Mosaic Model Achieves State-of-the-Art Weather Forecasting with Spectral Fidelity

A novel weather forecasting model, Mosaic, utilizes a probabilistic approach to resolve spectral degradation challenges in machine learning predictions. It creates ensemble members via learned functional perturbations and functions on native-resolution grids by employing block-sparse attention mechanisms. This method, aligned with hardware, effectively captures long-range dependencies at a linear computational cost by sharing keys and values among spatially adjacent queries. With 214 million parameters operating at 1.5-degree resolution, Mosaic either matches or surpasses models trained on data six times finer for key upper-air variables. The system generates well-calibrated ensembles, with individual members exhibiting nearly perfect spectral alignment across all resolved frequencies. A 10-day forecast with 24 members is completed in under 12 seconds, addressing two main sources of spectral degradation: training against ensemble means and compressive encoding that leads to information bottlenecks. Mosaic achieves leading results among weather forecasting systems at 1.5-degree resolution.