A weather model brings detail to global forecasts
Efficient probabilistic weather forecasting at quarter-degree resolution with two-sided transport attention · arXiv:2610.35545
In plain words
Astra is a weather model that makes several possible forecasts for the whole globe, with a detailed grid and updates every six hours. To share information across distant parts of the globe without comparing every grid point with every other one, it chooses places to read from and places to send information to.
In tests on 2020 weather, its probabilistic forecast score matched GenCast’s on five upper-air measures averaged over days 1–10, and was lower on 441 of 600 tested combinations.
The research
Can a probabilistic forecast keep detailed global weather information while avoiding the cost of comparing every grid point with every other? One-way selection can reduce that cost, but may leave some regions unread.
The paper asks whether a model can route information across the globe while retaining competitive forecast scores.
Astra combines two-way information routing with ensemble training to make detailed global forecasts, then compares its forecast scores and resource use with GenCast’s.
The challenge is to make probabilistic forecasts on a detailed global grid without the cost of comparing every location with every other. One-way selection can also leave some regions unread.
Astra routes information both to and from selected regions, then trains its forecasts using weather fields and features from a frozen copy of the forecaster. Against ERA5 on 732 2020 initialisations, its probabilistic forecast score matched GenCast’s averaged over five upper-air targets and days 1–10; across 600 target and forecast-time combinations, Astra scored lower in 441. The result supports competitive forecast skill on this test, alongside lower reported resource use.
The forecast comparison covers 2020 initialisations against ERA5, so it does not establish performance in other years or in operational forecasting.
What the evidence shows
- The global grid challenge
Astra is a probabilistic weather model for global forecasts on a 0.25° grid, updated in six-hour steps. The challenge is that weather can connect places thousands of kilometres apart, but comparing every grid point with every other is too costly.
One-way selection can also leave some places unread. The paper’s question is how to exchange information across the global grid while keeping this exchange manageable.
- Two-way information routing
Astra’s Kantorovich Attention layer, used at the network’s 1° and 2° levels, lets each tile select both tiles to read from and tiles to send information to. Every tile is therefore read at token resolution by other tiles, while the rest contribute through pooled mean values.
This design responds to the coverage gap in one-way selection without requiring dense comparisons across the entire grid.
- Testing the routing design
The authors checked the routing selections against an optimal plan computed for each trained-network score matrix, using the same number of tile pairs. Averaged over matrices from network blocks and heads at the 1° and 2° levels, the two-way selection achieved 94–98% of the optimal plan’s gain over random pairs.
A one-way control left some tiles unread in the routing comparisons, while the two-way construction left none.
- Comparing forecast scores
For forecast skill, the authors scored Astra and GenCast against ERA5 on the same 732 initialisations from 2020. Their probabilistic forecast scores matched when averaged over five upper-air targets and days 1–10.
Across 600 combinations of 20 upper-air and surface targets and 30 forecast times, Astra had a lower score in 441. The count shows how often this was true across those combinations, not that it held for every target or time.
- Reported resource use
The paper also compares resource use for the reported training and inference procedures. Astra’s hardware-normalised training compute was 9.8% of GenCast’s, based on the stated conversion factors.
For one 15-day forecast member, Astra’s counted operations were 1.9 PFLOP versus 75 PFLOP for GenCast’s released GPU implementation. These comparisons do not establish equivalent training procedures or runtime on identical hardware.
Deep
- Original paper authors
- Junu Choi, Youngjoon Hong
- Submitted to arXiv
- License
- arXiv non-exclusive