Google's AI-Boosted Weather Model Takes Flight with Improved Accuracy

Google has updated its AI-powered weather model, WeatherNext 3, to improve forecast accuracy. Unlike traditional models that rely on physical properties, machine-learning models like WeatherNext 3 train on past patterns and make predictions about future ones.

The team behind WeatherNext 3 has added a new feature: it takes into account physical information such as surface elevation and whether the location is land or ocean. This allows for more accurate calculations of surface temperature and dew point, leading to better forecast predictions.

According to the white paper, WeatherNext 3 shows significant improvements over its previous model, WeatherNext 2. For example, it improves upper atmosphere condition accuracy by about 5 percent, which translates to around six hours of more accurate forecast lead time. It also beats the European Centre for Medium-Range Weather Forecasts (ECMWF) AI model on this metric.

However, there are some curious exceptions. In a few cases, WeatherNext 3 performs worse than the initial six-hours-ahead forecast from other models before pulling ahead in the remaining 15-day forecast. Larger-scale patterns also show some weirdness, such as distinctly hexagonal blobs of precipitation and inconsistent global average temperatures.

Despite these quirks, the team says that WeatherNext 3 represents a major step forward for AI-based weather predictions by utilizing information-dense, low-latency observation data. The new model is now used across Google services, including Search, Gemini, and Maps.