3rd AtmoRep roadmap meeting

AtmoRep roadmap-meeting, 2024-11-04

General

• Participants: Nishant Kumar, David Greenberg, Martin Schultz, Asma Semcheddine, Michael Langguth, Julius Polz, Ankit Patnala, Kacper Nowak
• Start: 14:15, end: 14:45

Meeting notes

• Python environment for AtmoRep (Julius and Simon)
  • current issue: pip always installs latest package versions which may trigger incompatabilities
  • possible solution: fix Python package versions in setup
  • shared Python environment should be targeted (cetralized environment)
    • beneficial for the Hackathon
    • easier for onboarding of new partners
    • reduces inodes in the project
• HClimRep next to the Weather Generator (WG) project
  • most development will be steered from WG in the future
  • target: keep single code base
  • different development branches for projects
  • WG model and repo will be based on Christian's current repo for forecasting from satellite observations
    • presented e.g. at LSDLES workshop in Bonn
    • similaritites with requirement from WG project, i.e. fuse different (obs) data in joint latent space
    • skilfull forecasting for 2m temperature form satellite data demonstrated
    • ask Christian to give a presentation on OBS-based model (to be decided at the Hackathon in Cologne)
  • envisioned future model architecture for WG project
    • bunch of encoding networks for different datasets
    • fusion network/layers for joint latent space
    • forecast engine in the middle that operates data from joint latent space
    • random masking not suifficient: ask model to reconstruct specific data sources from others
    • research question: how to extend model to new datasets (fine-tuning or re-training)