weather forecasts

AI already is making inroads into the global weather forecasting business, more accurately and quickly predicting patterns around the world and, in turn, making it easier and faster to get alerts and warnings of severe weather out.

MIT Technology Review earlier this year took a look at a new 3D and high-resolution AI model from Huawei called Pangu-Weather that can predict weekly global weather patterns that are comparable to traditional computer simulations but much faster, and another deep-learning algorithm used to more accurately predict extreme rainfall with more notice.

In July, Nvidia wrote about ForeCastNet, its AI-based weather-prediction model that is as accurate as numerical weather prediction methods with “orders of magnitude greater speed and energy efficiency.” The Weather Company – currently owned by IBM, though a sale to private equity firm Francisco Partners is expected to close next year – also in July outlined how AI is used in weather forecasting and the benefits that come with it.

Now Comes GraphCast

Those are only a handful of examples. Now the folks at Google’s DeepMind AI research lab are rolling out GraphCast, a new AI-based forecasting model that the company boasts can deliver highly accurate 10-day weather predictions in less than a minute. It outperforms the European Centre for Medium-Range Weather Forecasts’ (ECMWF) High Resolution Forecast (HRES), “the industry gold-standard weather simulation system,” Remi Lam, staff research scientist at DeepMind, wrote in a blog post.

The center is running a live experiment with GraphCast on its website.

“GraphCast takes a significant step forward in AI for weather prediction, offering more accurate and efficient forecasts, and opening paths to support decision-making critical to the needs of our industries and societies,” Lam wrote. “And, by open sourcing the model code for GraphCast, we are enabling scientists and forecasters around the world to benefit billions of people in their everyday lives.”

Data, Not Equations

GraphCast, which uses machine learning and Graph Neural Networks, differs from numerical weather prediction by using data rather than physical equations to create a forecast system.

Numerical weather prediction “begins with carefully defined physics equations, which are then translated into computer algorithms run on supercomputers,” Lam wrote. “While this traditional approach has been a triumph of science and engineering, designing the equations and algorithms is time-consuming and requires deep expertise, as well as costly compute resources to make accurate predictions.

Google used four decades of weather reanalysis data from ECMWF’s ERA5 data set to train the deep-learning model, teaching what he described as the “cause and effect relationships that govern how Earth’s weather evolves, from the present into the future.” The data came from radar, weather stations and satellite images, and the scientists used numerical weather prediction to fill in gaps in the observation data.

According to Google, the model creates a forecast using more than a million grid points that cover the entire surface of the Earth, with each point predicting temperature, wind speed and direction, and mean sea-level pressure at the surface and such variables as humidity, temperature, and wind speed and direction at 37 levels of altitude. The model uses two points of Earths weather – the current time and six hours earlier – to predict what the weather will be six hours later, according to a Google research paper published in Science.

Accurate Predictions in Less Than a Minute

Based on this data, GraphCast on a single Google’s TPU v4 supercomputer can make 10-day forecasts in less than a minute. A similar 10-day forecast can take hours in a supercomputer with hundreds of machines, Lam wrote.

GraphCast’s capabilities go beyond weather forecasts, he wrote. The model “can also identify severe weather events earlier than traditional forecasting models, despite not having been trained to look for them. This is a prime example of how GraphCast could help with preparedness to save lives and reduce the impact of storms and extreme weather on communities.”

For example, Google can apply a cyclone tracker onto GraphCast forecasts to more accurately predict the movement of a cyclone. In the case of Hurricane Lee, which was moving up the Atlantic in September, a GraphCast model running on the ECMWF website accurately said the hurricane would make landfall in Nova Scotia nine days before it happened. Traditional forecasts predicted Nova Scotia would be where the hurricane would hit six days before it moved on shore.

In addition, the scientists said GraphCast can be used to forecast other complex situations, from climate and energy to agriculture and human and biological activity.

GraphCast is the latest piece in Google’s growing family of weather prediction systems that includes a regional Nowcasting model for predicting weather conditions 90 minutes ahead and MetNet-3, which is used in the United States and Europe for regional 24-hour forecasts.