Research map / Architectures and efficiency / Attention and transformers
(Sparse) Attention to the Details: Preserving Spectral Fidelity in ML-based Weather Forecasting Models
Maksim Zhdanov, Ana Lucic, Max Welling, Jan-Willem van de Meent
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| compress latent grid (enter at Nside=32 instead of 64) (ablation) | MOSAIC · 10m wind spectrum, ERA5 1.5deg | spectral ratio near Nyquist: 0.97 -> 1.02 | Hurt |
| predict residual x_{t+1}-x_t instead of next state directly (ablation) | MOSAIC · 10m wind spectrum, ERA5 1.5deg | spectral ratio at highest wavenumber: 0.97 -> 1.00 | Hurt |
| train probabilistically with ensembles instead of deterministically to the ensemble mean | ML weather models · 10m wind spectrum, 1.5deg | spectral ratio at highest wavenumber: 0.24 (Stormer) -> 0.97 (MOSAIC) | Improved |
| use block-sparse attention instead of dense FlashAttention and NSA | attention layer · forward pass runtime up to 512k tokens | speedup: 61.8x over dense, 9.4x over NSA | Improved |
| inject noise z as additive bias in SwiGLU gates instead of conditioning layer norm (ablation) | functional perturbations · ensemble calibration | spread-to-skill ratio: 0.83 to 0.95 | Improved |
| MOSAIC (1.5deg) vs 0.25deg-trained deterministic models | MOSAIC · Z500 10-day RMSE | RMSE: 610 vs Aurora 668, Pangu 800, GraphCast 748 | Improved |
| MOSAIC vs ArchesWeatherGen on Z500 10-day ensemble mean | MOSAIC · ERA5 1.5deg Z500 | RMSE: 610.36 -> 624.08 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Aurora — AuroRA: Breaking Low-Rank Bottleneck of LoRA with Nonlinear Mapping (NeurIPS 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Introducing Spectral Attention for Long-Range Dependency in Time Series Forecasting (NeurIPS 2024 · similarity 0.85)
- Tyche: One Step Flow for Efficient Probabilistic Weather Forecasting (NeurIPS 2026 · similarity 0.85)
- Probabilistic Weather Forecasting with Hierarchical Graph Neural Networks (NeurIPS 2024 · similarity 0.84)
- Calibrating Scientific Foundation Models with Inference-Time Stochastic Attention (NeurIPS 2026 · similarity 0.84)
- WeatherGFM: Learning a Weather Generalist Foundation Model via In-context Learning (ICLR 2025 · similarity 0.83)
- Learning with Calibration: Exploring Test-Time Computing of Spatio-Temporal Forecasting (NeurIPS 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
