Research map / Architectures and efficiency / Attention and transformers
EMFormer: Efficient Multi-Scale Transformer for Accumulative Context Weather Forecasting
Hao Chen, Tao Han, Jie ZHANG, Song Guo, Fenghua Ling, LEI BAI
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 |
|---|---|---|---|
| add accumulative context finetuning (ablation) | EMFormer · ERA5 1.4 deg, 10-day | RMSE: 0.5719 -> 0.5094 | Improved |
| add accumulative context finetuning (ablation) | EMFormer · ERA5 1.4 deg, 10-day | ACC: 0.4614 -> 0.5389 | Improved |
| add KV pruning memory module to accumulative finetuning (ablation) | EMFormer · ERA5 1.4 deg, 10-day | RMSE: 0.5153 -> 0.5094 | Improved |
| add KV pruning memory module (ablation) | EMFormer · ERA5 1.4 deg | GPU memory (G): 4.576 -> 4.913 | Hurt |
| add plain autoregressive finetuning (no accumulation) (ablation) | EMFormer · ERA5 1.4 deg, 6-hour | RMSE: 0.0626 -> 0.0645 | Hurt |
| add plain autoregressive finetuning (no accumulation) (ablation) | EMFormer · ERA5 1.4 deg, 10-day | RMSE: 0.5719 -> 0.5339 | Improved |
| use LoRA finetuning instead of full finetuning (ablation) | EMFormer · ERA5 1.4 deg, 10-day | RMSE: 0.5339 -> 0.5278 | Mixed |
| use EMFormer block instead of self-attention block (ablation) | EMFormer backbone · ERA5 1.4 deg, 10-day | RMSE: 0.5978 -> 0.5719 | Improved |
| use EMFormer block instead of self/windows attention block (ablation) | EMFormer backbone · ERA5 1.4 deg, 10-day | RMSE: 0.6016 -> 0.5719 | Improved |
| use multi-convs layer instead of plain multi-scale module (ablation) | EMFormer · ERA5 weather, 6-hour | training time (hours): 83 -> 60 | Improved |
| use multi-convs layer instead of plain multi-scale module (ablation) | EMFormer-B · ImageNet-1K | Top-1 accuracy (%): 84.3 -> 84.4 | No clear effect |
| use multi-convs layer instead of plain multi-scale module (ablation) | EMFormer · ERA5 weather, 6-hour Z500 | RMSE: 18.6 -> 18.6 | No clear effect |
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.
- MambaOut-Tiny — MambaOut: Do We Really Need Mamba for Vision? (CVPR 2025) · better on 3 of 3 shared results
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
- Scaling transformer neural networks for skillful and reliable medium-range weather forecasting (NeurIPS 2024 · similarity 0.90)
- OmniCast: A Masked Latent Diffusion Model for Weather Forecasting Across Time Scales (NeurIPS 2025 · similarity 0.87)
- SAMformer: Unlocking the Potential of Transformers in Time Series Forecasting with Sharpness-Aware Minimization and Channel-Wise Attention (ICML 2024 · similarity 0.87)
- TimePerceiver: An Encoder-Decoder Framework for Generalized Time-Series Forecasting (NeurIPS 2025 · similarity 0.86)
- WeatherGFM: Learning a Weather Generalist Foundation Model via In-context Learning (ICLR 2025 · similarity 0.85)
- Timer-XL: Long-Context Transformers for Unified Time Series Forecasting (ICLR 2025 · similarity 0.85)
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.
