Extreme Rapid Intensification of Hurricanes Otis (2023) and Patricia (2015): Machine Learning Diagnoses

Hurricane Otis (2023) presents a unique event for which all current operational hurricane models failed to predict its rapid intensification (RI) cycle after cycle. In contrast, the RI onset of Hurricane Patricia (2015), which occurred almost within the same month and location in the Eastern Pacific basin, was much more predictable across operational models. The fact that the previous generation of hurricane models could successfully capture the more extreme RI of Hurricane Patricia while more up-to-dated models could not predict RI for Otis apparently highlights some missing ingredients in Hurricane Otis that we wish to explore. Using different architectures of classification and deep learning models, we show that it is possible to find an effective combination of environmental features that control the probability of RI onset for both Hurricanes Otis and Patricia. Specifically, by trying various groups of environmental features along Otis’s track, we found that it is the combination of Otis’s bigger storm size, stronger vertical wind shear, and slower storm moving speed that is the main cause for the failure of Otis’s RI prediction in the model. Decreasing Otis’s size and increasing its movement in a weaker shear environment could help improve its RI onset prediction significantly. These results suggest that the failure of operational hurricane models may be due to the large-scale flow inherited from global models when imposing on the larger size of Otis. Our approach presents a new diagnostic approach for RI prediction based on machine learning, as well as a different way to understand RI onset variability beyond the traditional ensemble modelling methods.