Operational fine-tuning of the temporal downscaler

Determine whether fine-tuning the Varda-single temporal downscaler on operational analyses or operational forecast data improves its hourly forecast quality.

Background

The temporal downscaler generates hourly atmospheric states between the six-hourly outputs of the Varda-single forecaster. Unlike the forecaster, it is trained exclusively on reanalysis data because global operational analyses are not available at hourly resolution. This creates a potential distribution mismatch between training and operational inference data.

The paper identifies operational analyses and operational forecasts as possible sources for an additional fine-tuning stage, but does not evaluate whether either option improves the hourly forecasts.

References

Global operational analyses are not available at hourly resolution. For this reason, the temporal downscaler does not undergo fine-tuning and remains trained on reanalysis data only. One alternative would be to use operational forecasts for the global domain instead; we leave this exploration to future work.

— Varda-single-1.0: deterministic data-driven weather forecasting at 1 km resolution over Switzerland's complex topography  (2610.01835 - Pennino et al., 1 Oct 2026) in Section 4, Training curriculum; Section 6, Conclusions