Google’s AI weather forecast model is pretty darn good

GenCast, a new AI model from Google DeepMind, is accurate enough to compete with traditional weather forecasts. According to a recently published study, it managed to outperform a leading forecasting model when tested using 2019 data.

AI won’t replace traditional forecasting any time soon, but it could expand the arsenal of tools used to predict the weather and warn the public about severe storms. GenCast is one of several AI weather forecasting models currently being developed that could lead to more accurate forecasts.

“Weather fundamentally affects every aspect of our lives… it is also one of the major scientific challenges to predict the weather,” says Ilan Price, senior research scientist at DeepMind. “Google DeepMind is on a mission to advance AI for the benefit of humanity. And I think that’s an important path, an important contribution on that front.”

Price and his colleagues tested GenCast with the ENS system, one of the world’s leading forecast models operated by the European Center for Medium-Range Weather Forecasts (ECMWF). GenCast outperformed ENS 97.2 percent of the time, according to a study published this week in the journal Nature.

GenCast is a machine learning weather forecast model trained on weather data from 1979 to 2018. The model learns to recognize patterns across four decades of historical data and uses these to make predictions about what might happen in the future. This is very different from how traditional models like ENS work, which still rely on supercomputers to solve complex equations and simulate the physics of the atmosphere. Both GenCast and ENS produce ensemble forecasts that offer a range of possible scenarios.

For example, when it comes to predicting the path of a tropical cyclone, GenCast was able to warn an additional 12 hours in advance on average. GenCast was generally better at predicting the impacts of cyclones, extreme weather and wind power production up to 15 days in advance.

One limitation is that GenCast tested itself with an older version of ENS, which now operates at a higher resolution. The peer-reviewed study compares the GenCast forecasts with the ENS forecasts for 2019 and shows how close each model came to real-world conditions that year. According to ECMWF Machine Learning Coordinator Matt Chantry, the ENS system has improved significantly since 2019. Therefore, it is difficult to say how well GenCast might perform against ENS today.

Of course, resolution isn’t the only important factor when it comes to making meaningful predictions. ENS was already working at a slightly higher resolution than GenCast in 2019 and GenCast still managed to surpass it. DeepMind says it has conducted similar studies using data from 2020 to 2022 and found similar results, although these have not been peer-reviewed. However, there was a lack of data to make comparisons for 2023, when ENS began running at a significantly higher resolution.

GenCast divides the world into a grid and works at a resolution of 0.25 degrees – meaning each square in that grid corresponds to a quarter degree of latitude and a quarter degree of longitude. In comparison, ENS used a resolution of 0.2 degrees in 2019 and is now at 0.1 degrees.

Still, the development of GenCast “represents a significant milestone in the evolution of weather forecasting,” Chantry said in an emailed statement. In addition to ENS, the ECMWF says it also operates its own version of a machine learning system. Chantry says it’s “somewhat inspired by GenCast.”

Speed ​​is an advantage for GenCast. With a single Google Cloud TPU v5, a 15-day forecast can be created in just eight minutes. Physics-based models like ENS may require several hours to do the same. GenCast bypasses all of the equations that ENS needs to solve, so creating a forecast requires less time and processing power.

“Computationally speaking, running traditional forecasts is orders of magnitude more expensive than a model like Gencast,” Price says.

This efficiency could allay some of the concerns about the environmental impact of energy-hungry AI data centers, which have already contributed to Google’s rise in greenhouse gas emissions in recent years. But it’s hard to figure out how GenCast compares to physics-based models in terms of sustainability without knowing how much energy goes into training the machine learning model.

There are still improvements that GenCast can make, including possible scaling to higher resolution. Additionally, GenCast produces forecasts at 12-hour intervals, compared to traditional models that typically do so at shorter intervals. This can make a difference in how these forecasts can be used in the real world (e.g. to estimate how much wind power will be available).

“You want to know what the wind is doing throughout the day, not just at 6 a.m. and 6 p.m.,” says Stephen Mullens, an assistant professor of meteorology at the University of Florida who was not involved in the GenCast research.

While there is growing interest in how AI can be used to improve forecasting, it has yet to be proven. “People look at it. I don’t think the meteorological community as a whole is bought and sold by this,” Mullens says. “We are trained scientists who think in physics terms… and because AI is fundamentally not that, there is still an element where we kind of turn our heads, ‘Is this good?’ And why?”

Forecasters can try GenCast for themselves; DeepMind has released the code for its open source model. Price says he sees GenCast and other enhanced AI models being used in the real world alongside traditional models. “Once these models get into the hands of practitioners, trust will be further increased,” says Price. “We really want this to have a widespread social impact.”