MLCast

Nowcasting is the short-range prediction of weather from recent observations. Machine-learning approaches to precipitation nowcasting have often been developed separately at individual meteorological services, using national radar archives in different formats. MLCast builds shared infrastructure for this work. Its Python ecosystem combines a configurable ConvGRU ensemble model and interfaces for adding other architectures with a catalogue of high-resolution radar datasets converted to a common analysis-ready, cloud-optimised Zarr format. A published data specification, automated validator and source-specific converters allow archives maintained by different national services to enter the same data-access and training workflow. A sampler converts catalogued source data into training-ready sequences, while reproducible configurations support model training and inference. The current focus is observation-to-observation machine-learning nowcasting over short lead times, principally from radar precipitation data, with satellite data support under development.

MLCast began in September 2024 as a grassroots open-source effort among researchers comparing machine-learning nowcasting approaches and subsequently became Working Group 6 on Nowcasting within EUMETNET’s E-AI Optional Programme. This links an open research community to national meteorological services and operational requirements while keeping software, documentation and development public.

The data infrastructure is federated. MLCast specifies how contributed radar datasets should be structured in Zarr, provides automated checks for schema, metadata, spatial and temporal coverage, units and missing data, and maintains generic and source-specific conversion tools. Contributing services retain responsibility for their national archives while the converted datasets become accessible through a common catalogue. The catalogue currently covers five countries and is hosted through European Weather Cloud infrastructure.

Access to computing is uneven across participating institutes. Data preparation, validation, software testing and documentation do not require GPUs, while model training and benchmarking can use pooled resources. EUMETNET support currently includes an allocation of 10,000 GPU hours on CINECA’s Leonardo supercomputer. Training runs use composable configurations, and resolved experiment configurations are saved so that runs can be reconstructed.

The shared data layer is operational and expanding. The public catalogue currently contains six radar data sources from five countries, representing about 92 cumulative years and more than 6.9 million time steps. Each source is versioned, documented and accessible through the same Python catalogue, while the validator checks new contributions before they enter it. The common specification has therefore already been applied across independently maintained national radar archives rather than remaining only a proposed standard.

The modelling software is publicly released under permissive dual licensing and provides a ConvGRU baseline, configurable training and inference, standard losses and interfaces for external architectures. The package remains pre-1.0 and its documentation explicitly warns that APIs and functionality may change. MLCast also makes geographical incompleteness visible through its public coverage map rather than presenting the dataset as comprehensive.

Shared model evaluation is less mature than the data infrastructure. Models can be assessed with common machine-learning and meteorological metrics, but the dedicated community benchmarking suite is still under development. The project therefore has established reusable data and training infrastructure across institutions, while standardised community benchmarking and packaged pretrained models remain work in progress.

https://mlcast.org/

https://mlcast.org/docs/

https://mlcast.org/data

https://mlcast.org/software.html