Otter Weather: Skillful and Computationally Efficient Medium-Range Weather Forecasting

As AI-driven weather forecasting models increasingly outperform traditional Numerical Weather Prediction (NWP), they have created a new barrier to entry: massive computational requirements. This paper investigates whether high-performance weather forecasting can be democratized by replacing bespoke, Earth-specific architectural components with general-purpose, hardware-optimized methods used in large language models and computer vision.

The authors introduce Otter Weather, a spatiotemporal forecasting model built on a 2D Swin-UNet Transformer architecture. Instead of using specialized spherical geometry operators or custom graph neural networks, the model employs standard, highly optimized components: SwiGLU activations, Rotary Positional Embeddings (RoPE), and the Muon optimizer. By treating the weather forecasting task as a general spatiotemporal sequence problem, the researchers avoid the overhead of domain-specific inductive biases and leverage mature hardware acceleration. The model is trained on the ERA5 reanalysis dataset and evaluated using standard WeatherBench protocols, with both deterministic and probabilistic variants developed.

Otter Weather significantly advances the skill-compute Pareto frontier. The deterministic version outperforms the best NWP baseline by 9.6% at a 24-hour lead time while requiring fewer than 3.5 A100-days of training. The scaled-up version, Otter-XL, achieves a 9.7% improvement in the Continuous Ranked Probability Score (CRPS) over the IFS ensemble baseline, outperforming frontier architectures like GenCast while using an order of magnitude less compute. Furthermore, the authors demonstrate the model's versatility by applying it out-of-the-box to an acoustic scattering partial differential equation (PDE) task, where it outperforms specialized foundation models, suggesting that these architectural efficiencies are broadly applicable across scientific domains.