Unlike traditional weather models, which use the physical characteristics of a location to simulate atmospheric processes, machine learning systems typically operate as black boxes. They learn from historical patterns and use that data to predict future weather conditions. Google’s WeatherNext 3 model combines this data-driven approach with targeted physical information to improve location-specific forecasts.
For each requested location, WeatherNext 3 calculates surface temperature and dew point while also determining whether the point is on land or water. The model factors in surface elevation as well. According to the research team, training on historical weather-station data labeled with these geographic details produces more accurate and reliable weather predictions.
WeatherNext 3 improves forecast accuracy
The WeatherNext 3 white paper reports improved performance compared with both the previous WeatherNext 2 model and artificial intelligence forecasting systems from the European Centre for Medium-Range Weather Forecasts (ECMWF).
For upper-atmosphere conditions, the model delivers an accuracy improvement of approximately 5 percent compared with earlier versions. Google says that gain is equivalent to extending accurate forecast lead times by about six hours. Changes to the way WeatherNext 3 calculates surface temperatures for specific locations also increase accuracy by as much as 30 percent.
WeatherNext 3 generally outperforms ECMWF’s AI forecasting model across these benchmarks. However, the research identifies one unusual exception without offering a clear explanation for it. For many weather variables, WeatherNext 3 performs better when compared with other models’ initial six-hour forecasts, but its advantage changes when predicting conditions over the following 15 days.
The model also has some visual and technical quirks. In certain forecasts, including precipitation maps, the structure of its underlying grid can appear as distinct hexagonal patterns. Its method for generating multiple surface-temperature forecasts—each representing a different possible outcome—can also produce snapshot-like results. These variations may affect how local weather conditions are represented when calculating broader regional or global averages.
Despite these limitations, the researchers say WeatherNext 3 “represents a major step forward in AI-based weather prediction by leveraging information-dense, low-latency observational data rather than relying purely on analytics.”
WeatherNext 3 now powers weather forecast information across several Google services, including Google Search, Gemini, and Google Maps.
Source: arstechnica.com