For decades, researchers from a wide range of scientific disciplines have sought to develop better predictive models that can anticipate future changes. future weather conditions with increasing precision and for several days or even weeks in advance.
In recent years, advances in artificial intelligence have changed the way we approach the classic question: “Will it rain tomorrow?”, making it easier to make predictions based on Machine learning and historical data, which, while capable of providing a reasonable degree of accuracy, does not respond as effectively to unexpected events.
Now, Google researchers have developed what they say is the best weather forecasting model to date, which combines machine learning with more traditional data analysis techniques (big data), which could lead to accurate forecasts at a fraction of the current cost.
The model, which was named NeuronGCMaims to bridge the gap that has widened among experts in recent years, dividing them between those who favor more intensive use of AI and those who instead advocate “classical analysis” based on real-time empirical data.
Blurring the boundaries
As we noted at the outset, while new machine learning methods that predict the weather based on historical data are extremely fast and really effective, they often create problems with long-term forecasts.
On the other hand, models “general circulation of the atmosphere” which have dominated meteorology for the past 50 years, use complex equations to model changes in the atmosphere and provide accurate forecasts, but they are extremely slow and expensive to implement. So what does Google offer?
NeuralGCM still uses a traditional approach to calculate some of the large atmospheric changes needed for forecasting. Then it turns on artificial intelligence, which often works where larger models fail. usually for forecasts at scales less than 25 kilometersfor example, those related to cloud formations or regional microclimates.
The result, the researchers say, is a model that can produce better forecasts faster, using less computing power. So much so that they claim that NeuralGCM is already as accurate in its forecasts over a period of 1 to 15 days as European Centre for Medium-Range Weather Forecasts (ECMWF), research related organization.
In the medium term, the big promise of this model is not so much to compete with local weather forecasts, but to work on predicting major climate events at a scale that would be completely prohibitive using traditional methods. This includes everything from being able to predict the impact of hurricanes in advance to modeling the more complex climate changes that may play out in the coming years.
Google says NeuralGCM will be an open-source model of interest not only to meteorologists, but also to commodity traders, agricultural planners, insurance companies, and more.