Machine Learning for Earth System Modeling: the Anemoi framework

The main objectives of this course are to:

● Understand how to make a request and download data from the Copernicus Data Storage (CDS)
● Understand the landscape of earth system data formats and how to turn raw sources into AI-ready datasets
● Understand the architecture and design of the Anemoi framework and its end-to-end pipeline
● Build, scale, and operate on Anemoi datasets from meteorological source data
● Understand how graphs, model architecture and training are connected in anemoi-core and configure/launch a training run
● Run forecast rollouts from a trained checkpoint with anemoi-inference

Day 1 – 28/09/2026

10:30-13:30 – Module 0: Getting started with AI-Ready Geospatial Data: From raw data to ARCO Zarr data
• Introduction to geospatial data
• Installation of working environment and MN5 access
• Downloading data from CDSAPI
• Understanding the raw data
• Harmonizing spatial and temporal coordinates
• Producing ARCO Zarr datasets

Day 2 – 29/09/2026 

10:00-13:00 – Module 1&2: Introduction to the Anemoi framework and the Anemoi dataset dependency
• Introduction to Anemoi and the framework architecture
• Packages tour
• Anemoi datasets fundamentals
• Scaling and subsetting of an Anemoi dataset
• Reading and Querying an Anemoi dataset
• Best practices for Anemoi dataset

Day 3 – 30/09/2026

10:00-13:00 – Module 3&4
• From dataset to model construction
• Configure an Anemoi training run using Hydra
• Launch and observe an Anemoi training run
• Running forecast rollouts from a trained checkpoint with Anemoi inference

Day 4 – 01/10/2026

10:00-16:00 (1hr lunch break) – Hackathon
• Build an Anemoi dataset starting from raw data
• Configure and launch a short training run with anemoi-training
• Run inference and visualise the resulting forecast with anemoi-inference

• Joan Vedrí (BSC)
• Pai Peng Wang (BSC)
• Filippo Dainelli (BSC)

• Joan Vedrí is a Research Engineer at the BSC, where he develops AI-ready climate data pipelines of the BSC AI Factory (BAIF). He holds a degree in Physics and a Master’s in Remote Sensing from the University of Valencia. Before joining BSC, he worked as a researcher at the UV, developing AI models to estimate essential climate variables from satellite imagery.

• Pai Peng Wang is a Machine Learning Engineer at the BSC, where he works within the BAIF supporting companies in adopting AI while conducting machine learning research at BSC. He holds a B.S. double majored in Computer Science and Data Science from the University of Wisconsin–Madison and currently a candidate researcher and student to the M.S. in Artificial Intelligence Engineering at Carnegie Mellon University. His research focuses on deep learning for climate and VLM, currently developing neural architectures for Weather temporal interpolation.

• Filippo Dainelli is a Postdoctoral Researcher at the BSC, where he works on the development of a regional weather forecasting model covering the Western Mediterranean domain and supports BAIF activities. He holds a PhD in Applied Machine Learning for Climate Science from Politecnico di Milano.

This four-day, hands-on course introduces participants to machine learning for Earth system modeling, using ECMWF’s Anemoi framework as the guiding case study for going from data to forecast. Day 1 covers how to acquire and prepare AI-ready geospatial data, transforming raw sources (via CDSAPI) into harmonised, analysis-ready ARCO Zarr datasets. This is the general data step that underlies any ML-based Earth system model. Days 2–3 turn to the Anemoi pipeline specifically. Participants first learn the framework’s architecture and the anemoi-datasets package, creating, scaling, and querying training datasets, then move on to anemoi-core, configuring and launching a training run through its Hydra configuration hierarchy, and finally running forecast rollouts with anemoi-inference. Day 4 closes with a hackathon where participants apply the full pipeline to a guided problem. The course assumes basic Python proficiency and familiarity with weather forecasting concepts (NWP, GRIB, ERA5); no prior Anemoi experience is required. Lectures are paired with jupyter notebooks, so each concept is immediately exercised hands-on.

• Comfortable programming in Python (variables, functions, basic libraries)
• Familiarity with fundamental weather and climate concept (Numerical Weather Prediction, reanalysis, GRIB/NetCDF format, ERA5)
• A Climate Data Store (CDS) account

Certificate of Achievement

• Laptop with internet access
• CDS API account and API key
• MN5 access