Observation-driven Machine Learning Forecasts of Global Weather

Accurate weather forecasting is crucial for predicting and mitigating natural disasters, optimizing operations across agriculture, energy, and logistics, and managing daily activities. Recent strides in pure machine learning (ML) models for global weather forecasting have demonstrated impressive potential, often surpassing traditional physics-based numerical weather prediction (NWP) models while running at a fraction of the computational cost. However, these ML-based models have been entirely reliant on reanalysis data, leading to two notable limitations: (i) the models do not learn directly from observational data, and (ii) operationalizing these models depends on data assimilation processes carried out by a limited number of global weather agencies, such as ECMWF or NOAA. This dependency constrains the cadence, latency, and choice of observational datasets used for forecasts, thereby limiting the otherwise democratizing impact of ML-based weather modeling.

We present a deep learning system for global weather forecasting that generates probabilistic forecasts from observational data. Our system integrates diverse observational sources such as microwave and infrared radiances, ground station measurements, and other in situ data to capture the current atmospheric state and predict its evolution over time. This approach opens the door to (i) direct learning from extensive multi-modal observational data, including sources not currently leveraged in global weather forecasts, and (ii) the possibility for some degree of observation-to-forecast modeling to occur entirely outside of conventional weather forecasting centers. This points to a fuller realization of ML’s democratizing potential, in which individual researchers or organizations can develop and operate their own global forecast models.