8th AtmoRep core developer meeting

Date: 16.12.2024

Meeting notes

• work by Simon and Kacper on the configuration of AtmoRep
  • refactored configuration hierarchy introduced
    • all in one class
    • different distinct configuration objects created incl. doc-strings
    • objects include serialization and deserialization code
  • compatability to old code is maintained via ConfigFacade-object
    • allows iteratively updating the existing code to new configuration
  • further work to fully support compatability of configuration are still outstanding -> merge request will be done then
  • steps for adding new parameters to config:
    • add option to config.py, add way to serialize and deserailize (three lines of code required)
    • need to ensure proper (de-)serialization (by hand) -> test will check if new parameters are correctly introduced
    • hydra-package would do ensure (de-)serialization automatically, but not used
• work by Ankit and Michael on precipitation downscaling
  • question on temporal length of downscaling window
    • could either be the central subset of the 36h hour input data sequence or the whole temporal sequence
    • should be a configurable parameter of the downscaling application
    • Christian mentions that it would be an interesting ablation study to investigate the sensitivity to the temporal length of the downscaling window (was 3h or 6h in the 1st AtmoRep paper)
    • Martin mentions that we should have an eye on consistency on time sequences used (e.g. 6h for short-range forecasting, downscaling and temporal interpolation)
    • for now: go ahead with central time window (3h or 6h)
• scope of paper (all)
  • should the paper be a follow-up of the 1st AtmoRep paper
    • autoregressive rollout
    • AtmoRep as temporal interpolator or (de-)compression tool
    • advanced downscaling application
  • ...or an replacement of the 1st paper
    • would also include demonstration of zero-shot capacities
  • to be discusses in a separate meeting
• processing AtmoRep output (Michael T.)
  • current output format is hard to handle with Dask -> inefficient in terms of speed and memory
  • recent efforts to identify more proper output structure (create example zarr data store) -> adaptations to data writer required in the future
  • solution prototype targeted for mid of January
• autoregressive rollout (Nishant)
  • approach with GNN to project into latent space -> (from literature)
  • modeling of latent space with ELBO
  • instead of ELBO for approximating the latent space of GNN, use generative models like Normalizing Flows or Diffusion Models to model the accurate data distribution and not an approximate one
  • make such a generative model autoregressive for long rollouts that models climate (years or decades in future) and not weather that models only next few days
  • next step: implement generative models for climate application
• update by Kacper
  • AtmoRep presented at AGU: not sp much attention, but poster session was at the end of the conference
  • DSSIM as new metric for evaluation that is supposed to be context-aware; similar to SSIM
  • can incorporate physical constraints
  • can be used for evaluation and as loss