Samudra

Benchmark Results

Samudra is a global ocean emulator described in "Samudra: An AI Global Ocean Emulator for Climate" and updated in "Samudra 2: Scaling Ocean Emulators across Resolutions". Samudra efficiently emulates the ocean component of a state-of-the-art climate model, accurately reproducing key ocean variables including sea surface height, horizontal velocities, temperature, and salinity, across their full depth.

Ground truth (left) vs. Samudra 2 prediction (right) for sea surface temperature in the tropical Pacific.

We are actively and openly developing this emulator to support new tasks and data sources with the goal of building a broadly useful foundation model for ocean and climate. Please see our docs for more or our contributing guide to join in!

How to run

To configure, train, validate, and visualize Samudra 2 in a browser on public data with a Colab GPU, open the free-tier Colab quickstart.

Given uv is installed:

# Launch a training run with the default Samudra configuration
uvx samudra train samudra_om4/train.yaml --data @data/om4_demo.yaml

# Run a long autoregressive rollout against ground-truth data, collecting metrics.
uvx samudra eval samudra_om4/eval.yaml --data @data/om4_demo.yaml

# Generate maps, time series, and probability density plots from eval outputs.
uvx samudra viz samudra_om4/viz.yaml --data @data/om4_demo.yaml

To learn more, check out the installation and quick start guides.