The authors introduce GraphCast–Mamba, which adds a correction branch to a frozen pretrained 1° GraphCast backbone.
Memory-augmented Mamba branch improves GraphCast forecast error by 14 percent
Training a temporal memory correction on ERA5 data reduces RMSE across all variables at day 10 by 7 to 14 percent.
Big Tech
Ilya Vilkoviskiy · Liang-Hong Mo · Suozhi Huang · Dmitry A. Abanin
Princeton University · École Polytechnique Fédérale de Lausanne (EPFL) · Google Research
Research Digest··2 min read
The authors augment a frozen GraphCast backbone with a trainable Mamba correction branch that retains temporal memory across forecast steps.
§