Large regional reanalysis to train AI weather prediction models

AI Technology: Deep Learning

Machine learning weather prediction (MLWP) systems are progressing rapidly on all topics related to numerical weather prediction. However, they still struggle to represent fine scales and extreme events. 

Concurrently, extremes events are the most impactful on human activities and often occur at local scale, thus it motivates the need to improve the capacity of MLWP systems to provide skillful forecasts for extreme events at high resolution. 

This project aims to train a MLWP system on a regional reanalysis of unprecedented length. 

The study's regional reanalysis, ARRA, has a span of 50 years at hourly frequency and a resolution of 2.5km over Western Europe. 

The team adopt a methodology with successive training runs on length-increasing subsets to evaluate the improvement in the prediction of extreme events.

 Physical consistency of the forecasts and suitability of the system to operational use (including time resolution increase and training optimisations) are equally considered in the developments planned in this project. Such developments are carried out within the open-source collaborative framework Anemoi.