Newest papers

GraphCast-Mamba: Learning Non-Markovian Corrections for Weather Forecasting

Ilya Vilkoviskiy, Liang-Hong Mo, Suozhi Huang, Dmitry A. Abanin · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.28512 · cited by 0 works (OpenAlex) · open-access PDF

Portuguese

Os processos atmosféricos não resolvidos podem introduzir a dependência da história além das janelas de entrada curtas usadas por muitos modelos de previsão do tempo. Introduzimos o GraphCast--Mamba, que aumenta a espinha dorsal do GraphCast congelada, pré-treinada de $1\circ$ com um ramo de correção de Mamba treinável que combina a passagem de mensagens de gráfico e as camadas de Mamba. O ramo mantém um estado de memória temporal através de etapas de previsão de seis horas sucessivas, aprendendo correções informadas pelo contexto temporal além da entrada de dois quadros do GraphCast. O treinamento está confinado ao ramo de correção, usando dados ERA5 de 2015-2021. Depois de 2.500 atualizações do otimizador, incluindo aquecimento, o

Machine translation (Meta M2M100); the English original follows.

Abstract

Unresolved atmospheric processes can introduce history dependence beyond the short input windows used by many weather forecasting models. We introduce GraphCast--Mamba, which augments a frozen, pretrained $1^\circ$ GraphCast backbone with a trainable Mamba correction branch combining graph message passing and Mamba layers. The branch retains a temporal memory state across successive six-hour forecast steps, learning corrections informed by temporal context beyond GraphCast's two-frame input. Training is confined to the correction branch, using ERA5 data from 2015--2021. After 2,500 optimizer updates, including warm-up, the selected model achieves substantial improvements relative to the uncorrected GraphCast backbone. It reduces the GraphCast loss (normalized weighted mean-squared error; MSE) by 14.42% when aggregated over the full ten-day forecast and by 21.66% at day 10 (240 h after forecast initialization). At this final forecast lead time, root-mean-square error (RMSE) decreases by 7.09--14.60% across all 11 evaluated variable groups, spanning temperature, wind, pressure, geopotential, specific humidity, and precipitation. These results demonstrate broad improvements in ten-day weather forecasts through a short correction-training schedule while retaining the pretrained backbone.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Context (archaeology), Computer science, Schedule, Meteorology, Weather forecasting, Forecast skill

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange