Typhons : CausalVisionRI anticipates their rapid intensification at 24 hours thanks to a causal graph drawn from the observations

This deep model, guided by the cause-effect bonds discovered on 522 cyclones in the Northwest Pacific, better alerts the most dangerous phase than four reference systems.

Rapid intensification, when a tropical cyclone strengthens sharply in a matter of hours, is one of the most dangerous and difficult phases to predict. Its atmospheric, oceanic and structural causes are intertwined in observations. Researchers are proposing CausalVisionRI, a 24-hour warning system for the Pacific Northwest, which combines infrared satellite imagery, atmospheric and oceanic fields, and a causal graph.

AI constrained by physics

This graph was extracted from data from 522 cyclones that occurred between 2000 and 2017: it links atmospheric and ocean conditions, the state of the cyclone and the evolution of its intensity, with eight effects whose meaning has been verified. It serves as an auditable constraint during training. Tested over 2020-2024 (3,004 readings), the set of five models achieves a Brier skill score of 0.259, when the best statistical reference methods cap at 0.105, and detects more episodes with the same false alarm rate. Removing the graph loses part of the gain, which confirms its contribution.