Abstract
<jats:p>Abstract. Accurate global weather forecasting at increasing spatial resolution is a central challenge in atmospheric science. Machine learning models have recently matched or exceeded numerical weather prediction systems on standard medium-range benchmarks, with transformer-based architectures at their core. However, the self-attention mechanism underlying these models scales quadratically in memory and compute with sequence length — a critical bottleneck as training and inference resolutions approach operational standards. We describe Mamba-Stormer, which replaces the transformer backbone in the Stormer architecture with a bidirectional Vision Mamba (BiMamba) backbone. Our key design contributions are: (1) bidirectional Mamba scanning applied in an alternating horizontal–vertical pattern across 14 layers to capture 2D atmospheric structure, and (2) the integration of bidirectional SSMs into Stormer’s adaLN-Zero residual block by replacing the self-attention sub-block while retaining the feed-forward sub-block and zero-initialized lead-time conditioning. In a controlled local comparison on ERA5 WeatherBench 2 (240x121 grid, 69 atmospheric variables, multi-step finetuned), Mamba-Stormer outperforms the 24-layer Transformer baseline: +2.73 % mean per-variable RMSE improvement at 6h, +2.34 % at 72 h, and +1.07 % at 120 h (all p < 0.01), with BiMamba winning on 67/69, 69/69, and 66/69 variables respectively. Simultaneously, its O(L) backbone delivers a 1.10x computational throughput speedup at the WeatherBench 2 1.5° training resolution (121x240), growing to 2.59x at 512x512 as the O(L²) attention cost increasingly dominates. These results suggest that bidirectional Vision Mamba is a strong backbone for neural weather prediction — achieving better accuracy and growing computational efficiency, consistent with a more suitable spatial inductive bias for atmospheric modeling.</jats:p>