-
Notifications
You must be signed in to change notification settings - Fork 9
2nd AtmoRep roadmap meeting
Michael Langguth edited this page Oct 7, 2024
·
1 revision
- Start: 14:20, end: 15:10
- Participants: Christian Lessig, Martin Schultz, Michael Langguth, Enxhi Krespha, Siddhant Agarwal, Nishant Kumar, Ilaria Luise, David Greenberg, Kacper Nowak, Asma Semcheddine, Michael Tarnawa
- datasets
- OceanRep status
- training data is prepared at AWI (zarr-output), models are currently training
- first coupled run expected in the next weeks (FESOM with ERA5 forcing)
- running at 1°deg, daily output
- architetcure/model configuration
- working at DKRZ
- 8 nodes for training
- 12 layers in encoder and decoder (by contrast to more light-weight current config with six layers)
- Kacper will share his specific config with Christian and Ilaria
- issues with memory consumption and masking
- other: total precipitation data is available in
sfc_data
in zarr
- OceanRep status
- downscaling
- current status:
- initial branch set-up, see here
- first integration of PerceiverIO-module and simple downscaling network
- current focus: get training to work
- prepare IMERG precipitation data
- adapt/use dataset sampler to read in IMERG data for training
- other notes
- in AIFS: no added value from high-resolved topography -> focus on getting training to run, not ingesting additional data
- partition between convective and large-scale precipitation beneficial? -> ask Paula
- test if output from intermediate layers is beneficial (down to encoder)
- first Multiformer prototype excl. total precipitation will be available by end of this week
- follow-up Multiformer incl. total precipitation one/two weeks later
- consider training/working on BSC (I have access)
- current status:
- forecasting/ roll-out
- Christian is implementing on forecasting (deterministic) performed in latent space
- first results by end of this week -> branch on github
- further clean-up afterwards
- get latent representation from all patches and merge into global latent representation
- do forecasting with global latent representation
- latent space representation includes at least three hours of data, but can be more
- can probably trained end-to-end (incl. core model) with the lightweight model architecture
- Nishant is working on forecasting with diffusion model, see this issue-branch
- Christian suggest to condition diffusion model with data from latent space, i.e. after encoder
- ideally: have a pre-trained, multi-purpose encoder-decoder and put a forecast engine (e.g. a diffusion model) in between working with latent representation -> would fit the concept of a Foundation Model
- more details to be discussed in a personal talk between David, Nishant, Christian and Ilaria
- Christian is implementing on forecasting (deterministic) performed in latent space
- Hyperparameter tuning (work by Nishant)
- paper on Tensor Programs V by Microsoft that could be implemented along with AtmoRep
- is orthogonal to other work
- be carful with resource usage, but is labelled as not requiring excessive resources
- JUREAP
- not on the lightening track -> AtmoRep is not highly scalable
- some scaling experiments required as an entry card to JUPITER
- benchmarks required, see issue-branch #32
- benchmark on the inference and training
- use Multiformer/Singleformer trained by Ilaria
- NIC Symposium paper
- JSC have been invited to submit a short paper
- draft with focus on Asma's data compression work and downscaling under preparation
- update discussion internally at JSC this Wednesday
- status update on parallel processing of AtmoRep output
- will also be discussed internally on Wednesday
The AtmoRep Collaboration - last update: April 2024