The effects of climate change, as we have seen with DANA in Letur and the province of Valencia, make It is essential to have improvements in weather forecasting with sufficient advance notice to take disaster prevention measures. Scientists and meteorological agencies from various countries are committed to this, and they are already capable of making fairly accurate predictions several days ahead. Unfortunately, their predictions are not as accurate as they would like, and it is very difficult for them to make forecasts more than a few days in advance.
Precisely what Google DeepMind wants to collaborate with the help of its AI model GenCastwhich they have presented through a publication in the journal Nature. It is a high-resolution model developed to improve weather prediction, as well as extreme events that the most prominent weather prediction system, the ENS of the ECMWF (European Center for Medium-Range Weather Forecasts), with an advance of up to 15 days .
DeepMind has already confirmed that they are going to release the model code, as well as all its related data and details, so that the research community dedicated to weather prediction can use it in their tasks.
GenCast is an evolution of a deterministic DeepMind prediction model, which offered the best possible estimate of future weather. But GenCast produces a weather forecast that includes a minimum of 50 predictions. Each one of them represents a possible evolution and trajectory of time.
It is a diffusion model, but instead of using the same system as those of its type used in image, video or music generation; It is adapted to the spherical geometry of the Earth, and learns to generate the complex probability distribution of future weather scenarios when given the most recent weather as input.
GenCast Training and Accuracy
For GenCast training, Google DeepMind has used data from four decades of historical weather archivesoriginating from the ECMWF ERA5 archive. Among these data are variables such as temperature, wind speed and pressure at various altitudes. The model learned global weather patterns, with a resolution of 0.25º, directly from this processed data.
When rigorously evaluating GenCast’s performance, the researchers who developed it trained it with historical weather data up to 2018, and tested it with data from 2019. GenCast managed to make a more accurate weather prediction than ENS of ECMWF, the main pension system. This system is the basis for many countries and regions to make weather decisions around the world on a daily basis.
To test both systems, the researchers reviewed predictions of different variables at different key moments. In total, they worked with 1,320 combinations. GenCast turned out to be more accurate than ENS in 97.2% of cases, and in 99.8% for predictions over 36 hours.
The model requires a single TPU v5 from Google Cloud, and in 8 minutes it is capable of generating information and forecasts for the next 15 days. It is also capable of generating all the forecasts of the set, in parallel. To give us an idea of the magnitude of this advance, to obtain physics-based forecasts, such as those produced by the ENS, with a resolution of 0.1º or 0.2º, it is necessary for a supercomputer to work for several hours , employing tens of thousands of processors.
Another field in which GenCast obtained better results than ENS was in the prediction of extreme weather events. Among them, abnormally high or low temperatures, tropical cyclones, typhoons or hurricanes. In the case of these last three phenomena, GenCast made more precise forecasts than the ENS of its trajectory.
GenCast is part of Google’s suite of next-generation AI-based weather models. These include AI-based medium-range deterministic forecasting models from DeepMind, Neural GCM from Google Research and SEEDs, as well as several models that have been developed to predict floods.
These models are beginning to be incorporated into the user experiences of Google’s search engine, as well as its Maps application. They are also improving prediction of precipitation, wildfires, floods and extreme heat. But from DeepMind they want to focus their work as a cooperation between AI and traditional meteorologyand have confirmed that they will continue to work with meteorological agencies in the development of AI-based methods that improve predictions.
