Otter Weather Model Boosts Medium-Range Forecasts with High Efficiency

State-of-the-art AI weather models have shown superior performance over traditional Numerical Weather Prediction (NWP) for medium-range forecasts. However, their massive training computational requirements often limit their accessibility and rapid iteration, especially for groups with fewer resources. This new research introduces "Otter Weather," a highly efficient spatiotemporal forecasting model designed to democratize high-performance AI-driven weather prediction. The Otter family of models significantly advances the balance between predictive skill and computational cost. The deterministic version of Otter Weather surpasses the best NWP baseline by 9.6% at a 24-hour lead time, achieving this with fewer than 3.5 A100-days of training. This represents a twofold efficiency gain over other lightweight AI models and a remarkable 100-fold reduction in compute compared to more resource-intensive frontier architectures. The efficiency gains extend to probabilistic forecasting, where Otter-XL, a larger architecture, improves upon the IFS ENS baseline by 9.7% in Continuous Ranked Probability Score (CRPS). This translates to nearly double the predictive skill of comparable lightweight models at similar compute budgets and outperforms frontier architectures like GenCast by over 2% with an order of magnitude less compute. The model's applicability also extends beyond weather, demonstrating superior performance in a complex acoustic scattering PDE task, suggesting its potential across various scientific domains.