Research map / Architectures and efficiency / RNNs and normalization
Scaling Laws of Global Weather Models
Yuejiang Yu, Langwen Huang, Alexandru Calotoiu, Torsten Hoefler
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 |
|---|---|---|---|
| Increase training data by 10x | Aurora · ERA5 validation loss | validation loss: up to 3.2x lower loss | Improved |
| Use Aurora instead of other models for data scaling | Aurora · ERA5 data-scaling exponent | beta: 0.51 vs 0.30-0.46 for others | Improved |
| Widen network at fixed parameter count (instead of deepening) (ablation) | Aurora, GraphCast, Pangu, SFNO · ERA5 validation loss | validation loss | Improved |
| Increase width to 256 with fewer message-passing steps vs width 128 with more steps (ablation) | GraphCast · ERA5 validation loss, 1.6M params | validation loss: width=256 vs width=128 | Improved |
| Reduce depth to 1 (ablation) | GraphCast, SFNO · ERA5 validation loss | validation loss | No clear effect |
| Double dataset size (15.0 TB to 30.0 TB) (ablation) | all five models · parameter-scaling law | exponent delta | Improved |
| Increase dataset from 15.0 TB to 30.0 TB (ablation) | Pangu · parameter-scaling exponent | delta | Improved |
| Use maximal update parameterization (muP) (ablation) | AIFS · AIFS largest model (width 1024) validation loss | validation loss | Hurt |
| Allocate extra compute to data rather than model size | all models · compute-optimal frontier | validation loss | Improved |
| Use transformer attention (Aurora) instead of GNN message passing (GraphCast) | Aurora vs GraphCast · H100 training utilization | GPU utilization (%): 1.03 -> 37.2 | Improved |
| Use GraphCast for 2m temperature | GraphCast · ERA5 2T RMSE | RMSE | Improved |
Closest papers by meaning
- Scaling transformer neural networks for skillful and reliable medium-range weather forecasting (NeurIPS 2024 · similarity 0.87)
- Generalizing Weather Forecast to Fine-grained Temporal Scales via Physics-AI Hybrid Modeling (NeurIPS 2024 · similarity 0.87)
- WeatherSyn: An Instruction Tuning MLLM For Weather Forecasting Report Generation (ICML 2026 · similarity 0.86)
- Scaling Storm-Resolving Atmospheric AI Simulation to the Entire Planet (NeurIPS 2026 · similarity 0.86)
- EWMoE: An Effective Model for Global Weather Forecasting with Mixture-of-Experts (AAAI 2025 · similarity 0.85)
- Scaling Law for Time Series Forecasting (NeurIPS 2024 · 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.
