New AI from Deepmind could help predict severe weather faster

• Deepmind’s GraphCast AI can produce an accurate 10-day weather forecast in less than a minute.
• The Google-owned company’s AI model is the first that can outperform traditional forecasters.
• Deepmind experts told the Financial Times that they were able to predict the landfall of Typhoon Lee three days before it made landfall.

An artificial intelligence model developed by Google-owned Deepmind has outperformed traditional weather forecasters for the first time, new research shows.

The peer-reviewed study, published in the journal Science, said Deepmind’s GraphCast AI model “significantly outperformed” the traditional weather forecasting system, which is run by the European Center for Medium-Range Weather Forecasts.

The AI ​​model can produce an accurate 10-day forecast in less than a minute, and represents a “turning point in weather forecasting,” according to the newspaper.

“Weather forecasting is one of the most challenging problems that humanity has been working on for a very long time,” Pushmeet Kohli, vice president of research at Deepmind, told Business Insider.

“If you look at what’s happened in the last few years in terms of climate change, this is a very important issue,” he said.

GraphCast is trained on 39 years of historical weather data from ECMWF. according to Financial Timesthe research showed that GraphCast’s three- to 10-day weather forecasts were more accurate than traditional forecasts.

But Deepmind cautioned that the model has some limitations compared to non-AI forecasts when dealing with the uncertainty inherent in long-term weather forecasts.

Speaking to the Financial Times, project engineers said GraphCast was able to predict where Hurricane Lee, a powerful storm that hit North America earlier this year, would make landfall three days earlier than traditional forecasters.

Matthew Chantry, a machine learning specialist at ECMWF, told Insider that GraphCast represents an “important and welcome step forward” for the industry.

Once trained on a wide range of historical data, models like GraphCast will prove much cheaper than current weather forecasting methods, which rely on powerful (and expensive) supercomputers, he told the Financial Times.

“We’re probably talking about a thousand times cheaper in terms of energy consumption,” he said. “This is a miraculous improvement.”