The WeatherGenerator foundation model - pilot application and service call

The success and wide-spread availability of machine learning approaches in earth and environmental sciences has resulted in a proliferation of deep learning models adapted for weather prediction. Neural networks have proven highly successful for a multitude of data-driven tasks such as bias correction, downscaling and nowcasting. 

Inspired by recent advances in generative modeling of textual data through large language models, the EU WeatherGenerator project aims to develop the leading European AI foundation model for weather and atmospheric climate modeling. This model is pre-trained with petabytes of multi-modal data (reanalyses, station observations, satellite products, etc.) necessitating training on powerful computing clusters, including Europe’s first exascale-class supercomputers. 

In pursuit of this “foundation model” status, and in order to evaluate the flexibility and added value of the model, it is necessary to integrate the WeatherGenerator to a wide range of existing forecast pipelines. The pilot application we present here is nowcasting of heavy precipitation in Western Africa. Dense networks of (openly available) automated weather stations and radars cannot be found in every region of the world. In sub-Sahara Western Africa, a lack of available radar data means that nowcasting is mostly limited to available satellite products, complicating efforts to predict the risk of flash floods from high-intensity precipitation. 

We investigate fine-tuning the pre-trained WeatherGenerator to SEVIRI output, training a tail network that predicts rainfall retrieval from the MSG-CPP product. We also explore transfer learning with WeatherGenerator, using a decoder trained to EURADCLIM over the European continent with SEVIRI input and assessing its accuracy over the target region. 

Finally, we present the details of our upcoming service call, through which the Netherlands eScience Center plans to bring the WeatherGenerator technology to potential stakeholders. The service call will be open to applicants from a broad range of sectors including the European research community, public institutions and industry, and provide for both (short term) deployment support as well as (longer term) explorative research applications.