantages off those different model physics then dynamically combine them.

00:20:42: And the nice thing is not necessarily you need to have like the most perfect super duper models for everything because You can see it kind of as if you have the truth in the middle, then you put the models As a function of their physics or some variables are whatever and they're kind of scattered around that.

00:21:06: And sometimes It might even be good to have models worse for this one variable because they are then opposite of the truth towards the other models.

00:21:21: And since we're dynamically combining them, We are kind of centering them.

00:21:27: so you can see it like as we scatter all the other modules around a truth and when we dynamically combine them will end up in middle Yeah,

00:21:44: so you're sort of balancing out the mistakes that the models do.

00:21:49: That's a bit the point.

00:21:55: and then also another perspective or point of it is that non-linear behavior is theoretically captured by this combination.

00:22:08: So non linear behavior in this case has like lot of those sub grid scale processes.

00:22:15: they are then coupled to Yeah, or there are somewhat slave to the bigger processes.

00:22:22: So if we represent those better.

00:22:25: We hope to improve like a bigger picture.

00:22:40: so you're using this huge model set up To trying to improve predictability of end zone which is very important because it has sometimes devastating impacts on agriculture people living in coastal areas and so on.

00:22:54: Can I go little bit more into details?

00:22:57: What methods you use to generate results in your master thesis?

00:23:04: So what I am actually looking at now is about forty years of model data and i'm normally using the so-called free run.

00:23:14: We just start a model, then it started in nineteen eighty and runs until twenty twenty In between.

00:23:22: there's no assimilation with any observations or So it's just the model is evolving on its own.

00:23:31: Additionally to this I look at so-called hind casts, where we reforecast one year... We have about thirty of these where a forecast, a reforecaster started basically for example in January first and then it runs until the end of or beginning of one.

00:23:58: And so like this I can evaluate how good model is at predicting this one year.

00:24:07: basically

00:24:08: because you know,

00:24:10: exactly.

00:24:11: we have observations.

00:24:12: now compare it to observations and that's what i'm doing with a free run and behindcast simulations comparing them two observations but also say if its better than normal models, how we can call them.

00:24:28: We also have the exactly same models that were used for the supermodel but just run on their own so where they are not coupled or exchanging information.

00:24:41: So here I am then looking at those three models And the supermodel itself, because those three models are not exactly the same.

00:25:01: We still have tiny differences between them even though they're synchronized in a Supermodel.

00:25:08: That's why I'm looking at them individually as well but also at average of them and then also non-interactive versions.

00:25:19: so where there just run alone.

00:25:20: And then here I am comparing different variables, for example sea surface temperature precipitation winds or like wind stress and also ocean currents.

00:25:53: Did you see big differences between the interactive ensemble and non-interactive ensemble?

00:26:02: We did see some, for example in the tropical mean state what we call it.

00:26:07: So where I just look at... Yeah!

00:26:10: Where basically average to free run over the years and there you can see that the interactive ensemble.

00:26:18: so this supermodel is actually improving The pattern i would say?

00:26:27: Basically an example of sea surface temperature.

00:26:32: As I said before, at the beginning you have normally this warm pool in the west close to the coast of Australia.

00:26:39: And then you've got the equatorial cold-tongue – how it is called?

00:26:43: It's basically a cold water tongue extending from the South American coast towards the West and what we see a lot of times on normal models that these cold tongues are extending too far into the West.

00:26:59: And that is then important, for example.

00:27:02: For those ENSO processes because as I said before in ENSO there's this disruption of the winds you have warm water coming to east and all these interactions.

00:27:15: If cold-water tongs are already too far west at the beginning Because the model simulates it way too large or too much this will kind of shift the responses a bit and therefore impact also, um... ...the results on how an El Nino or El Ninio will evolve.

00:27:40: Or what impacts the ENSO will have globally because it's not only the face but also position where an El Niño is happening that gives us information about largest impacts, for example.

00:27:58: Yeah I can imagine if you have cold water much more towards the east then there is less air rising in the East sort of and that will change wind patterns that are resulting from...

00:28:11: That's exactly the point yeah because at the end of this walker circulation you have this rising branch where they normally also have cloud building rain.

00:28:26: And this convection center during an El Nino would be placed more over the Pacific because you have this shifting, or this disruption of the winds.

00:28:39: so it shifts to the central Pacific and then you've increased rain basically over the pacific.

00:28:46: That is not as strong.

00:28:49: in case if this cold tongue is preventing especially for the moisture to be taken into the air and then raining, it has to exceed a certain threshold.

00:29:03: And if those waters are too cold this threshold will not be exceeded!

00:29:07: Then we have a lack of shift on convection cell... That's an nice translation to next topic because also there is the intertropical convergence zone that maybe some of you know from geography lessons in high school or so.

00:29:29: So it's this band of rainfall that is normally north of the equator.

00:29:33: And what we see in normal, or I always say normal... The models are that they tend to have a bit of bias there and also a southern branch of this ITC Zad.

00:29:53: There's some rain, there's just this rainband North of the Equator meandering a bit about.

00:30:01: And there's this southern band, which is not really that strong in the real world and In The Supermodel we can actually see That This Band Is Reduced.

00:30:16: Oh So Thats A Huge Success.

00:30:18: Then To Make It Closer to Reality Yeah?

00:30:22: It Is Very Nice Though Also The Problem That The SuperModel Tends To Rain Too Much.

00:30:28: Anyway.

00:30:29: So the rain that's in the north is too strong.

00:30:32: But, you know... We would take it when there's always something!

00:30:38: You have to bring an umbrella if you work with supermodels?

00:30:40: Okay noted!

00:30:41: Yeah

00:30:43: exactly.

00:30:50: And for example what I saw was If we take a certain index which is called the center of heat index which basically just gives you the longitude of an ENSO month where it is the strongest and It also gives you an amplitude And that we can see, that this supermodel was able to improve The positioning of the ENSO.

00:31:17: So if you compare it with observations We can see normally more stronger Elinio events are located in East Pacific So more close to the South American coast and a lot of times because partly because of this Cold Tongue buyers, also these double ITCZ buyers.

00:31:41: Normal models would locate strong LNU events closer to the Central Pacific.

00:31:50: so there will be a bit of divergence between them.

00:31:53: but in the supporter now we can see that this is definitely improved, so the stronger events are more located into the east.

00:32:02: but we see that the amplitude is reduced in supermodel.

00:32:05: So again have a different kind of problem or a different question that's now posed.

00:32:12: well... We will look at how to improve it then again.

00:32:18: But its quite promising results for us already because one of the first ones that's using those big earth system models.

00:32:31: And I think it is already quite exciting to have results like this, which actually show its working!

00:32:39: It just needs to

00:32:41: be

00:32:42: improved.

00:32:55: Yes, it sounds like this supermodel does what you hope.

00:33:00: It would do by for example introducing this southern rain band or improving the location of the center of Enzo.

00:33:12: Are there any other?

00:33:14: But are they... You also mentioned some limitations Or

00:33:19: stuff

00:33:20: where doesn't so good For example simulating much more rain.

00:33:25: Yeah, yeah.

00:33:28: Well it's not perfect for sure No but there are still some limitations that we can see just when building this model For example.

00:33:40: what is done now so far Is only interacting with the ocean.

00:33:47: So only the sea surface temperature has exchanged.

00:33:50: So the question is maybe because the El Nino for example, it's a phenomenon that's taking place between atmosphere and ocean.

00:34:01: There are lots of interactive processes there and feedbacks happening.

00:34:06: so For example if we would also let the atmosphere interact That might be already give us completely different picture.

00:34:15: The problem with that Atmosphere processes are way faster, so you would need to let the model interact a lot more times.

00:34:28: And that's taking of course a bit of computational power because for this yet they kind-of need to be stopped.

00:34:38: then You take out data and you'll let models exchange Then restart them.

00:34:44: That is technically just difficult.

00:35:00: Do you want to share some more thoughts on the process of writing your master thesis and doing this research?

00:35:08: Definitely, so for me it was The structure in the end that was most important I would say.

00:35:16: So because... ...I've been working on this now quite a time And sometimes i'm just too focused like small things.. ..and then kind-of forget That I have other stuff.

00:35:31: I just need a structure to work around.

00:35:34: So, i don't know...I lost maybe a week because I was too much going into one direction of literature research and then noticed in the end yeah well if i would have structured it more clearly at the beginning i would've noted that like this direction is not leading me anywhere..i dont'need go that deep into it!

00:35:54: Because its not answering any questions than ive actually.

00:36:00: Yeah, I think this is also a little bit... This risk when you have an open exploration of data.

00:36:08: Whereas compared to your setting up an experiment and trying to prove or disprove a hypothesis yeah i can relate sometimes the similar case

00:36:21: at some point it's just the question Of structure And You Just Need To Have A Real Good.

00:36:29: I think, yeah you should have a good research question in the beginning to start with and then structurally.

00:36:37: Think about it first before you go into data processing or data evaluation because what i tend to do is sometimes too fast... ...I want look at the data so I just plot something.. ..and that ends up me plotting five hundred plots per week And being like okay!

00:36:55: I will probably use one of those

00:36:59: And you

00:37:01: have to find that one.

00:37:02: Yes, exactly!

00:37:03: So

00:37:04: it's just a bit exhausting sometimes.

00:37:08: and then another thing.

00:37:09: I was a bit in the beginning difficult for me but we also talked about last time is this huge amount of data especially because now i have three models ones interactive version.

00:37:24: Then I have that for the free run, which is already forty years of data and also For the hindcast where it's actually only thirty years Of data.

00:37:34: but in The hindcast there Is even Also additionally to case That we Have ten members per model.

00:37:41: so There are Actually Ten versions of each Model.

00:37:46: It's a huge amount of Data.

00:37:49: That you have To process.

00:37:53: sometimes I was really struggling to keep the overview of which data am i actually looking at right now because it's just so confusing.

00:38:06: Yeah, imagine yes but also think um...I assume its same for you as is first time ever working with that amount of data.

00:38:19: So a lot these data management tools or yeah, tricks that you know about in the end.

00:38:31: You don't know them at the beginning.

00:38:32: it's something we develop over time and I'm very hopeful that next time We both start to work with a new large data set Or multiple large datasets And then yes be much smoother right from the beginning

00:38:51: I'm

00:38:53: sure.

00:38:54: Yeah, no i think that's a very important part of the learning process too.

00:38:59: yeah it's the skills you acquire during such a process.

00:39:02: yeah

00:39:02: i definitely.

00:39:03: yes yeah cool.

00:39:07: well lennart thank you very much for giving all these insights and great explanations on enzo la niña el nino how you worked with super models.

00:39:21: It was a pleasure, I hope it wasn't uncomfortable.

00:39:23: I thought...I

00:39:24: thought it was very insightful and interesting.

00:39:28: so thank you very much And with that i'm looking forward to our next episode!

00:39:34: We will talk more about models.

00:39:38: once again.

00:39:39: we'll

00:39:39: ask some questions on how reliable they are.

00:39:44: why need them in addition to predicting Enzo?

00:39:49: Alright, then thank you very much and I'll

00:39:52: see you next time.

00:40:15: Bye!