GenCast, a new AI model from Google DeepMind, is accurate enough to compete with traditional weather forecasting. In tests using 2019 data, it outperformed a leading forecasting model, according to recently published research. 🌦️✨

AI won't replace traditional meteorology anytime soon, but it can add valuable tools to predicting the weather and alerting the public about severe storms. GenCast is one of several AI models in development which could lead to more accurate forecasts. 🤖⛈️

“The climate affects virtually every aspect of our lives… and it is also one of the biggest scientific challenges predict it,” says Ilan Price, a senior scientist at DeepMind. “Google DeepMind is on a mission to advancing AI for the benefit of humanity. I think this is an important way, a key contribution in that regard.” 🌈📊

Price and his team conducted tests with GenCast against the ENS system, one of the world's most advanced forecasting models, managed by the European Centre for Medium-Range Weather Forecasts (ECMWF). GenCast outperformed ENS 97.2% of the time, according to the research. published this week in the magazine Nature. 📈🧐

GenCast is a machine learning-based weather prediction model trained on climate data from 1979 to 2018. This model learns to recognize patterns in four decades of historical data and uses that information to predict what might happen in the future. This is very different from how traditional models like the ENS work, which still rely on supercomputers to solve complex equations and simulate the physics of the atmosphere. Both GenCast and the ENS produce forecasts as a whole, which offer a variety of possible scenarios. 🌌🌪️

For example, when predicting the path of a tropical cyclone, GenCast can provide an average of 12 hours of advance warning. GenCast is generally better at predicting cyclone paths, extreme weather events, and wind energy production up to 15 days in advance. 🌪️⏳

One caveat is that GenCast was tested against an older version of the ENS, which now operates at a higher resolutionThe peer-reviewed research compares GenCast's predictions with ENS forecasts for 2019, analyzing how closely each model matched that year's actual conditions. According to ECMWF Machine Learning Coordinator Matt Chantry, the ENS system has improved significantly since 2019, making it difficult to assess how GenCast might compare to ENS today.

It's clear that resolution isn't the only important factor in making solid predictions. ENS was already working at a slightly higher resolution than GenCast in 2019, yet GenCast still managed to outperform it. DeepMind claims to have conducted similar studies with data from 2020 to 2022 and found similar results, though this hasn't been peer-reviewed. However, there was no data to compare for 2023, when ENS began operating at a significantly higher resolution. 📊📈

Dividing the world into a grid, GenCast operates at a 0.25-degree resolution, meaning each square on that grid measures a quarter of a degree of latitude by a quarter of a degree of longitude. By comparison, the ENS used a 0.2-degree resolution in 2019 and is now at 0.1 degrees. 🌍🗺️

However, the development of GenCast “marks a significant milestone in the evolution of weather forecasting,” Chantry said in an emailed statement. Alongside ENS, ECMWF is also running its own version of a machine learning systemChantry comments that he “takes some inspiration from GenCast.” 🌟🤓

Speed is an advantage for GenCast. It can generate a 15-day forecast in just eight minutes using a single Google Cloud TPU v5. Physics-based models like ENS could take several hours to do the same. GenCast skips all the equations ENS has to solve, which explains why it requires less time and computing power to produce a forecast. ⚡️💻

“In computational terms, it is magnitudes more expensive run traditional forecasts against a model like GenCast,” Price says. 💰📉

That efficiency could help alleviate some concerns about the environmental impact of AI data centers, which have already contributed to Google's rising greenhouse gas emissions in recent years. However, it's difficult to assess how GenCast compares to physics-based models in terms of sustainability without knowing how much energy is used to train the machine learning model. 🏭♻️

There are still improvements GenCast can make, including the ability to scale to higher resolution. Additionally, GenCast issues predictions at 12-hour intervals, compared to traditional models, which typically issue predictions at shorter intervals. This may impact how these forecasts can be used in the real world (e.g., to assess how much wind power will be available).

“You’d want to know how the wind will behave 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's growing interest in how AI can be used to improve forecasts, it has yet to prove its effectiveness. “People are looking at it. I don't think the meteorological community as a whole is convinced by it,” Mullens says. “We're trained scientists who think in terms of physics… and because AI isn't fundamentally physics, there's still an element where we're asking, 'Is this good? And why?'”

Forecasters can try GenCast for themselves; DeepMind published the code of its open source model. Price says he sees GenCast and others improved AI models being used in the real world alongside traditional models. “Once these models are in the hands of practitioners, it builds more trust and confidence,” Price says. “We really want this to have a broad societal impact.” 🌍💡