A specialized artificial intelligence system can predict cyclone trajectories more than 30 hours in advance without sacrificing the accuracy of established methods.
Predicting where a cyclone will pass usually relies on several weather models. These calculate how the atmosphere will evolve based on the laws of physics. Forecasters then compare their results and account for their uncertainties. This method requires extensive computing, especially when many possible scenarios must be produced.
Image Wikimedia
The system presented in Nature takes a different approach. It learns from vast collections of weather data and cyclone archives. It can then produce numerous scenarios describing the possible evolution of the atmosphere and the cyclone's future position. These forecasts extend up to 15 days.
The main benefit comes from the time saved. With comparable trajectory accuracy, the system reaches certain quality levels more than 30 hours earlier than reference methods. In other words, information about a cyclone's probable path can become sufficiently reliable more than a day earlier.
This head start matters directly for threatened regions. A few dozen additional hours can provide more time to prepare evacuations, protect infrastructure, or organize emergency response. A trajectory nevertheless remains a probabilistic forecast: several paths remain possible, especially several days before the event arrives.
To account for this uncertainty, the model does not calculate just one evolution of the atmosphere. It generates a set of scenarios. If many of them follow a similar route, forecasters have a more robust signal. If they diverge widely, that spread instead indicates that the trajectory remains uncertain.
The system has already moved beyond the laboratory stage. Google DeepMind worked with the US National Hurricane Center to provide its forecasts to specialists during one hurricane season. According to a NOAA presentation devoted to this experiment, the data served as operational guidance for forecasters.
Artificial intelligence models do not solve every difficulty, however. Earlier studies show that they can perform very well at tracking a cyclone's path but be less reliable at forecasting its strength. Maximum wind speed and its rapid evolution remain particularly difficult to reproduce accurately.
The next step is therefore to evaluate these forecasts across more cyclones and seasons, especially for rapid changes in intensity. Their operational use will also depend on the reliability of the probabilities provided, since warning decisions rely as much on uncertainty as on the most likely trajectory.