Few things please us more than knowing the weather forecast. Meteorology is a science that obsesses millions of people, and over the years, meteorologists have been improving their accuracy… until now.
Google DeepMind, the artificial intelligence arm of the search giant, has just announced a new weather prediction model that outperforms traditional systems in over 90% of cases.
Dubbed as GraphCast, this machine learning model promises 10-day forecasts that are superior, quicker, and more energy-efficient than the tools currently powering your weather app.
“We believe this marks a turning point in weather prediction,” write the Google researchers in a study published this Tuesday.
A new, cheaper, and more accurate model: AI serving science
The current prediction model is generally referred to as “numerical weather prediction (NWP).” NWP involves inputting current meteorological conditions into vast models that simulate forthcoming changes based on the principles of fluid dynamics, thermodynamics, and other atmospheric sciences. It’s complex, expensive, and demands significant computational power.
In contrast, GraphCast breaks tradition by emphasizing historical data instead of simulating how molecules fly and collide. In other words, it’s a machine learning model that predicts based on past occurrences. While there’s a lot of computer science involved, it’s generally simpler in terms of the level and number of calculations required.
GraphCast starts with the current state of weather on Earth and data from six hours ago. It then predicts the weather six hours ahead. Subsequently, GraphCast reintroduces these predictions into the model, conducts the same calculations, and issues longer-term forecasts.
Google’s team compared GraphCast’s results with the current medium-term weather prediction model, known as HRES. According to the study, GraphCast “significantly” outperformed HRES in 90% of the test objectives.
Surprisingly, GraphCast also excelled in predicting extreme weather phenomena, such as tropical cyclones and sudden temperature changes, despite not being specifically trained for these scenarios.
The study’s authors emphasize that their work is intended to complement the standard systems relied upon by meteorologists. “Our approach should not be seen as a substitute for traditional methods of weather forecasting,” the study authors wrote.
