When Helene hit Florida earlier this year, 234 people lost their lives in the worst hurricane to hit the US mainland since Katarina in 2005. It’s natural disasters like this, and their increasing intensity due to climate change, which has pushed scientists to develop more accurate weather forecasting systems. On Wednesday, Google’s DeepMind division announced what could be considered the most significant advance in the field in nearly eight decades of work.
In a post on Google Keywords BlogDeepMind’s Ilan Price and Matthew Wilson detailed GenCast, the company’s latest AI agent. According to DeepMind, GenCast is not only better at providing daily and extreme weather forecasts than its previous AI weather program, but it also outperforms the best forecasting system in use at the moment, the one run by the European Center for Medium-Range Weather Forecasts (ECMWF ). In tests comparing the two systems’ 15-day forecasts for weather in 2019, GenCast was, on average, more accurate than ECMWF’s ENS system 97.2% of the time. With turnaround times exceeding 36 hours, DeepMind proved to be even more accurate, at 99.8%.
“I’m a little reluctant to say this, but it’s like we’ve made decades of improvements in one year,” Rémi Lam, the chief scientist of DeepMind’s former AI weather program, said The New York Times. “We’re seeing really, really rapid progress.”
GenCast is a diffusion model, which is the same technology that powers Google’s generative AI tools. DeepMind trained the software on nearly 40 years of high-quality weather data curated by the European Center for Medium-Range Weather Forecasts. The forecasts generated by the new model are probabilistic, in the sense that they take into account a series of possibilities which are then expressed as a percentage. Probabilistic models are considered more nuanced and useful than their deterministic counterparts, which only offer the best guess of what the weather might be like on any given day. The former is also more difficult to create and calculate.
Indeed, what is perhaps most striking about GenCast is that it requires significantly less computing power than traditional physics-based ensemble forecasts such as ENS. According to Google, only one of its own TPU Tensor Processing Unit v5 can produce a 15-day GenCast forecast in eight minutes. In contrast, it may take a supercomputer with tens of thousands of processor hours to produce a physics-based prediction.
Of course, GenCast isn’t perfect. One area where the software could provide better predictions is the intensity of hurricanes, the DeepMind team reported The times he was confident he could find solutions to the agent’s current shortcomings. Meanwhile, Google is making GenCast an open model, with sample code for the tool available on GitHub. GenCast forecasts will soon arrive on Google Earth.
