On Wednesday, Google unveiled its latest advancement in weather prediction technology with the introduction of GenCast, an artificial intelligence (AI) model developed by its AI research branch, Google DeepMind. The company has also released a research paper that outlines the model’s capabilities, particularly in medium-range weather forecasting. According to Google, GenCast has demonstrated a superior performance compared to existing leading forecasting models, particularly in resolution and accuracy. The model is capable of predicting weather conditions up to 15 days in advance with a resolution of 0.25 degrees Celsius.
Features of Google GenCast
In a blog post, Google DeepMind presented insights into the capabilities of this high-resolution AI ensemble model. It noted that GenCast not only forecasts day-to-day weather but also predicts extreme weather events. The model outperformed the established Ensemble (ENS) system from the European Centre for Medium-Range Weather Forecasts (ECMWF). Results from these comparisons were published in the journal Nature.
Unlike traditional deterministic forecasting methods, which yield a single specific forecast based on defined initial conditions, GenCast utilizes a probabilistic approach. Deterministic models rely on precise physical and chemical equations to generate their forecasts. In contrast, probabilistic models, such as GenCast, offer multiple potential outcomes by simulating a variety of initial conditions and parameters, a technique known as ensemble forecasting.
Google DeepMind emphasized that GenCast operates as a diffusion model that adjusts to the Earth’s spherical geometry, allowing it to produce a complex probability distribution of future weather scenarios. The training of this AI model involved four decades of historical weather data sourced from the ECMWF’s ERA5 archive, enabling the model to learn global weather patterns at a resolution of 0.25 degrees Celsius.
In its published research, Google assessed GenCast’s effectiveness by training the model with historical data until 2018 and subsequently challenging it to make predictions for the year 2019. The researchers analyzed 1,320 combinations of diverse variables across different time frames and discovered that GenCast surpassed the ENS system in accuracy for 97.2 percent of the targets, achieving an impressive 99.8 percent accuracy for predictions made more than 36 hours in advance.
Furthermore, Google DeepMind announced its intention to provide the GenCast AI model’s code, weights, and predictions to foster collaboration and advancement within the weather forecasting community.
