For many years, improving weather forecasting models has been a priority for scientists worldwide. Accurate weather forecasts are essential for choosing appropriate attire. They’re also key for the transport, agriculture, and energy sectors. More importantly, better weather predictions can help mitigate significant risks, including the potential loss of life.
Google’s new tool. Google DeepMind, the company’s division focused on artificial intelligence, aims to revolutionize the meteorology field with GenCast. On Wednesday, the company published a study in Nature to introduce its new model to the world.
The new open-source tool is an AI-based model designed to predict weather and assess risks associated with extreme weather conditions. Google says its new AI model is comparable to the most advanced meteorological systems available today.
15-day forecasts. GenCast is particularly interesting for its ability to make medium-term forecasts. According to Google, its new model maintains accuracy even 15 days in advance, a timeframe where uncertainties typically challenge contemporary models.
Probabilistic models. Medium-term forecast models generate predictions based on different scenarios. The results are then weighted to create a probabilistic forecast, assigning different probabilities to the various scenarios predicted by the model.
However, this approach has a major drawback. It requires substantial computational resources, which means a supercomputer must spend hours performing these calculations. According to Ferran Alet Puig, a senior researcher at Google DeepMind and co-author of the study, processing the latest atmospheric data and obtaining results can take about two hours.
From supercomputers to AI. Google’s new weather forecasting model doesn’t rely on theoretical models based on the laws of physics. Instead, it draws from observations. Its creators used historical weather records up to 2018, which allowed the model to adjust and “learn” from this data.
Alet Puig told the Spanish science news site SINC, “Machine learning models like GenCast operate very differently from classical models. The [probabilistic forecast system] ENS of the European Centre for Medium-Range Weather Forecasts essentially simulates the laws of physics using supercomputers. While we believe we understand the laws of fluids in theory, in practice, we face sensor errors and limitations in computing power. Many parameters in the models remain unknown.”
In contrast, GenCast requires significantly less computing power and time, producing results in just minutes.
Testing the model. As mentioned earlier, the meteorological records up to 2018 were used to train the model. Meanwhile, the model was then tested with later data, yielding satisfactory results.
Is this a revolution or evolution? While Google’s new approach is innovative, it’s still too early to abandon traditional models completely. Although the methodology is cutting-edge, the model still depends on established meteorological methods.
Experts point out that GenCast continues to rely on the traditional integrated forecasting system model for initial conditions and for training the machine learning algorithms. Hybrid systems combining both approaches might also soon be available. This dual methodology could allow scientists to use models based on physical equations to verify, train, and improve the system. Meanwhile, it would also enable models based on observations to optimize prediction calculations.
Image | Brian McGowan
Related | The Best Weather Apps for Your Phone
View 0 comments