Skip to content

Milestones

List view

  • Milestone: Multimodal capabilities ensembled. UX research completed, findings presented, and white paper drafted. Suggested deliverables: - We aim to start developing a benchmark platform for comparing existing models if we have additional REG support. - Finalise partners who can operationalise this work - Deliver a modular multimodal data ingestion framework - Deliver a prototype of an ensemble learning technique for integration to the pipeline (potentially Anemoi has this feature by adding nth dimension). Using CRPS loss function - User Research completed, present findings to drive product outputs. (wait for the MoU in placed) - smarter encoder (multimodal capability without preprocessing) if EASE5 works

    Due by September 25, 2026
  • Milestone: Systems adapted to intake local regional datasets for finetuned precise forecasts. We will design product output formats per UX research recommendations. Suggested deliverables: - Create forecasts for applications requiring sea ice thickness, sea ice motion – for navigation, migration - Set training schedule to combat data drift (Automation) - Create one framework with Aardvark to cover weather, sea ice, and ocean - Final report and next steps

    Due by February 29, 2028
  • Milestone: Full infrastructure handling multi data sources and multi models with multi predictions capability (MIMO). Manuscripts drafted. Suggested deliverables: - Raw data observation (need to determine common format, or who would be the sample case study / actual use case) - Technical paper on the system - Collaborators identified on who can provide long term maintenance

    Due by February 26, 2027
    0/1 issues closed
  • Milestone: Intelligent handling of non-gridded data, models beat persistence, manuscript outline drafted, work to integrate diffusion model Suggested deliverables: - Integrate masks and data reprojection, potentially consider other methods to improve models (James' list if these are sufficient to beat persistence) - Krigging or another method beyond nearest neighbour gridding - make the diffusion model autoregressive - outline of manuscript - Diffusion on Multimodal pipeline Stretch goals: - Looking at a way to integrate the diffusion model into pipeline by separating encoder / decoder from processor training - Work on the documentation

    Due by June 26, 2026
    1/13 issues closed