Keep up with the innovative tech transforming business
Tech Brew keeps business leaders up-to-date on the latest innovations, automation advances, policy shifts, and more, so they can make informed decisions about tech.
Google DeepMind has found a new way of using generative AI to more accurately model weather predictions, and it could take the forecasting world by storm.
The search giant’s research arm recently unveiled GenCast, an AI model that researchers say can beat the world’s leading forecasting system on projections up to 15 days in advance, including everyday weather and extreme weather events.
The release marks the latest milestone in the evolution of AI-based weather forecasting, which researchers hope can push the science beyond the limits of traditional physical simulations. Doing so could be especially important at a time when the climate crisis has made accurate predictions tougher but also more critical to decision-makers.
Beating its own record: GenCast improves upon a previous model that Google DeepMind released about a year ago that the company also said could outperform leading forecasting computers. Unlike that last model, which provided a single best guess of the weather, GenCast offers something called an ensemble forecast, a set of 50 or more possible weather scenarios.
It relies on a diffusion model, the same technique that’s led to breakthroughs in image, video, and audio generators. In this case, Google researchers wrote in the paper, the projection it generates is tuned to “the spherical geometry of the Earth” and ingests the most recent weather data as a prompt. The model is also trained on historical weather information from 1979 to 2019.
Edging out supercomputers: Google put GenCast to the test by pitting it against the European Centre for Medium-Range Weather Forecasts (ECMWF)’s ensemble system (ENS), which it says is the top ensemble system operating today. The ECMWF and other leading forecasters use supercomputers to crunch complex physics-based equations based on current weather data, a process that Google claims is slower and more resource-intensive.
Google said it tested GenCast against ENS on 1,320 different combinations of variables and lead times. GenCast outperformed ENS in 97.2% of those cases, with the figure increasing to 99.8% when lead time was above a day and a half, the researchers wrote.
Google is releasing GenCast as an open model with code and model weights freely available.
“We are eager to engage with the wider weather community, including academic researchers, meteorologists, data scientists, renewable energy companies, and organizations focused on food security and disaster response,” authors Ilan Price and Matthew Willson wrote in the blog post.