EWGLAM Meeting 2021 - Day 1
szeptember 27, 2021
(8:30) Guest 2: Good Morning
(8:44) Milan, Marco: good morning
(8:45) Mike Bush: Good morning everyone!
(8:45) Anke Finnenkoetter: Good morning! 
(8:49) rontu: Huomenta Helsingistä!
(8:53) Maria Monteiro (IPMA): Good morning !
(8:55) Patricia Pottier: bonjour
(8:56) Dmitrii Mironov (L): Balazs, this is Dmitrii. Can you here me?
(8:57) Milan, Marco: Hi Balazs. You small image is moving. The big one is fixed, at least for me.
(8:57) Ekaterina Kurzeneva, FMI: Good morning! 
(8:59) Claude Fischer: bonjour tout le monde
(8:59) ECMWF Florian Pappenberger: Good morning
(8:59) Simon André: Good morning to everybody from Bratislava!
(8:59) Jure Cedilnik (ARSO): Good morning.
(9:00) Susanna Hagelin (SMHI): God morgon!
(9:00) Petra Smolíková: Dobré ráno!
(9:00) Inna Rozinkina RHMC: Bonjour, Good Morning to everybody from Moscow!
(9:01) martina tudor: dobro jutro
(9:01) Aline Kraai (KNMI): Goedemorgen
(9:02) SRNWP-EPS Alfons Callado: Bon dia !!! (Catalan)
(9:07) Marion Mittermaier: Morning!
(9:08) Anastasia Bundel: Are there emojis to applaude in blueJeans?:)
(9:08) Elena Astakhova, RHM: Good morning!
(9:15) Reima Eresmaa (FMI): Hei
(9:17) Siham Sbii: Good morning
(9:35) Raschendorfer Matthias: Mike, does "vertically distributed canopy layer" mean that a kind of porous-medium flow simulation within the roughness-layer formed,e.g., by buildings or trees is faced?
(9:41) Mike Bush: Development of a vertically distributed urban canopy: This will require a consistent distributed representation of canopy fluxes (including radiative exchange) and drag. Another challenge is how to incorporate the effects of isolated tall buildings or groups of tall buildings which lead to wake effects.
(10:00) Roger Randriamampianina: @Claude: Can you please inform us about the cost of the porting to GPU in term of power energy?
(10:01) Raschendorfer Matthias: @Mike: Thanks Mike, I completely agree; and this is among our aims as well.  However by doing this, the formal definition of model layers is an issue, and the porous-medium approach with model layers being intersected by rougnness-elements is one possibility.
(10:14) Claude Fischer: @Roger: you raise an important question. I do not have myself figures on this, and I will forward your question to our experts. Keep in touch.
(10:14) Marco Arpagaus --- MeteoSwiss: @Jeanette: Can you elaborate a bit more as to why SPPT was abandoned?
(10:16) Inger-Lise Frogner: SPPT adds very little to the variability in our system when combined with the initial, LBC and surface pert. See my presentation about this last year, or wiat for the paper :)
(10:16) Marseille, Gert-Jan (KNMI): @Jeanette, Sander. Is it possible to experiment with UWC-West before 2023?
(10:17) Marco Arpagaus --- MeteoSwiss: Thanks, Inger-Lise!
(10:19) Sander Tijm: We are already testing with a UWC-west domain at ECMWF. But not yet with full DA as the structure functions are still being derived
(10:21) Marseille, Gert-Jan (KNMI): Thanks Sander, that sounds like DA with UWC-W should be possible in 2022
(10:23) Sander Tijm: I hope a little bit sooner
(10:27) Marseille, Gert-Jan (KNMI): Great! thanks
(10:33) Javier Calvo: LACE does operate a common deterministic suite, right? If so what is the reason for that?
(10:39) Martin Bellus: @Javier: No, LACE doesn't operate a common deterministic suite (however, each member state runs their own deterministic model/s). We have a common LAM EPS suite A-LAEF running on ECMWF's HPCF as TC2.
(10:46) bazile: what about dynamics impact on wind quality at 500n ?
(10:49) Florian Meier: Do you know, if there are also problems with T2m/RH2m in the alps with 500m resolution or is it only wind? We faced some problems with 2m values in valleys recently, when increasing resolution of orography/model.
(10:52) Roberts, Nigel: Dmitrii, Really interesting presentation. I'd say km-scale models are pretty good compared to coarser resolutioin if you look at rain structures compared to radar. I'm wondering what is the definition of good and poor for variables of interest.
(11:34) Dmitrii Mironov: Florian's  sounds is often broken. Is it only my problem?
(11:35) Martin Leutbecher: same sound issue here too
(11:35) Usuario AEMet: No, I think it's general.
(11:37) Macpherson, Bruce: slides not moving on?
(11:38) Alena Trojakova: slides move ok for me
(11:39) Usuario AEMet: The problem was only in his sound, the slides move ok. 
(11:56) Dmitrii Mironov: Peter, could you give some detail what  is meant by two items starting with "Learn" on your "prospect for machine learning" slide?
(12:12) Dmitrii Mironov: Peter, I was to quick, sorry, much of he answer to my questions is in your slides. Thanks.
(12:13) Dmitrii Mironov: Not really.
(12:13) Dmitrii Mironov: OK
(12:15) Roberts, Nigel: Will here ever be enough observations to learn fine-scale weather?
(12:19) Hannah Christensen: Great talk, Peter! You showed the ENSO results from Ham et al, nature - but they don't compare their CNN to ECMWF forecasts. Do you happen to know if they beat you too?
(12:19) Roberts, Nigel: Thanks Peter, very interesting.
(12:20) Peter Dueben: @Dmitrii: Sorry, I am not sure I understand your question, but please reach out and we can discuss offline peter.dueben@ecmwf.int
(12:22) Raschendorfer Matthias: Thank you very much for this very good overview, Peter! My question: Wouldn't it be a wise approch anyway to use as much as possible of valid physical knowledge, in order to employ ML only to describe the remaining uncertainty introduced into the model by wrong or missing closure assumtions. So why are  approches beyond the "delta"- or "error"-approach an issue at all?
(12:23) Peter Dueben: @Hannah: Sorry, I do not know. But I am very much looking forward to the results of the WMO challenge on S2S predictions in this context: https://s2s-ai-challenge.github.io/
I can also recommend the paper by Jonathan Weyn to compare ECMWF seasonal predictions with machine learning solutions:  https://arxiv.org/abs/2102.05107
(12:25) Peter Dueben: @Matthias: Well, we have seen that physics informed apporaches can help a lot to improve results. However, in other domains we have also seen that an "all-in" approach can get you a long way, and even be better than the conventional tools. I think we should do both. Bottom up to learn how to build small tools that are physics informed and also top down and build large tools to see how far we get.
(12:49) Christoph Gebhardt (DWD): 👍
(12:56) Claude Fischer: Resume at 14:00 CEST
(13:54) Jana Sanchez: yes!
(14:29) Pascal Marquet: does it means that ML is better for dimilating sqalll lines ?
(14:29) Pascal Marquet: simulating
(14:50) Peter Dueben: The simulations within DYAMOND seem to behave quite differently in the vertical (e.g. for w). Is this a problem? Or is it maybe even a good thing as it will help to understand sources of errors based on differences of the schemes used.
(14:53) martina tudor: would the systematic error for tropics above sea have any relation to systematic error in midlatitudes over land? should we recompute it forveach area, horiz and vert res. and timestep?
(15:01) Tsyrulnikov Michael: Coarse-grained hi-res fields may not lie on the low-res model attractor. The resulting transient (spinup) when you start the low-res forecast may mask the small model error. How do you cope with this effect?
(15:01) Pirkka Ollinaho: Hi Hannah, nice talk 🙂 Is the 3rd "perfect for this synoptic situation" SPPT scale equal to a static in time and space perturbation (infinite, infinite)? If so, does this imply SPPT should (in this case) apply skewed perturbations to the tendencies?
(15:05) martina tudor: thank you
(15:10) Hannah Christensen: Hi Pirkka! Thanks for your question. With the size of Cascade, I wasn't able to robustly estimate the spatio-temporal scale of that third scale, though I could detect the need for a larger scale pattern. But your question about skew is spot on - when I analysed the distribution of 'e' I found that it was indeed skewed - an average skewness of 0.6, higher over land, and higher at night time
(15:11) Pirkka Ollinaho: Thanks Hannah, that's a really interesting finding!
(15:16) Tsyrulnikov Michael: @Hanna: if you compare the two forecasts starting from lead time 1h, then the two tendencies you compare do not start from the same point in phase space, which is required by the definition of model error. So the difference you obtain may be not only due to model error but also due to the difference in initial states.
(15:26) Hannah Christensen: Hi Michael - thanks for your comment. I'm very interested to hear all suggestions for improvements to the protocol. Discarding the first hour is, in my view, necessary, as the error statistics are not representative of the error statistics within the bulk of the forecast model simulation (needed for stochastic parametrisation development), if what we are seeing over that first hour is the rapid spin-up of the forecast model.
(15:37) Carlos Geijo: The software for identification and association of objects is home-made or off-selvre ?
(15:38) Roberts, Nigel: Jason, do you do this with radar as well for the same cases. would that give you extra information?
(15:40) Daniel Santos Muñoz: https://dtcenter.org/community-code/model-evaluation-tools-met
(15:42) Dmitrii Mironov: 👏
(15:42) Claude Fischer: Resume next session 15h55
(15:44) Claude Fischer: Reminder for the ML side meeting organized by Accord+Ewglam: tomorrow 16h-18h (on the Tuesday BJ link)
(16:18) Ana Mihalovic: Balazs, whwrw I can find writtenrecommendation about verification?
(16:23) Ana Mihalovic: where I can find written recommendation