CryoCast is a multimodal machine-learning framework for sea-ice forecasting. It combines satellite observations, Argo float sensor data, and ERA5 reanalysis fields to produce short-term Arctic and Antarctic sea-ice concentration forecasts.

Example forecasts are shown below.

Arctic

Antarctic

The encode-process-decode architecture translates each input dataset into a shared latent space, allowing new data sources and model components to be added without changing the full pipeline.

Key capabilities

• Multimodal data fusion across satellite observations, reanalysis fields, and in-situ sensor data.
• Extensible encode-process-decode design for adding new input sources and prediction targets.
• Multiple model configurations, including UNet, vision transformer, diffusion, and persistence-baseline approaches.

Project context

CryoCast is developed at The Alan Turing Institute as a research system for Arctic and Antarctic sea-ice forecasting. The current codebase supports research, benchmarking, sensitivity experiments, and case-study analysis; it is not intended for operational forecasting, safety-critical decision making, or public warnings. See the model card for intended use and scope.

Quick start

git clone git@github.com:alan-turing-institute/cryocast.git
cd cryocast
uv sync --managed-python

Create a local config in cryocast/config/ (see Configuration for details):

# cryocast/config/my.local.yaml
defaults:
  - base
  - _self_

base_path: /path/to/my/data

Then download datasets and train:

uv run cryocast datasets create --config-name my.local
uv run cryocast train --config-name my.local

Evaluate a checkpoint:

uv run cryocast evaluate --checkpoint /path/to/checkpoint.ckpt --config-name my.local

Documentation

See the project documentation for the full user guide and reference material.

• Installation for prerequisites, uv setup, and HPC-specific steps
• Configuration for local config files, model overrides, and custom datasets
• Commands for datasets create, datasets inspect, train, and evaluate
• Add a model for tensor format and model architecture guidance

Jupyter notebooks

The notebooks/ folder contains demonstrator notebooks. Run them with:

uv run --group notebooks jupyter notebook

Start with notebooks/demo_pipeline.ipynb for a worked example of the full pipeline.

Contributing

See CONTRIBUTING.md for the development workflow, coding conventions, and test instructions. For project questions, contact SeaIce@turing.ac.uk.

License

CryoCast is released under the MIT License.