Highs of AI: DeepMind says its weather model beats leading systems

Can AI predict the weather more quickly and accurately than the forecasting field’s top supercomputers?

That’s the contention of a new paper from Google DeepMind, in which the AI-focused research unit introduces a new weather projection model trained on decades of historical weather data across hundreds of variables.

The team claimed this method, called GraphCast, outperforms existing prediction systems across more than 90% of 1,380 measures and can better account for the types of extreme weather events that are growing more common amid the climate crisis.

The findings come as weather data is becoming a priority for businesses trying to gird their operations for increasingly erratic climate patterns. DeepMind is far from the first team to attempt to apply AI and historical data to the problem. But the researchers claim GraphCast sets a new benchmark for accuracy and speed of calculation, the latter of which reduces the resources needed to produce a computation.

“It’s definitely an inflection point in weather forecasting,” Rémi Lam, staff research scientist at DeepMind and a lead author on the paper, said in a press call. “It shows the extent to which machine learning and AI can really deal in and really process extremely complex physical phenomena.”

Sunny with a chance of algorithms: Most modern weather forecasting is based around sets of physics-based governing equations combined with data on weather conditions, which are then crunched by powerful supercomputers. While these techniques are “a triumph of science and engineering,” the authors wrote, incorporating historical data into the calculus is “not straightforward” and can be resource-intensive.

GraphCast uses reams of historical data, observations, and analysis to fashion deep learning algorithms and capture “patterns in the data which are not easily represented in explicit equations,” the paper said.

GraphCast produces forecasts of up to 10 days in advance in less than a minute, and can better account for extreme events like cyclones, atmospheric rivers, and intense heat, according to the authors. DeepMind is open-sourcing the model’s code, which is trained on 39 years of historical data.

“Revolution” in the forecast: Matthew Chantry, machine learning coordinator at the European Centre for Medium-Range Weather Forecasts, said in the press call that thinking around AI and weather forecasting has undergone a transformation in the last year or so. Researchers had previously looked for ways to supplement physics models with AI. But projects like GraphCast demonstrate how deep learning might be used to predict weather directly through algorithms molded by neural nets and data.

“What we’ve seen with GraphCast is a significant step forward in the abilities of these models, such that they now have, in our eyes—we’ve done some assessments—some legitimate claims [to] equal and sometimes outperform the physical models for deterministic forecasts.

“There’s probably more work to be done to create reliable operational products, but this is likely the beginning of a revolution,” he added.