README.md

metadata

title: Samudra 2
emoji: 🌊
colorFrom: green
colorTo: gray
sdk: gradio
sdk_version: 6.25.0
app_file: app.py
python_version: '3.12'
short_description: Roll a neural ocean emulator forward against GFDL-OM4
startup_duration_timeout: 30m
models:
  - M2LInES/Samudra2
tags:
  - ocean
  - climate
  - emulation
  - weather
license: apache-2.0

🌊 Samudra 2 — global ocean emulator

Interactive rollouts of M2LInES/Samudra2, an autoregressive neural emulator of the global ocean.

The 1° checkpoint is a ConvNeXt U-Net (84 M parameters) that maps the ocean state at two consecutive 5-day snapshots — potential temperature, salinity, zonal/meridional velocity over 19 depth levels, plus sea-surface height (77 fields → 154 channels) — together with surface forcing (wind stress, heat flux and its day-of-year anomaly) to the next two 5-day snapshots. One model step therefore advances the ocean by 10 days, and steps are chained autoregressively.

What this Space does

1. Streams a real GFDL-OM4 initial condition and surface forcing from the public 1° store on the NYU OSN pod (m2lines-pubs/Samudra/v2026-07/om4_onedeg), pulling only the handful of chunks a rollout needs.
2. Runs the released EMA checkpoint forward on ZeroGPU for up to 36 steps (~1 year).
3. Animates the emulated field next to the OM4 ground truth and the error field, and scores the rollout (area-weighted RMSE) against a persistence baseline.

Initial conditions are restricted to the held-out inference window used upstream (2014-10-10 → 2022-12-24), so nothing shown here was seen during training.

Normalization, masking, channel ordering and the derived hfds_anomalies forcing channel follow samudra.datasets.InferenceDataset in m2lines/Samudra. The heat-flux day-of-year climatology needed for that channel is precomputed from the same public store and shipped here as hfds_anom_stats.npz.

Credits

• Model & code: the Samudra authors (M²LInES) — CC-BY-4.0 weights, Apache-2.0 code.
• Paper: Samudra 2: Scaling Ocean Emulators across Resolutions, arXiv:2606.02610.
• Data: GFDL-OM4 preprocessed by M²LInES, public-read on the NYU OSN pod.