In a world increasingly impacted by climate change and frequent extreme weather events, precise and reliable weather forecasting has become essential. As demonstrated during the recent DANA phenomenon in Spain, how this information is managed can profoundly affect our way of life and human safety.

Mid-term weather forecasting remains highly complex and challenging to perfect, as predictions become notably uncertain beyond just a few days in advance.

To address this, scientists and national and international meteorological agencies rely on probabilistic ensemble predictions—an established model that forecasts a series of likely weather scenarios. Such data proves far more useful than single predictions by offering decision-makers a comprehensive view of potential weather conditions and the likelihood of each scenario unfolding over the coming days and weeks.

With the goal of improving this information, a scientific team at Google DeepMind, the British artificial intelligence research and development company, has published a study in Nature introducing GenCast, a new high-resolution AI-based ensemble forecasting model.

“Machine learning models like GenCast work very differently from traditional models. The ENS (a probabilistic forecasting system) at the European Centre for Medium-Range Weather Forecasts (ECMWF) essentially simulates the laws of physics using supercomputers. In theory, we believe we know the laws governing fluid dynamics, but in practice, sensor errors and finite computational capacity create limitations. There are many model parameters we don’t fully understand,” explains Ferran Alet Puig, senior researcher at Google DeepMind and co-author of the study.

How Does This AI Model Work?

GenCast and other machine learning models learn directly from data. “They receive numerous examples in which they must predict tomorrow’s weather based on today’s conditions. In doing so, they learn relationships that account for computational, sensor, and parameterization limitations,” adds Alet Puig.

The researchers report that GenCast provides “more accurate forecasts, both for daily weather and extreme events, than the leading operational system, the ENS from ECMWF—a forecasting ensemble system that informs many national and local decisions daily—up to 15 days in advance.”

This enables GenCast to surpass the most effective traditional medium-range weather forecasting systems while also outperforming them in predicting extreme weather, tropical cyclone trajectories, and wind energy production.

“GenCast generates 50 forecasts, representing 50 potential scenarios. With these 50 futures, we can estimate all sorts of probabilities, from marginal ones (e.g., will there be a heatwave in Seville?) to joint probabilities (e.g., how much wind energy will Spain generate in three days?),” Alet Puig elaborates.

Álvaro Sánchez González, another co-author of the study, notes, “For the traditional ENS system, producing 50 forecasts takes about two hours once atmospheric analysis is available, using supercomputers with thousands of processors pushed to their technological limits.”

By contrast, GenCast can generate a forecast in eight minutes using a single hardware device not much larger than a standard computer. “This means that if we used 50 such devices (instead of thousands), we could produce all 50 forecasts in just eight minutes. This is significant not only because predictions are ready nearly two hours earlier, but because more than 50 forecasts could be generated, yielding even more precise probabilities,” Sánchez González continues.

Enhanced Forecasting Capability

To evaluate GenCast’s performance, the scientists trained it with historical weather data up to 2018 and tested it using 2019 data. The new model demonstrated superior forecasting capabilities compared to the ENS.

“We tested both systems exhaustively, analyzing forecasts for different variables at various lead times: 1,320 combinations in total. GenCast outperformed ENS in 97.2% of these targets, and in 99.8% for lead times exceeding 36 hours,” the researchers emphasize.

Forecasting Extreme Weather Events

When it comes to extreme weather—such as heatwaves or high winds—the new tool also appears to outperform ENS, offering potential to save lives, prevent damage, and reduce costs. “When we tested GenCast’s ability to predict extreme heat, cold, and high wind speeds, our model consistently surpassed ENS,” they note.

“Interestingly, while we trained GenCast to predict general weather, it performs exceptionally well with extreme events, like heatwaves or hurricane tracking. Moreover, our model estimates the probabilities of these events, providing countries with vital information to prepare for potential impacts,” says Alet Puig.

In the case of tropical cyclones—commonly known as hurricanes or typhoons—anticipating landfall locations offers “invaluable” benefits, as the new tool delivers superior predictions for these often-devastating storms.

Renewable Energy Planning

Another application experts propose for these predictions is renewable energy planning. For instance, improvements in wind energy forecasts directly enhance its reliability as a sustainable energy source and could accelerate its adoption. In a proof-of-concept experiment analyzing wind power predictions from global wind farms, GenCast outperformed ENS.

To foster broader collaboration and accelerate research and development within the meteorological and climate communities, the authors have made GenCast an open model, publishing its code and parameters. “Just as we did with our deterministic global medium-range weather prediction system,” they emphasize.

Communicating Forecasts to the Public

One major challenge in weather forecasting lies in the chaotic nature of the atmosphere, where small differences in the current state can lead to large-scale effects days later. “For one- or two-day forecasts, this isn’t a significant issue. However, as we move into medium-term forecasts, absolute certainty becomes impossible, and probabilities must come into play,” says Sánchez González.

Probabilistic systems like GenCast allow for calculating probabilities of nearly any event, from simple ones—such as the likelihood of a maximum temperature exceeding 30 degrees Celsius on Tuesday—to more advanced probabilities, like the chances of having more than three consecutive days next week where the maximum temperature exceeds 30 degrees and the minimum stays above 25 degrees.

The challenge, however, lies in presenting this probabilistic information intuitively. “When sharing a forecast with the public, probabilities must often be simplified, which can make the prediction seem less consistent. Beyond developing more accurate models like GenCast, we need to rethink how forecasts are presented to preserve as much information as possible when summarizing probabilistic predictions,” Sánchez González explains.

The team plans to release historical and real-time forecasts from GenCast and earlier models soon, enabling anyone to integrate this meteorological data into their own models and research projects.

“We look forward to collaborating with the broader meteorological community, including academic researchers, meteorologists, data scientists, renewable energy companies, and organizations focused on food security and disaster response. Such collaborations bring invaluable insights and constructive feedback, as well as significant commercial and non-commercial opportunities—all of which are crucial to our mission of applying these models for the benefit of humanity,” the Google DeepMind team concludes. (Agencia Sinc)