HealDA: Highlighting the importance of initial errors in end-to-end AI weather forecasts

TL;DR

HealDA is a novel ML-based data assimilation system that provides initial atmospheric states directly from observations, enabling ML weather forecasts with minimal loss of skill and without reliance on traditional NWP infrastructure.

Contribution

HealDA introduces a simple, direct observation-to-state ML data assimilation method that effectively initializes ML weather forecast models, reducing dependence on NWP data assimilation.

Findings

HealDA-initialized forecasts lose less than one day of effective lead time.

HealDA-initialized FCN3 ensembles trail ECMWF IFS ENS by less than 24 hours.

Forecast error growth remains unchanged from HealDA initialization, with skill primarily limited by initial analysis error.

Abstract

AI weather models now rival leading numerical weather prediction (NWP) systems in medium-range skill. However, almost all still rely on NWP data assimilation (DA) to provide initial conditions, tying them to expensive infrastructure and limiting the practical speed and accuracy gains of ML. More recently, ML-based DA systems have been proposed, which are often trained and evaluated end-to-end with a forecast model, making it difficult to assess the quality of their analysis fields. We introduce HealDA, a global ML-based DA system that maps a short window of satellite and conventional observations directly to a 1{\deg} atmospheric state on the HEALPix grid, using a smaller sensor suite than operational NWP. We treat HealDA strictly as a DA module: its analyses are used to initialize off-the-shelf ML forecast models without any fine-tuning of either. For a variety of off-the-shelf ML forecast models, including FourCastNet3 (FCN3), Aurora, and FengWu, HealDA-initialized forecasts lose less than one day of effective lead time when scored against ERA5. HealDA-initialized FCN3 ensembles similarly trail those of the ECMWF IFS ENS system by < 24 h. We find that forecast error growth in these models is unchanged from HealDA initialization, and the skill gap primarily arises from the larger initial error of the HealDA analysis. Spectral analysis reveals that this stems from overfitting to the large scales and upper-tropospheric fields. We also demonstrate that small changes in the verification setup can shift apparent skill by 12--24h, underscoring the need for consistent scoring. Taken together, these results clarify the current performance of ML-based DA systems and show that a relatively simple, direct observation-to-state network can already provide initial conditions that are usable by state-of-the-art ML forecast models with only modest loss in medium-range skill.