IceNet is a deep learning sea ice forecasting system developed by an international team and led by the British Antarctic Survey and The Alan Turing Institute.
The original IceNet research model, published in Nature Communications was trained on climate simulations and observational data to forecast the next 6 months of monthly-averaged sea ice concentration maps.
This version advanced the range of accurate sea ice forecasts, outperforming a state-of-the-art dynamical model (ECMWF SEAS5) in seasonal forecasts of summer sea ice, particularly for extreme sea ice events.
Since then, the IceNet team has focussed on building an operational version of the model which forecasts on a daily resolution.
The original research code was refactored into
icenet: a library for operational forecasting.
icenet library can support further research efforts into AI-based operational sea ice forecasting
In addition, several use cases and an ecosystem of service components are contained within this organisation, supporting execution and downstream analysis.
For further information about the team involved, please look at the project pages at BAS or The Alan Turing Institute.
How do I get started?
Please visit the icenet-notebooks repository and check out the instructions there! :wink:
The repositories listed below are the key ones to pay attention to.
Example notebooks showing how to use the library.
The IceNet library, containing documentation sources. Probably easiest to start with the icenet-notebooks or icenet-pipeline repositories for guidance in the first instance.
A detailed explanation of the machine learning pipeline can be found here. Fundamentally, there is a reusable and continuously executable pipeline from data ingestion through to forecast production and upload. This repository complements the icenet-notebooks repository which explains both CLI and programmatic usage, by actually implementing a load of workflows in bash.
This high level diagram depicts the structuring of the pipeline:
The daily forecasting pipeline centers around a refactored version of the original research model for monthly forecasts, located in this library. An example to using the library can be found here which leverage a BAS developed tool for running model ensembles. on HPCs.
This project is for the code that provisions the Azure cloud infrastructure. It sets up various Azure resources and is currently under heavy development.
Forecast NetCDFs are deposited into the cloud infrastructure for dissemination and processing.
For further details look at the repository here.
This project has been merged with the icenet-etl project.
This is the public project repository where we manage the collaboration between the British Antarctic Survey (BAS) and Alan Turing Institute (ATI). We’re always happy to have additional contributors so please feel free to drop by!
- Scott Hosking of BAS and The Alan Turing Institute highlighting the importance of AI approaches on Al Jazeera.
- James Byrne described how the IceNet project reflects best practice software engineering for climate science at Climate Informatics 2023…
- …and demonstrated the original proof of concept pipeline approach online for AI UK 2022 with Tom Andersson and James Robinson
- Tom Andersson describing IceNet…
- …and giving a more detailed talk on the research.