ICON Meeting 2023/3 (26 September 2023)

Venue

Hybrid: ETH Zurich (L 17.1) and via Zoom

Participants (on-site)

Michael Jähn (MJ), Jonas Jucker (JJ), Annika Lauber (AL), Fulden Batibeniz (FB), Clarissa Kroll (ClK), Jacopo Canton (JC), Doris Folini (DF), Emmanuelle Russo (ER), Stefan Rüdisühli (SR), Guillaume Bertoli (GB), Jan Zibell (JZ), Nadja Omanovic (NO), Andrea Stenke (AS)

Participants (Zoom)

Matthieu Leclair (ML), Arash Hamzehloo (AH), Corina Keller (CoK), Marco Arpagaus (MA), Brigitta Goger (BG), Katrin Ehlert (KE), Mikael Stellio (MS), Chia Rui Ong (CRO), William Saywer (WS)

Minutes by Annika Lauber

Reports

C2SM (Michael Jähn, Annika Lauber, Jonas Jucker, Matthieu Leclair)

MJ welcomes all the participants to the meeting and presents the news from C2SM about ICON.

Discussion about presented results of the speed-up of the two-moment microphysics scheme after EuroHack23

MA asks about the difference between the implicit and the explicit scheme. AL replies that the implicit scheme depends on the state of the next time step and the current time, while the explicit scheme does not.

WS asks if it is a fair to compare the timing of the explicit scheme with the implicit scheme. He reports on behalf of Anurag Dipankar that he is concerned about stability issues with the explicit scheme.

JJ replies that these are preliminary results and it’s not a completely fair comparison. He suggests doing some academic work, such as a master’s thesis, to compare the explicit and implicit schemes.

Fulden Batibeniz

FB started to work for the University of Bern, which wants to start using ICON. She is doing regional climate modeling and plans to use the ECMWF seasonal forecast for extreme events running with many ensembles.

Clarissa Kroll

ClK is working with R2B10, running it on an uncoupled setting.

Jakopo Canton

JC doesn’t use ICON yet, but wants to start with another LES code and run ICON at a very high solution.

Doris Folini

DF reports on behalf of Boriana Chtirkova who is using ICON with ECHAM physics. She is looking at SST patterns and their effect on the internal variability of other variables.

Emmanuelle Russo

ER uses ICON with ECHAM physics to study the drivers of heat waves.

Stefan Rüdisühli

SR is helping GB solve a 300 K warming bias in their machine learning algorithm. There are inconsistencies in the training data, so they are trying to get better training data.

Guillaume Bertoli

GB is working on two new tools for machine learning emulation of radiation in ICON. There is a huge temperature error at the top when switching to the neural network. This may be due to the pressure intervals at the surface and top layers, which is 10^4 higher at the top. ICON goes so high in the atmosphere because it needs a damping layer.

Jan Zibell

JZ is working with an aquaplanet ensemble to study storm tracks and the role of latent heat release.

Nadja Omanovic

NO uses the two-moment microphysics scheme. The goal is to run ICON in LES mode to improve the precipitation forecast.

WS asks if it’s the same scheme as presented by the C2SM core team. NO confirms that she is using the same scheme.

Andrea Stenke

AS uses ICON-ART to simulate the atmospheric cycling of various trace elements. She is working on aersol uptake. Unfortunately, the ART documentation is not very comprehensive. Progress is slow as there are only two people working on it.

Corina Keller

CoK works at Empa. In their project they are using ICON-ART to simulate air quality in Europe and Switzerland. They run full chemistry simulations to capture the formation of ozone or other air pollutants.

Arash Hamzehloo

AH is working on porting ICON-ART to GPUs. They are almost done with one of the submodules for modeling immersed gases like methane. They want to make the code GPU ready for more complex parts of the code.

Katrin Ehlert

KE has just started as scientific coordinator for numerical weather prediction at MeteoSwiss.

Mikael Stellio

MS is working on optimizing ICON-HAM for GPUs. A missing interface with ICON-HAM and rtmp (radiation scheme) is now included and ported. The next step is to test ICON-HAM with different resolutions.

Marco Arpagaus

MA reports on the status of ICON-22, which is replacing the numerical weather prediction engines. They are reaching the milestone to go pre-operational next Monday. They are testing if the system is technically robust on the new machine (not yet convinced). ICON is proving to be a good model so far. It can consistently improve the score for wind, clouds, precipitation, and temperature. However, there are seasons where ICON is consistently worse. There is a systematic bias in 2m temperature and global radiation. So not everything gets better. Overall, MA is confident that ICON will be better than COSMO.

ER asks if the biases in temperature and radiation are positive. MA replies that yes, it’s too warm, especially in the valley, by 1 K or more. Probably the cooling at night is not efficient enough. The bias is not present on mountain tops. There is a strong bias in the radiation in the morning and evening. Wind is too low on mountain tops and in valleys. MA can send plots on request. ER says that in ICON-CLM there are some biases in temperature and radiation that are counterintuitive and wonders if the problems are related.

William Sawyer

WS works for CSCS and is supposed to keep track of what is going on in the ICON world. He mentions two other ICON-related works at EuroHack23 (besides the two-moment microphysics scheme). One of the projects is about the ocean, which had a lot of success. In fact, it looks so good that WS may propose to run ICON in a coupled ocean-atmosphere version. The other project was on ICON-ART where they ported parts of the atmospheric chemistry to GPUs. The results were very good, but the code contained Fortran 2008 features that are difficult to port to GPUs. They had a speedup that was 20 times better than the CPU speedup that WS was working on.

WS also reports on the EXCLAIM effort on behalf of Anurag Dipankar. Anurag Dipankar tries to run the first version with R2B9, but the model explodes after a few months. The ALPS platform should be fully operational around February next year. First tests with ICON on ALPS look very good performance-wise, but moving to a new machine is always a bit dramatic. Users will be supported by C2SM and CSCS.

Brigitta Goger

BG is interested in the mountain boundary layer. She uses high resolution over the Alps down to 500m/250m using both turbulence schemes. BG is testing whether higher resolutions improve the model. It seems that some things are better, but others are not.

Chia Rui Ong

CRO is porting the graupel scheme to GT4Py.