A new AI program predicts weather and storm paths more accurately than traditional models, offering faster and more reliable forecasts.

Google DeepMind’s new AI-based weather forecasting program, GenCast, has demonstrated an impressive ability to predict day-to-day weather and storm paths more accurately than existing models, including the European Centre for Medium-Range Weather Forecasts (ECMWF) widely-regarded ENS system. In a head-to-head comparison, GenCast outperformed the ENS forecast by up to 20%, marking a significant step forward in the use of AI for weather prediction.

While traditional weather forecasting relies on solving complex equations and using supercomputers to generate predictions, GenCast takes a new approach. Trained on 40 years of historical weather data, the AI system can predict weather changes with high accuracy over 15 days and in 12-hour intervals. It analyses a vast array of variables—such as wind speed, pressure, humidity, and temperature—to offer predictions at a granular level, down to 28km by 28km squares.

The AI’s speed and efficiency set it apart. While a traditional weather forecast can take hours to run on supercomputers, GenCast produces forecasts in just eight minutes using a single Google Cloud TPU—a chip designed specifically for machine learning. This speed makes it a promising tool for supporting existing forecasting systems, rather than replacing them outright.

Ilan Price, a research scientist at Google DeepMind, emphasised that this advancement represents a turning point in AI’s role in weather prediction. While the program is expected to complement traditional models in the short term, its ability to quickly generate multiple forecasts makes it an invaluable asset for predicting extreme events, such as hurricanes, cold fronts, and heat waves. This could assist not only weather forecasters but also energy companies, helping them better predict the output of wind farms and other renewable energy sources.

However, despite the impressive performance of GenCast, questions remain regarding its full potential. Critics like Sarah Dance, professor of data assimilation at the University of Reading, have pointed out that while the AI system excels in generating reliable predictions, it is not yet clear whether it can capture the “butterfly effect”—the rapid cascade of uncertainties that can impact forecast accuracy. Traditional models rely heavily on sophisticated maths and physical simulations to account for such uncertainties, and it remains to be seen whether machine learning can entirely replace these methods.

Nevertheless, the success of GenCast offers promising possibilities for the future of weather forecasting. With its ability to produce large ensembles of forecasts, it could provide more reliable estimates and better predict the outcomes of extreme weather events. Steven Ramsdale, the Met Office chief forecaster with responsibility for AI, described the advancement as “exciting,” while the ECMWF acknowledged the potential impact of GenCast and noted that elements of its approach were already being incorporated into their AI-based forecasts.

Google has been at the forefront of AI-driven weather prediction for several years, building on previous models such as NeuralGCM, which combined AI with traditional physics for long-range climate forecasting. The success of GenCast follows the 2023 introduction of GraphCast, which generates a single best-guess forecast and builds on it by creating a range of potential forecasts with probabilities for different weather events.

While the breakthroughs in AI-powered forecasting are encouraging, experts stress that it will take time before machine learning systems like GenCast can fully replace traditional, physics-based methods. There remains much to learn about how AI can account for the complexities of weather systems, and a long path ahead before its implementation can become ubiquitous in global forecasting.