Google’s AI weather prediction model is pretty darn good

Google DeepMind’s new AI model, GenCast, is accurate enough to compete with traditional weather forecasting. According to recently published research, it outperformed leading predictive models when tested on 2019 data.

While AI won’t immediately replace traditional forecasting, it could lead to more tools being used to predict the weather and warn the public about severe storms. GenCast is one of several AI weather forecasters Model under development It can lead to more accurate predictions.

“Weather affects basically every aspect of our lives, and predicting it is one of the great scientific challenges,” said Ilan, a senior researcher at DeepMind. Price says. “At Google DeepMind, we have a mission to advance AI for the benefit of humanity, and I think this is one of the important ways, and one of the important contributions, on that front. ”

Price and his colleagues tested GenCast against the ENS system, one of the world’s best predictive models, run by the European Center for Medium-Range Weather Forecasts (ECMWF). According to our research, GenCast outperformed ENS 97.2% of the time. Published in this week’s magazine nature.

GenCast is a machine learning weather prediction model trained on weather data from 1979 to 2018. The model learns to recognize patterns in 40 years of historical data and uses them to predict what will happen in the future. This is very different from how traditional models like ENS operate, which still rely on supercomputers to solve complex equations to simulate the physics of the atmosphere. Both GenCast and ENS generate Ensemble predictionprovides various possible scenarios.

For example, when predicting the path of a tropical cyclone, GenCast was able to provide an additional 12 hours of advance warning on average. In general, GenCast was good at predicting cyclone paths, extreme weather events, and wind power generation up to 15 days in advance.

One caveat is that GenCast tested against an older version of ENS, so it now works at higher resolutions. The peer-reviewed study compares GenCast’s predictions with ENS’s predictions for 2019 to see how close each model came to real-world conditions that year. ECMWF Machine Learning Coordinator Matt Chantry said the ENS system has improved significantly since 2019. Therefore, it is difficult to judge how well GenCast can perform against ENS today.

Indeed, resolution is not the only important factor in making strong predictions. ENS was already running at a slightly higher resolution than GenCast in 2019, but GenCast was still able to outperform it. DeepMind says it conducted a similar study on data from 2020 to 2022 and found similar results, but it has not been peer-reviewed. However, there was no data available for comparison for 2023, when ENS began running at significantly higher resolution.

GenCast divides the world into a grid and operates at a resolution of 0.25 degrees. This means that each square on the grid is 4 degrees of latitude and 4 degrees of longitude. In contrast, ENS used a resolution of 0.2 degrees in 2019 and now has a resolution of 0.1 degrees.

Nevertheless, the development of GenCast “represents an important milestone in the evolution of weather forecasting,” Chantry said in an emailed statement. ECMWF, along with ENS, has its own version of machine learning system. Chantry says it is “inspired by GenCast.”

Speed ​​is an advantage for GenCast. Create a single 15-day forecast in just 8 minutes using a single Google Cloud TPU v5. Physics-based models such as ENS can take hours to do the same thing. GenCast bypasses all the equations that ENS has to solve, so it requires less time and computing power to generate predictions.

“Computationally, traditional predictions are orders of magnitude more expensive to perform than models like Gencast,” Price says.

This efficiency may alleviate some concerns about the environmental impact of energy-hungry AI data centers. AI data centers have already contributed to an increase in Google’s greenhouse gas emissions in recent years. But without knowing how much energy is used to train machine learning models, it’s difficult to determine how GenCast compares to physically-based models in terms of sustainability.

GenCast still has room for improvement, including the possibility of scaling up to higher resolutions. Additionally, GenCast issues predictions at 12-hour intervals, compared to traditional models that typically make predictions at shorter intervals. That could make a difference in how these predictions can be used in the real world, for example when assessing the amount of wind power available.

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

There is growing interest in how AI can be used to improve predictions, but the effectiveness of AI needs to be proven. “People are looking at it. I don’t think the entire meteorology community is buying or selling it,” Mullens says. “We are scientists trained to think in terms of physics… and AI is fundamentally not, so there are still parts where we are scratching our heads, is this a good thing? ?And why?

Forecasters can check GenCast themselves. DeepMind released: code For open source models. Price said he believes GenCast and better AI models are being used in the real world alongside traditional models. “Once these models are in the hands of experts, there is an added level of trust and confidence,” Price says. “We really want this to have a broad social impact.”