A technology capable of anticipating the trajectory of a cyclone or the accuracy of a heatwave better than ever: this is the ambition of GenCast, a Google DeepMind AI model, now leading the way in weather forecasting.
GenCast stands out by generating not just one, but fifty simultaneous forecasts, each representing a weather scenario that is then compared to propose the most likely outcome. This "ensemble" approach allows for a nuanced understanding of the range of possibilities, a key innovation in a field where uncertainty dominates beyond a few days.
But how does the AI achieve such results? By combining the power of diffusion models, often used to generate images, with a specific adaptation to Earth's spherical geometry. Thanks to this structure, GenCast can map complex climatic evolutions with unprecedented accuracy.
This model was trained using four decades of data from the ERA5 archives of the European Centre for Medium-Range Weather Forecasts. Among the integrated variables: temperature, wind speed, and atmospheric pressure. These data enable the AI to learn major climate patterns and deliver detailed forecasts with a resolution of 0.25 degrees.
During testing, GenCast outperformed the ENS system, the current global benchmark, in 97.2% of cases, and even in 99.8% for timeframes exceeding 36 hours. In just eight minutes, the model produces a 15-day forecast using a single Google Cloud TPU processor, whereas traditional methods require supercomputers and hours of computation.
This data shows specific humidity and the projected trajectories of Typhoon Hagibis before it reached Japan. GenCast's forecasts, initialized one day or 15 days prior, remain sharp, though ensemble averages and perturbed states appear blurry due to uncertainty. Power spectra confirm a loss of resolution for these blurred states. The trajectories highlight significant long-term uncertainty, reduced as the event horizon approaches.
Beyond daily weather predictions, GenCast's capabilities prove crucial for extreme phenomena like cyclones and high winds. By anticipating these events more accurately, the AI could save lives and minimize material damage.
The potential impact extends to renewable energy as well. Better wind forecasts could optimize the use of wind turbines, thereby supporting the transition to clean and sustainable energy.
However, this progress does not entirely eliminate limitations. GenCast relies on past data, and its effectiveness in the context of a changing climate remains uncertain. Researchers stress the need for frequent updates to adapt to new climate dynamics.
By sharing GenCast's codes and data, Google hopes to accelerate innovation in this field. The goal? To merge artificial intelligence with traditional meteorology.
Article author: Cédric DEPOND