Aardvark Weather Model Gains Probabilistic Capabilities, Separating Observation and Model Uncertainty

FraunhoferHHI · hf · 2026-09-01

Fraunhofer HHI introduces uncertainty-aware end-to-end AI weather forecasting, making the Aardvark weather model probabilistic. By adding learned observation noise and Monte Carlo dropout to its encoder and processor, the method yields a calibrated nested ensemble. This approach improves mean forecasts and effectively disentangles aleatoric uncertainty from epistemic uncertainty.

Original post →

More from Research

Research channel →