Google's first-party AI weather forecasting models are asserting their position as global leaders in accuracy, with the company's GenCast model significantly outperforming traditional systems. Bryan Beal, a Google executive, recently highlighted this development on social media, stating, > "Most people don’t know that Google has our own first party AI weather forecasting models, and they’re proving to be the most accurate in the world." This claim is supported by recent research from Google DeepMind.
The GenCast model, an AI ensemble system, provides 15-day weather predictions with high resolution (0.25°). It demonstrated superior forecasting skill compared to the European Centre for Medium-Range Weather Forecasts’ (ECMWF) ENS, a top operational system, in 97.2% of tested scenarios. For lead times exceeding 36 hours, GenCast's accuracy rate rose to 99.8%, indicating a substantial improvement in medium-range predictions.
Beyond general weather, GenCast has shown advanced capabilities in forecasting extreme events, including tropical cyclones, high wind speeds, and temperature extremes. A single 15-day forecast can be generated in just eight minutes using a Google Cloud TPU, a stark contrast to the hours required by traditional physics-based models on supercomputers. However, it's noted that these comparisons were often made against an older version of ENS, and AI models are expected to complement, rather than fully replace, existing meteorological tools.
Google's broader AI weather initiatives, including Weather Lab and other models like GraphCast and NeuralGCM, are aimed at enhancing disaster preparedness and supporting renewable energy planning. These AI-driven forecasts are also beginning to integrate into user experiences on platforms like Google Search and Maps. The company continues to collaborate with weather agencies to refine these technologies and ensure their responsible deployment for global benefit.