DeepMind AI outperforms top weather forecasts, with one caveat

AI company Google DeepMind says AI can more accurately predict the weather 10 days out than current state-of-the-art simulations, but meteorologists are still trying to build weather models based on actual physical principles. It warns against abandoning and relying solely on patterns in the data. Expose the shortcomings of AI approaches.

Existing weather forecasts are based on mathematical models, which use physics and powerful supercomputers to definitively predict what will happen in the future. These models have gradually become more accurate by adding more details, but this requires more calculations, more powerful computers, and higher energy demands.

Rémi Lam and his colleagues at Google DeepMind took a different approach. The company’s GraphCast AI model is trained on his 40 years of historical weather data from satellites, radar, and ground-based measurements to identify patterns that even Google DeepMind can’t understand. “As with many machine learning AI models, it’s not that easy to interpret how the model works,” Lamb says.

To make the predictions, actual weather measurements taken at two points six hours apart from more than one million locations around the globe are used to predict the weather six hours into the future. These predictions can be used as input for another round that predicts another 6 hours into the future.

DeepMind researchers carried out this process using data Produces a 10-day forecast from the European Center for Medium-Range Weather Forecasts (ECMWF). They say it outperformed ECMWF’s “gold standard” High Resolution Forecasting (HRES) by providing more accurate forecasts on more than 90 per cent of the data points tested. At some altitudes, this accuracy increased to 99.7%.

Matthew Chantry He, who worked with Google DeepMind at ECMWF, said his organization had previously seen AI as a tool to complement existing mathematical models, but in the past 18 months it has seen a shift in the way AI can actually provide predictions on its own. He said that he has become able to do so.

“We at ECMWF believe this is a very exciting technology that has the potential not only to reduce energy costs when making forecasts, but also to improve them. Creating a reliable operational product “Probably more work is needed, but this is likely the beginning of a revolution in the way weather forecasts are made, and this is our assessment,” he says. According to Google DeepMind, using GraphCast to make a 10-day forecast takes him less than a minute on a high-end PC, but with HRES it can take several hours on a supercomputer.

But some meteorologists are wary of trusting weather forecasting to AI. Ian Renfrew According to researchers at the University of East Anglia in the UK, GraphCast currently lack the ability to marshal data into its own starting state, a process known as data assimilation. In traditional predictions, this data is carefully incorporated into simulations after thorough checks on physics and chemistry calculations to ensure accuracy and consistency. Currently, GraphCast must use a starting state prepared in the same way by ECMWF’s own tools.

“Google won’t be doing weather forecasts any time soon because they can’t assimilate the data,” Renfrew said. “And data assimilation is typically one-half to two-thirds of the computation time for these forecasting systems.”

He says there are also concerns about completely abandoning deterministic models based on chemistry and physics and relying solely on AI output.

“Even if you have the best predictive model in the world, what’s the point if the public doesn’t trust you and you don’t take action? We ordered the evacuation of 30 miles of Florida’s coastline. “If nothing happens, it will blow away the trust that has been built over decades,” he says. “The advantage of a deterministic model is that you can investigate it. If you get a bad prediction, you can investigate why that prediction is bad and target those aspects for improvement.”