In plain words
Many weather models, including GraphCast, predict from two recent snapshots of the atmosphere. Earlier weather may hold clues that those snapshots miss.
GraphCast, Mamba adds a trainable branch that remembers information between forecast steps and corrects GraphCast’s predictions, while leaving the original model fixed.
In tests on 122 forecast starts from 2023, its weighted error was lower than GraphCast’s across ten-day forecasts. That result applies to the tested model and setup; it does not mean every forecast improved.
The research
Can earlier weather states help correct a forecast when a model sees only two recent atmospheric states? Those inputs may omit unresolved influences, but the history of resolved weather could hold clues. The paper tests whether carrying that history forward improves forecasts.
GraphCast–Mamba adds a trainable branch that carries memory between forecast steps and corrects a frozen, pretrained GraphCast forecast.
The authors trained only the correction branch on ERA5 data from 2015–2021, selected a checkpoint using 2022 development forecasts, then evaluated it against uncorrected GraphCast on 122 forecast starts from 2023 at 1° resolution. Across the pooled ten-day rollout, its normalized, weighted squared-error loss was 14.42% lower.
At the final 240-hour lead, losses pooled across those starts were 21.66% lower, showing that the aggregate gain persisted to day 10.
The main comparison adds the whole correction branch, so it cannot attribute the full forecast gain to memory alone.
The reset controls support a memory benefit for the tested configuration and training run; they do not measure variation across training runs or establish transfer to other forecasters and resolutions.
Analysis
Deep
- Original paper authors
- Ilya Vilkoviskiy, Liang-Hong Mo, Suozhi Huang, Dmitry A. Abanin
- Submitted to arXiv
- License
- arXiv non-exclusive