av555/aardvark-weather

Aardvark Weather

This repo contains a machine learning ready version of the observational dataset developed for training and evaluating the Aardvark Weather model. We gratefully acknowledge the agencies whose efforts in collecting, curating, and distributing the underlying datasets made this study possible.

Dataset

We provide two types of data: a timeslice of sample data in the format expected for ingestion to the model for demo runs and a training dataset for those interested in developing their own end to end models.

Sample data

A timeslice of sample data is available under /sample_data. This is already in the format to run with the model weights provided at https://github.com/annavaughan/aardvark-weather-public providing a demo of running Aardvark Weather.

Training data

To allow others to explore training their own end-to-end systems, we additionally provide a training dataset for 2007-2019 for each of the observational sources utilised in the paper in ML ready format under /training_data. Note that the Aardvark weather model additionally uses ERA5 for training targets and climatology. These are not included here in the interests of minimising storage space, however can easily be downloaded from the Weatherbench project at https://weatherbench2.readthedocs.io/en/latest/data-guide.html. The observation sources included are as follows:

• AMSU-A (gridded satellite data)
• AMSU-B/MHS (gridded satellite data)
• HIRS (gridded satellite data)
• ASCAT (gridded satellite data)
• GRIDSAT (gridded satellite data)
• IASI-PCs (gridded satellite data)
• HadISD (station observations)
• ICOADS (ship observations)
• IGRA (upper atmosphere observations)

For details of processing, data sources and an ablation of the impact of different sources on model performance see https://www.nature.com/articles/s41586-025-08897-0.

FAQ

Is the model/data available in real time?

The datasets used in this initial prototype are not available in real time. We are however in the process of building a fully operational system, Aardvark 2.0.

Why is there missing data?

There are complex patterns of missingness in the data resulting from a variety of reasons, including satellite launch dates, sensor failure, quality control and lack of availability in the source datasets. All missing values are indicated by nans.

Will further versions of the dataset be provided?

Yes, this project is ongoing and we aim to continue updating this dataset. Updates currently in progress include the development of a real time data feed, extension of current historical dataset to 2025 and inclusion of further instruments. If you would like to receive updates as new products become available please email [email protected] with subject line 'Aardvark dataset updates'.

Can this dataset be used commercially?

The dataset is released under a non-commercial no-derivatives license. This is due to the lisencing inherited from the source datasets.