A Google AI model can outperform conventional weather forecasting up to days ahead of time, according to a peer-reviewed paper published in Science this week – and it does so at a fraction of the cost and in under a minute.

GraphCast is a machine learning-based approach to weather prediction that leans on historical weather data. It takes the current state of the Earth’s weather and the state six hours earlier, then outputs how the weather will be in six hours' time.

Where GraphCast differs from conventional weather forecasting is in the use of historical data.

According to the Science paper, most weather prediction is done by “solving the governing equations of weather using supercomputers” like the $2.3 billion one the UK announced it was building in 2020.

The scalability of adaptive numerical weather prediction has meant that, as computing power has increased, so has the accuracy of forecasting “to the point where the path of a hurricane can be predicted many days ahead”.

Where GraphCast excels is in its efficiency.

In under a minute, it can produce “an accurate 10-day forecast” using just one Google Cloud tensor processing unit (TPU) v4 chip. For reference, the hourly on-demand cost of a TPU v4 pod is just under $5.

This dramatically lowers the hardware requirements for making accurate predictions, so long as the training data is accurate.

Google’s researchers plugged 39 years of historical data (1979 to 2017 inclusive) into its model and used existing data from the remaining years to test its accuracy.

The results were surprising. Not only did it outperform results from the European Centre for Medium-Range Weather Forecasts (ECMWF) supercomputer, GraphCast was also able to accurately predict severe weather events like cyclones and extreme temperatures.

“These are key downstream applications for which GraphCast is not specifically trained, but which are very important for human activity,” the paper says.

The models can also be re-trained with recent data to keep up with things like climate change without using their utility or accuracy.

Overall, the researchers believe their machine learning system “marks a turning point in weather forecast” and will “strengthen the breadth of weather-dependent decision-making by individuals and industries, by making cheap prediction more accurate, more accessible, and suitable for specific applications”.

But it has limitations. GraphCast wasn't as accurate in predicting stratospheric conditions as its counterpart. AI models like GraphCast missed the strength of Hurricane Otis which struck southern Mexico with ferocity last month, killing more than 350 people.

Google’s researchers also noted that its weather prediction models are limited with greater uncertainty over longer periods of time, which can be especially important for capturing extreme events. They said that GraphCast is trained “to spatially blur its predictions in the presence of uncertainty” and that this is where ‘ensemble forecasting systems’ are better suited.

The application of AI to weather forecasting seems to be speeding up. Wired reported that the ECMWF is working on its own AI model that leverages its data and internal expertise.

GraphCast’s creators were modest in their appraisal of the model, saying it “should not be regarded as a replacement for traditional weather forecasting models” and instead proves that machine learning can “complement and improve” existing weather prediction methods.