IceNet Multimodal Pipeline

IceNet-MP is an AI/ML framework for multimodal sea-ice forecasting.

IceNet-MP fuses satellite observations, Argo float sensor data, and ERA5 reanalysis fields to produce short-term sea ice concentration forecasts. The encode-process-decode architecture translates each input dataset into a shared latent space, allowing new data sources and ML model components to be added without changing the full pipeline.

Quick start

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

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

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

base_path: /path/to/my/data

Then download datasets and train:

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

Evaluate a checkpoint:

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

Documentation

• Installation — prerequisites, uv setup, HPC-specific steps
• Configuration — local config files, model overrides, custom datasets
• Commands — datasets create, datasets inspect, train, evaluate
• Add a model — tensor format, standalone vs. processor model architectures

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.