Google's WeatherNext 3 boosts forecast accuracy with physical data

Traditional weather models simulate physical processes, while AI models like WeatherNext 3 learn patterns from historical data. The new model bridges these by feeding in physical data about the surface—land versus ocean, and elevation—during training on past station observations. This hybrid approach yields measurable gains.

The paper claims improvement over both its predecessor and the European Centre for Medium-Range Weather Forecasts model. However, the white paper itself notes unexplained quirks, such as worse performance in the first six hours of a forecast for some variables, and visual artifacts like hexagonal blobs in precipitation maps. The global average temperature inconsistency in multi-run surface forecasts also raises questions about the ensemble method's calibration.

The stakes for Google are high, as this model is now the default forecast source across key products. The team calls it a major step forward by using information-dense observation data. Whether these oddities affect real-world reliability in a meaningful way is uncertain, but they are acknowledged areas where the model does not yet behave as expected—potential targets for refinement in future iterations.