Uncertainty-Aware End-to-End AI Weather Forecasting: Disentangling Observation and Model Contributions
The paper turns Aardvark Weather into a probabilistic forecaster and separates uncertainty from observations versus learned dynamics.
The authors add learned input-dependent noise to the observation encoder and Monte Carlo dropout to the processor. That nested ensemble lets them attribute forecast spread with a law-of-total-variance decomposition, checked by withholding observation streams. Probabilistic finetuning improves the mean forecast by 4.2% on average, with calibration against ERA5 through medium range. It still trails the operational ECMWF ensemble. HF Daily Papers' note
The authors add learned input-dependent noise to the observation encoder and Monte Carlo dropout to the processor. That nested ensemble lets them attribute forecast spread with a law-of-total-variance decomposition, checked by withholding observation streams. Probabilistic finetuning improves the mean forecast by 4.2% on average, with calibration against ERA5 through medium range. It still trails the operational ECMWF ensemble. HF Daily Papers' note
score 5