Google introduced its weather forecast artificial intelligence (AI) model Gencast on Wednesday. The AI model was developed by Google DeepMind, an AI research arm of the mountain view-based technology giant. The company’s researchers also published a paper on the technology that highlights its capabilities in medium-range weather forecasting. The company claims the system can outperform existing latest predictive models in terms of solutions and accuracy. It is worth noting that Gencast can make weather forecasts over the next 15 days with a resolution of 0.25 degrees Celsius.
Google Gencast Features
In a blog post, Google DeepMind details the new high-resolution AI ensemble model. It emphasizes that Gencast can predict daily weather and extreme events, thus claiming that it can perform better than Europe’s Medium Weather Forecasting Center (ECMWF) Overall (ENS) system. The performance of the model is now published in Nature Daily.
It is worth noting that Gencast does not use the traditional deterministic method to predict weather, but uses the probabilistic method. A weather prediction model based on a deterministic method produces a single, specific prediction for a given set of initial conditions and relies on precise equations of atmospheric physics and chemistry.
On the other hand, a probabilistic method-based model produces multiple possible results by simulating a series of initial conditions and model parameters. This is also called ensemble prediction.
Google DeepMind stresses that Gencast is a diffusion model that adapts to the Earth’s spherical geometry and produces complex probability distributions of future weather conditions. To train AI models, researchers provided for forty years of historical weather data from ECMWF’s ERA5 archive. In this way, the model is taught at 0.25 degrees Celsius (resolution 0.25).
In a published research paper, Google evaluated Gencast’s performance by training 2018’s historical data and then asking it to predict 2019. Using different lead times, a total of 1320 combinations of different variables were used, and the researchers found that Gencast believes that Gencast believes that 97.2% of these targets are more accurate than ENS, 99.8% in more than 36 hours of delivery time.
It is worth noting that Google DeepMind announced that it will release code, weights and forecasts for the Gencast AI model to support the weather forecast community.
