Varda-single-1.0: MétéoSuisse is testing an AI forecasting system accurate to the nearest kilometre across the Alpine terrain

This data-driven weather forecasting model rivals the Swiss operational numerical models for up to five days, but it still smooths out thunderstorms too much and underestimates certain local gusts.

A team working with MeteoSwiss presents Varda-single-1.0, a machine learning weather forecasting system designed for the Alpine region. It produces hourly forecasts on a grid one kilometre above Switzerland, nested in a global forecast at 31 km. The whole thing is based on two graph transformers developed with the Anemoi framework: one advances the atmosphere in six-hour steps, the other reconstructs the intermediate hours.

A three-step learning process

The model was first pre-trained on the ERA5 global reanalysis, then on twenty years of regional reanalysis at the kilometre scale, before refining on operational analyses. Evaluated over a full year, it globally equals the ICON-CH1 numerical model at 1 km up to 33 hours of maturity and generally exceeds ICON-CH2 at 2 km for up to 120 hours.

The authors do not hide its limitations: local wind maxima are underestimated and convective precipitation appears too smooth, a classic bias of models trained to minimize square error. Three case studies in particular show weaknesses in winds in complex terrain. Varda-single is designed as a complement to physical models and as a pre-trained basis for other applications.