ICON Meeting 2022/3 (1 September 2022)

Venue

Hybrid: ETH Zurich (L 17.1) and via Zoom

Participants (on-site)

Michael Jähn (MJ), Annika Lauber (AL), Christina Schnadt Poberaj (CSP), Nikolai Ponomarev (NP), Athena Nghiem (AN), Boriana Chtirkova (BC), Andrea Stenke (AS), Sebastian Schemm (SS), Stefan Rüdisühli (SR), Brigitta Goger (BG)

Participants (Zoom)

Tamara Bandikova (TB), Emmanuele Russo (ER), Marco Arpagaus (MA), Michael Steiner (MS), Lionel Constantin (LC), Samuel Kellerhals (SK), Arash Hamzehloo (AH), Kseniya Ivanova (KI), Stephan Henne (SH), Guillhaume Bertoli (GB)

Minutes by Annika Lauber

Reports

C2SM (Michael Jähn, Annika Lauber, Tamara Bandikova)

MJ welcomes all participants to the meeting and presents C2SM ICON news together with AL.

AL talks about the ICON task about the stable RBF interpolation. SS asks where the RBF interpolation is used in the code. AL answers that the RBF interpolation is used for the interpolation from a grid point to a point not on the grid. KI asks what is meant with unstable. AL explains that unstable means the calculation has different results for every compiler and machine. KI wonders if this this part of ICON tools. AL clarifies that it is part of the actual ICON code and is being calculated during every run.

Brigitta Goger

BG tries to understand the turbulence representation better, especially the surface layer and turbulence scheme.

Boriana Chtirkova

BC is working with changes in the surface radiation.

Athena Nghiem

AN starts working with ICON-ART.

Andrea Stenke

AS is also working on the ICON-ART development. AS asks if anybody knows what the status of the ART license? CSP answers that she is in contact with Corinna Hose. The lawyers are currently looking into it. AS says that she is about to sign an ART license with Eawag.

Nikolai Ponomarev

NB will be working in Dominik Brunners group in Empa. He uses ICON-ART for simulation of Zürich with a 500m resolution. For now, NB is trying to get familiar with ICON and learning Python.

Michael Steiner

MS uses ICON-ART in limited area mode. He implemented an online approach to bring emissions to ICON.

Arash Hamzehloo

AH started working for Empa in January and is porting ART to GPU. He is currently running test cases to make sure he gets same results as on CPU. This is part of the HAMAM project.

Kseniya Ivanov

KI from Empa is coupling FLEXPART to ICON and also runs ICON-ART simulations.

Stephan Henne

SH from Empa is interested in turbulence in ICON. He is part of a project with MCH and the University of Frankfurt.

Guillhaume Bertoli

GB shows preliminary results for using machine learning to predict radiation with ICON. SS says that they work in collaboration with the SDSC. The offline performance is close to the performance to what is possible. Currently machine learning based radiation schemes are being linked to ICON. GC shows plot about memory vs. performance for the neural network and a random forest. The precision of the random forest is close to the precision of the neural network. For reaching such close results, the complexity of the random forest needs to be increased, which makes it too large to be used. SS says that the important lesson learnt is that the neural network consumes much less memory given a certain error tolerance. The memory aspect is also very import for EXCLAIM.

CSP asks if they intend to use it in later simulations and if there is any connection with EXCLAIM. SS says that it is part of the effort to accelerate the code. They want to use an aqua planet to see if it generalizes and to see if the original code and the neural network both predict radiation also in summer.

MA asks what the general motivation is. SS says that there are two motivations, one is the performance aspect and the other is that ecRAD is expensive. They could reduce the number of calls with ecRAD and speed up and improve it at the same time.

MA asks if it improves because of a higher resolution in time and space. SS says that the comparison of ecRAD vs. satellite data shows a larger error, than the comparison of machine learning against ecRAD.

AH asks how big the training algorithm is, if it is unsupervised and how general it is. SS says that training the random forest doesn’t take longer than a day and it takes several days to train the neural network on SSDs.

AH asks how they generalize the neural network, if they need to retrain from scratch or if they can train an already trained one. SS says that they are testing the generalization abilities. The quality of the results is relatively low for completely unseen data. Thus, it is not useful for climate simulation yet but for weather prediction. Retraining the neural network is not so expensive to hinder them from doing it, but the generalization is the big question mark.

AH asks if they are using 3D data for the neural network. SS says that they cannot afford to use 3D for long simulations, but one gains something by taking the neighbors into account by using convolutions.

BG asks if there will by a hybrid format for ICON or if it is planned to be used as a full scaled replacement for the ecRAD scheme. SS wants to use a hybrid version and not get rid of ecRAD in the sense of keeping the physics but accelerating it as much as possible. They will not replace the 3D part, which is the most expensive one.

Stefan Rüdisühli

SR is looking into time step effects of ICON vs. COSMO (same PDF as before). For testing, he uses an idealized convective system. He shows plots, which show time step sensitivity, especially for short time steps. It is still ongoing work. SS says that COSMO is more sensitive to different time steps. There is a paper from Axel Seifert called One Step at a Time.

Sebastian Schemm

SS is developing an aqua planet with a summer and winter hemisphere (still same PDF). The summer is imposed by radiation and SSTs. This produces a double ITCZ cell, which is seen as a bug in the model community. They figured out how to control the occurrence of a double ITCZ by changing SSTs. ER asks why they are not imposing seasonality. SS answers that the jet stream does something very unreasonable when doing that. CSP asks if shifting SSTs does not relate to different seasons. SS says that it would be late summer vs. early summer. Changing the radiation has almost no effect on double ITCZ.

SS says that for ICON v2.6.4 the sensitivity to compiler optimization is still very large and wonders if this may be caused by the unstable RBF interpolation.

Emmanuele Russo

ERs goal is to better understand heatwaves in Europe to enhance their seasonal to sub-seasonal predictability. Therefore, ER is looking at three things: 1) The dry dynamical core, 2) The aqua-planet configuration, 3) The amip configuration. ER brings the aqua planet closer to a land planet by modifying the heat capacity of mixed layer and moisture availability. There are problems with a mixed layer ocean in ICON. A bug was found in ICON v2.6.4 related to the initialization of heat fluxes. They fixed the bug and introduced new parameters (dmixsea, rmois) but the model is still very unstable. The model crashes as there are no horizontal ocean heat transport and no sea ice. They tested the same with a minimum threshold of 273.15 K and get a stable model with that.

Samuel Kellerhals

SK joined EXCLAIM in May and talks about the current state of the gt4py implementation. It allows to define high level stencil kernels in ICON. They are mainly working on porting the dry dynamical core. There are seven to ten stencils left out of 100. They hope to have it ready in the coming month.

Marco Arpagaus

MA talks about the last upgrade which promised to be much faster but is not. They can’t even compile with the new upgrade on ALPS. ICON still not as fast as COSMO and substantial slower on GPU. ICON simulations take twice as long as for COSMO.