EWGLAM Meeting 2021 - Day 2
szeptember 28, 2021
(8:55) Karoliina Hämäläinen (FMI): God Morning - Hyvää Huomanta :)
(8:55) Mike Bush: Good morning!
(8:57) Simon André: Good morning!
(8:59) Elena Astakhova, RHM: Morning!
(8:59) Anastasia Bundel: Доброе утро!
(9:00) Ekaterina Kurzeneva, FMI: Доброе утро! Huomenta!
(9:03) Dmitrii Mironov: Guten Morgen! Доброе утро! 🌞
(9:05) Aline Kraai (KNMI): Goedemorgen!
(9:16) Marco Arpagaus --- MeteoSwiss: Dumb question of an outsider: If I got that right yesterday, cy46 is the latest model version. You are using cy43. How old is this cycle, and why do you use such an old version (assuming it is "so old" ...)?
(9:20) Hannah Christensen: Thanks Clemens for your talk. A question about the new surface parameter perturbation scheme. Could you talk about the theoretical motivation for the scheme? As the parameters you are stochastically perturbing sound as if they represent properties that be constant over the duration of the forecast, in which case a fixed parameter perturbation approach could be more appropriate.
(9:22) Hannah Christensen: Thanks Clemens - my misunderstanding!
(9:24) Clemens Wastl: Yes, in fact it is a parameter perturbation scheme. you take the parameters from the surface assimilation and perturb them stochastically. this procedure is repeated with every assimilation cycle
(9:41) Marco Arpagaus --- MeteoSwiss: very interesting talk, Anne, thanks a lot! - I would have two questions ... ;-)
(9:42) Martin Leutbecher: Why should initial perturbations scale up to larger scales & lead times in a LAM?
(9:42) Dmitrii Mironov: Anne, thanks for the talk. Could you briefly explain what (which variables) are perturbed in the boundary layer  and how. 
(9:43) Karoliina Hämäläinen (FMI): 👏
(9:46) Henrik Feddersen: Thanks for an interesting talk! Could you simply inflate the initial perturbations and thereby increase the spread from the beginning of the forecast?
(9:47) Markku Kangas: Thanks for a great and inetersting talk!
(9:48) Marco Arpagaus --- MeteoSwiss: Sorry for taking up too much discussion time. - And thanks again for an excellent talk.
(9:54) Mccabe, Anne: In reply to Dmitrii, the perturbations in the boundary layer are to temperature and moisture.  They are applied over an 8x8 grid and are correlated in time.
(9:55) Dmitrii Mironov: Anne, how are the perturbations specified? Any physical reasoning behind them? I mean, what processes are those perturbations meant to describe?
(9:55) Mccabe, Anne: In reply to Henrik, I think it would be interesting to do some experiments like you suggest.  Ideally we would like to set the initial conditions up using a DA system that balances the spread and error at the beginning of the forecast (I say this with only a limited understanding of DA systems).
(9:57) Mccabe, Anne: In reply to Martin, do you think I shouldn't have this expectation?  We have see this upscale for systems that come from the south of the UK and tend to have more energy.  I'd be interested in hearing your thoughts more on what would be reasonable to expect.  Thanks for you comment.
(10:07) Mccabe, Anne: Dmitrii, the motivation for the perturbations is that an update to the boundary layer scheme a few years ago resulted in the model struggling to initiate convection from the sub-grid scale which resulted in a reduction in the number of small scale showers and some missed heavy storms.   The perturbations are random with the magnitude linked to the surface heat flux.  They are only applied where cumulus is diagnosed.
(10:08) Marco Arpagaus --- MeteoSwiss: Thanks, Inger-Lise, nice talk!
(10:08) Markku Kangas: 👏
(10:09) bazile: !
(10:09) Mccabe, Anne: 👏
(10:11) Dmitrii Mironov: Thanks, Anne. Linking the perturbations to the surface flux(es) sounds interesting. Do you have something written on this point, e.g. a short description with a few formulae? 
(10:15) bazile: Thanks Inger-lise, for the single precision you have some crashes .. where in the forecast  and for some members or for all the members ? The reduction of CPU for the full system is about 30% ?  
(10:19) Inger-Lise Frogner: The forecast model runs 30% faster. We do not know so much on the cause of the crashes yet, but it does happen in individual members. 
(10:21) Mccabe, Anne: Dmitrii, can you access this link?  If so, it's section 6. https://code.metoffice.gov.uk/doc/um/latest/papers/umdp_081.pdf
(10:24) Dmitrii Mironov: Anne, thanks, but I seem to need a pass to access your MO pages. Could you send me that section 6 pdf by email? dmitrii.mironov@dwd.de 
(10:25) bazile: Thanks Inger-lise, Single precision is only used in the forecast model ?
(10:25) Inger-Lise Frogner: Yes.
(10:25) Anastasia Bundel: Just a comment: What the Israelian colleagues found that we can give a forecast of intense precipitation when 10% of the members predict it reflects the fact that our model underestimated more rare intense precipitation events. It reflects the usefulness of the ensemble as it can capture such situations better.
(10:26) Mccabe, Anne: Dmitrii, there are details in this paper: https://journals.ametsoc.org/view/journals/atsc/78/3/JAS-D-19-0291.1.xml The scheme that is operational is referred to as the Lock scheme in and is described in section 5.  This paper is a more physically based scheme that we will consider using in the future.
(10:29) Martin Leutbecher: @Anne, yes I think that continued upscale perturbation growth due to initial perturturbations only is an unrealistic expectation for small model domains as the ensemble is very much constrained by the lateral boundary conditions.
(10:29) Matthias Raschendorfer: I have a comment to the turbulent BL perturbatiion.
(10:31) Roberts, Nigel: Chiara, I like your chessboard, we should do that!
(10:31) Mccabe, Anne: Chiara, I really like your checkerboard approach to guide forecasters.  I'd be interested to hear the forecaster feedback.
(10:32) Marion Mittermaier: Agree. Like the checkerboard approach as well. 
(10:32) Roberts, Nigel: Do the random perts really shift the precipitation positioning systematically. Do you know why? Maybe I didn't understand.
(10:33) Mccabe, Anne: In reply to Martin, thank you.  One of the things we wanted to test was whether a larger domain would result in more upscale of the IC perturbations.
(10:37) Dmitrii Mironov: Anne, I downloaded the JAS paper. Thank you. 
(10:37) Balazs Szintai - C-SRNWP: we resume at 10:50 CEST
(11:12) Susanna Hagelin (SMHI): Do you get any jumpiness in the IMPROVER forecast when e.g. you add a new global ensemble cycle?
(11:15) Hannah Christensen: Thanks for your talk, Nigel. How do you optimise the blending weights? Would there be scope for changing these dynamically - e.g. if you know the global or LAM ensemble is better for some forecast situations than others?
(11:19) Roberts, Nigel: Hi Hannah, we optimise at short range from verification results. At longer range it is a pragmatic transition to get from one model to another. Yes it would be good to change dynamically in principle - although forecasters might get nervous if the weights change day to day.
(11:24) Hannah Christensen: Thanks!
(11:25) Xiaohua Yang: @nigel, does IMPROVER make direct use of observation infomation (cloud, precipitation)  at the nowcasting range?
(11:30) Gabriella Szepszo (OMSZ): @Nigel: how often do you update the IMPROVER forecasts? I mean e.g. in every 15 mins in the nowcasting time range, hourly up to 12 hours etc.?
(11:36) Roberts, Nigel: Xiohua, only precipitation via optical flow extrapolation. Seperate work on nowcasting may provide obs-based nowcasts IMPROVER can use in future. 
(11:37) Xiaohua Yang: Thank you @nigel
(11:39) Roberts, Nigel: Gabriella, it's every 15 minutes for the blend out to around 4h ahead (but only output hourl at present) and then it's every hour out to 5 days and then every 6 hours for forecasts beyond 5 days.
(11:39) Ekaterina Kurzeneva, FMI: Thank you @Anastasia
(11:39) Gabriella Szepszo (OMSZ): Thanks!
(11:48) Marion Mittermaier: Martina, this is interesting. Is there any final report or publication explaining this work?
(11:58) Balazs Szintai - C-SRNWP: we resume at 13:20 CEST
(13:36) Carl Fortelius: Yuji: could you give the reference for modifying the surface heat flux in stable conditions, thank you?
(13:37) Yuji Kitamura: We can find the reference at https://journals.ametsoc.org/view/journals/atsc/77/8/jasD190255.xml
(13:38) Carl Fortelius: Thank you!
(13:38) Kristian Pagh Nielsen: 👍
(13:38) Ekaterina Kurzeneva, FMI: thank you!
(13:42) Roger Randriamampianina: Your system is good to evaluate the accuracy of AROME coupled with different hor. resolutions. Can you comment on its performance with different res?
(13:43) Ekaterina Kurzeneva, FMI: What is the resolution of physiography, which you use for 1.3 km resolution of AROME?
(13:45) Christoph Gebhardt (DWD): Ghislain, maybe I missed it, sorry. When will the suite with identical resolution for EPS and det be operational?
(13:59) Ghislain Faure: You are right, I forgot to mention the schedule of the e-suite: it should be in operations by mid 2022. It is a longer e-suite than usual due to the amount of changes.
(14:03) Ghislain Faure: We use orography from GMTED2010 at 7.5- arc-second (~ 250m), and Ecoclimap I for other physiography fields (which is at 1 km resolution if I remember well).
(14:04) Ekaterina Kurzeneva, FMI: Thank you!
(14:11) Patrick Samuelsson: @Jurgen: Do you assimilate any other snow control variables in the new snow scheme than snow depth?
(14:16) Gianpaolo Balsamo: @Jurgen: great presentation! Do you expect also NWP impact from the vegetation photosynthesis revision (the two big leaf)?
(14:16) Jeanette Onvlee: @Jurgen: can you say more about the new urban canopy parameters which you mentioned would be provided?
(14:17) Souhail Boussetta (ECMWF): @Jurgen any initial results related to the change of the Jarvis approach ?an what is the motive 
(14:19) Marco Arpagaus --- MeteoSwiss: @Patrick: We currently do not use a snow assimilation scheme at all: The free running SNOWPOLINO performs better than the current snow assimilation scheme. 
(14:19) Gianpaolo Balsamo: Thanks a lot and congrats on the snow multilayer work
(14:21) Patrick Samuelsson: Thanks Marco!
(14:25) Ekaterina Kurzeneva, FMI: The basic resolution of ESA-CCI is 300m. You project it for the coarser resolution first?
(14:25) Jürgen Helmert (DWD): @Jeanette Onvlee: For details about the urban canopy parameters I would like to refer to Jan-Peter Schulz as principal investigator of CITTA.
(14:33) Jürgen Helmert (DWD): @Souhail Boussetta: The problem is that the LAI does not respond to water stress. Moreover, the “one big-leaf” approach has a disadvantage which is related to the impossibility of accounting for the difference of the physiological properties between sunlit and shaded leaves (Dai et al., 2004). In order to overcome the limitations, it was decided to change the empirical Jarvis approach with the physically based Ball-Berry (Ball, 1988; Ball and Berry, 1991) approach coupling with photosynthesis (Farquhar et al. 1980 and Collatz et al., 1991 models for C3 and C4 plants) and introduced a “two-big leaf” canopy (Thornton and Zimmermann, 2007).
(14:34) Dmitrii Mironov: Gianpaolo, in your 5-layer snow scheme, how deep is the uppermost layer? related question: do you have an explicit or implicit coupling of he snow temperature equation to the temperature equation in the atmospheric surface layer? I assume you use tile approach and allow fractional snow cover.   
(14:34) Ekaterina Kurzeneva, FMI: What will be the soil depth for the 10-layer model? 
(14:36) Souhail Boussetta (ECMWF): Thanks @Jurgen, that's a substantial change toward more physical representation and I guess the next step would be to adjust parameters for better NWP impact
(14:36) Matthias Raschendorfer: @Gianpaolo and Jürgen: Actually the current treatment of plants in COSMO and ICON is not single-leafe approach, as the leave-temperature just equals the over-all surface-temperature at the top of the soil. Apart from the mentioned initiative, there is already a canpy scheme being implemented into ICON, treating roughness-elements (like plants) as an additional semi-transparent cover-layer, which has a rather strong beneficial impact particularly to the diurnal cycle of T2m and Td2m.
(14:38) Souhail Boussetta (ECMWF): @Ekaterina initial testing are for a 10m depth
(14:40) Ekaterina Kurzeneva, FMI: @Souhail, thank you!
(14:47) Souhail Boussetta (ECMWF): @Matthias Thanks for the clarification, can you give some more details or a reference for the canopy model that is  "treating roughness-elements (like plants) as an additional semi-transparent cover-layer"
(14:51) Gianpaolo Balsamo: @Dmitrii: The top layer of the ML snow scheme is 7.5 cm. In the original paper of Arduini et al. 2019 we tried 5cm, but during the operationalisation we found better numerical properties with a slightly bigger first layer (less oscillation in 4d-var which is very beneficial).
(14:55) Gianpaolo Balsamo: @Ekaterina: Souhail has replied, just wish to add that we separate the depth for hydrology (kept to 3m to avoid unrealistic storage) and the depth for thermal diffusion (the 10m there is beneficial for the annual/seasonal cycle of soil temperature). Adding more layer can be important for Data Assimilation and with @Patricia we are keen to test this further to see if beneficial.
(14:58) Ekaterina Kurzeneva, FMI: @Gianpaolo, thank you! For us, it is very important, because we are also moving to multi-layer and deep soil. And for LAM, we usually initialize the cold start from ECMWF fields. For us it is important, how deep you are ... and another question, how soon we can expect this deep soil will become operational at ECMWF?
(15:02) Dmitrii Mironov: Patrick, looks like many ocean model are being used/coupled... what is the ACCORD general stratey along this line?
(15:04) Dmitrii Mironov: Gianpaolo, thanks for your answer. Please answer my 2nd question (in the same chat message I posted). Thanks in advance. 
(15:06) Martin Leutbecher: @Patrick, you mention that the change in physiography requires tidy work on tuning. Does this point to uncertainties that should be represented in ensembles?
(15:08) Gianpaolo Balsamo: @Matthias thanks for the comment, we would be interested in the Canopy Cover layer. Is there a report or a paper we can read? Many thanks in advance
(15:09) Jan-Peter Schulz: @Martin: Yes. There you are also in line with Chiara.
(15:11) Gabriele Arduini: Hello @Dmitrii, regarding your questions on the coupling of the snow scheme, the coupling of the snow temperature eq to the surface energy balance is with an "explicit flux", and a snow cover fraction parametrization is still used using a relatively simple parametrization based on snow depth
(15:11) Patrick Samuelsson: @Martin, partly yes but also no beacuse long (say 1 month validation runs) should still show good statistics in my mind with respect to e.g. wind and we should at least be close to observations over such a time period and not rely on the EPS perturbations to "solve" that...
(15:14) Jan-Peter Schulz: @Patrick: You showed a lot about SURFEX and all the developments around. Are the HIRLAM formulations meanwhile somehow "merged" into this, or is this still going on, or will it (partly) not happen? I remember that you worked a lot on e.g. snow and vegetation, with good results.
(15:15) Dmitrii Mironov: Thanks, Gianpaolo. With the explicit coupling, a 5 cm deep uppermost layer is likely a good choice. I feel a more advanced  treatment of snow fraction  could do good to our models, but we still do not know now to do it right, I am afraid. 😕
(15:16) Patrick Samuelsson: @Jan-Peter: Yes, in collaboration with SURFEX collegues the HIRLAM snow-vegetation work is now availbale in SURFEX as Multi-Energy Balance by Boone et al.
(15:18) Patricia de Rosnay: Ekaterina, very surprised to hear about the physiography inconsistency between model and DA. Beyond spurious rejections as you mentioned, does it cause spurious increments?
(15:19) Jan-Peter Schulz: @Patrick: Thanks.
(15:20) Gianpaolo Balsamo: @Thanks Dmitrii, @Gabriele has snow fraction on his radar :-) for next. it will be nice to chat about that (there seems to be no easy choice). He also thinks about application of snow ML on cryosphere (frozen-lakes and sea-ice).
(15:23) Jan-Peter Schulz: @Gianpaolo: I am surprised that you did not need to do this "tidy" work of retuning after introducing a new land use classification. The BATS-based was relatively outdated, or? Therefore, I would expect many difference in land use, which you also mentioned, and therefore also in the model results. You showed annual verifications, what about seasons, or different regions?
(15:23) Stefan Schneider (ZAMG): Helga Toth is working in Hungary
(15:23) Helga Toth: :)
(15:23) Dmitrii Mironov: Gianpaolo, having cht(s) is a good idea. Let's meet at breakout session first. We can also have a phone chat (I am at the office very day). I am nt sure I have any constructive ideas as to the fractional snow cover, but I can share possibly useful thought as the the snow over lake/ocean ice.
(15:23) Patrick Samuelsson: @Patricia: Katya can explain more but, yes, we have experienced problems in snow analysis close to water where bad increments destroy the snow evolution if not looked into.
(15:29) Gianpaolo Balsamo: @Jan-Peter: you are right the vegetation calibration is just starting as is planned to last for 1 more year. I agree this is a major work. Happy to discuss further. @Souhail is looking after this development.
(15:34) Gianpaolo Balsamo: @Ekaterina: impressive work. Interesting also the experience with the simple ice scheme and assimilation. Which observations did you assimilate to constrain ice temperature?
(15:34) Ekaterina Kurzeneva, FMI: @Patricia: yes, it leads to spurious increments, and this is the main problem. As a results, we have serious problems in operational runs in spring, because errors accumulate.
(15:37) Ekaterina Kurzeneva, FMI: @Gianpaolo: observations from VIIRS, from OSI-SAF product. Yurii Batrack can provide more details: yuriib@met.no.
(15:37) Gianpaolo Balsamo: Thanks 👍
(15:39) Ekaterina Kurzeneva, FMI: @Helga: sorry! of course!
(15:39) Ekaterina Kurzeneva, FMI: :)
(15:48) Roberts, Nigel: Given the recent record heat wave in Canada, do you have to b cautiuos about range checks?
(15:48) Roger Randriamampianina: @Matt: did you test mainly Titan to quality control Temperature data? Thanks for the very good overview
(15:48) Patrick Samuelsson: @Matt: You think it is possible, or is a potential, to inspire more people to connect their weather staions to WOW 
(15:48) Karoliina Hämäläinen (FMI): Do you think some parameters are more reliable than the others?
(15:49) Dow, Gareth: Did you consider any other variables eg RH, Wind, Precip ?
(15:49) Roberts, Nigel: thanks Matt
(15:49) Marion Mittermaier: Very nice Matt.... having this QC control software sounds like a large step forward. I would say though that DA is more resilient to QC issues. Verification is very unforgiving, which is why we use DA as a QC control mechanism. 
(15:50) Markku Kangas: Excellent and clear presentation! Thanks 🙂
(15:52) Roger Randriamampianina: @Matt: at MET Norway we use it also for precipitation, as far as I'm aware. If you are interested in QC of wind data we have QCwind at KNMI. 
(15:53) Florian Meier: In Austria we did some tests with Netatmo and Titan: wind was quite OK, but MSLP is very sensitive to errors in station height.
(15:54) Roger Randriamampianina: Thanks @Florian. Very interesting. 
(15:54) Marco Arpagaus --- MeteoSwiss: @Marion: Can you let me know what QC your DA is doing? - We always feel we are more reluctant to use fishy obs for DA than for verification ... ;-)
(15:56) Balazs Szintai - C-SRNWP: ML side meeting starts here at 16:00 CEST
(15:56) Jan-Peter Schulz: Do you also save the chat?
(15:56) Matt Nagle: @Roger: Ah thank you for the clarification, I wasn't aware that Met Norway was also using TITAN for precipitation, that is very interesting!
(15:56) Matt Nagle: Thanks everyone for the questions!
(16:16) Bogdan Bochenek: Can you share screen with quaestions?
(16:19) Bogdan Bochenek: question
(16:26) Jeanette Onvlee: question
(16:27) bazile: ?
(16:52) martina tudor: can we use the model with existing phy params to generate ml version to emulate tangent linear and adjoint of the code?
does this needs to be repeated for each change in domain, rresolution, etc?
(16:53) Axel Seifert: Martina, maybe you want to take a look at this paper on deep emulators  https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2021MS002554
(16:57) Lesley De Cruz: Once you have the deep learning emulator of a parametrization scheme then I suppose you can compute TL/AD exactly/ analytically?
(16:58) Peter Dueben: https://doi.org/10.1029/2021MS002521
(16:59) Lesley De Cruz: Thank you.
(17:04) Magnus Lindskog: question
(17:24) Phillip Scheffknecht: Question. Since I don't have a microphone, I'll put it here:
Considering that the metric for performance will probably be forecast quality, will the entire NWP be run for every training cycle, depending on the use case? And is this even feasible?
(17:26) Phillip Scheffknecht: I could go on my phone, just a moment
(17:31) martina tudor: can you have numerical instability when using ml emulated phy schemes?
(17:31) Axel Seifert: Some of these topics are discussed here https://raspstephan.github.io/blog/optimization-dichotomy/
(17:35) martina tudor: thank you for the link
(17:48) bazile: the instabilities are not linked with the potantial inconsistency of the ML for some conservative variables ?
(17:56) Magnus Lindskog: Do you see a risk/worry that you introduce correlations of model errors and observation errros when applying a ML based obop?