Reproduce “Efficient probabilistic weather forecasting at quarter-degree resolution with two-sided transport attention” on OpenResearch

About this paper

Junu Choi, Youngjoon Hong

Astra produces quarter-degree probabilistic weather forecasts using two-sided tile routing and a training objective that scores weather fields and multiscale representations from the model itself. Its average CRPS, where lower is better, matches GenCast and is 7.1% below IFS ENS across five upper-air targets at days 1–10. IFS ENS uses its own analysis for verification; Astra and GenCast use ERA5.

Key insights

• Two-sided routing addresses an information-coverage issue in sparse attention: each tile selects where it reads from and where it sends information. Its tested attention plans capture 94–98% of the optimal transport plan’s gain over random pairings.
• The feature-space CRPS supplements pointwise and pyramid-scale field scoring with learned representations, while keeping the representation model fixed during each scoring update.
• Astra’s architecture concentrates expensive global attention at coarse resolutions, enabling direct quarter-degree forecasts without a separate latent encoder or mesh.

Summary by alphaXiv.