Original title: Uncertainty-Aware End-to-End AI Weather Forecasting: Disentangling Observation and Model Contributions
Rodrigo Almeida, Noelia Otero, Jost Arndt, Simon Baur +2 · August 2026
Modern AI models can forecast global weather straight from raw Earth observations cheaply and accurately. However, they only give a single deterministic guess without showing how uncertain they are. To fix this, researchers updated the Aardvark Weather model to make probabilistic forecasts. They tackled uncertainty by adding two random mechanisms: one injects learned noise where raw observations enter the model to catch measurement errors, and another randomly drops connections inside the main processor to catch errors in the learned weather physics. This setup creates an ensemble of forecasts that separates these two sources of error. The new probabilistic training improved average forecast skill by 4.2%. It also correctly matches real-world spread and produces more reliable, transparent weather predictions.
Imagine you're planning a picnic tomorrow. You look at two apps: one says sunny, the other says 60% chance of rain. You don't know why they disagree. Is it because the sensors are noisy or because the models are different? In this paper, the weather AI tells you exactly: part of the uncertainty comes from the instruments that measure the atmosphere, and part comes from how the model simulates the future. So instead of just a probability, you get a breakdown of why it's uncertain.
Forecasters can now trust not just the prediction but also its reliability. Knowing whether uncertainty comes from noisy satellite data or incomplete physics helps prioritize improvements. This makes end-to-end AI weather models more useful for planning and supports observation-driven digital twins of the atmosphere.
The ensemble trails the operational ECMWF ensemble in skill. The decomposition into observation and model contributions is validated only by withholding complete observation streams, not by independent ground truth. The approach is specific to the Aardvark architecture and may not transfer to other end-to-end models without modification.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Uncertainty-Aware End-to-End AI Weather Forecasting: Disentangling Observation and Model Contributions — Rodrigo Almeida, Noelia Otero, Jost Arndt et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.