One of the first machine-learning frameworks to predict tropical-cyclone rapid intensification from satellite data, the moment a survivable storm becomes a catastrophic one. Built at NASA JPL; now part of operational practice.
On this page. Results summarize published, peer-reviewed work with NASA JPL. The forecast replay shown here is an illustrative recreation for the web, not the operational system.
A replay of the core result: where the operational baseline misses a rapid-intensification event, the ensemble tracks the ramp into Category 5, and quantifies the probability while there's still time to act.
Track prediction has steadily improved for decades. Intensity, and especially rapid intensification, a jump of 35 mph or more in 24 hours, has not. It's the failure mode behind the worst surprises: Hurricane Patricia went from Category 1 to Category 5 in a single day. The physics are nonlinear and the interactions are messy, so the operational models, built on largely linear assumptions, systematically under-call the ramp.
Most prior work also ignored the satellite's view of the storm's inner core, exactly where the signal of an imminent ramp hides.
I formulated ML.HIFS, a machine-learning framework that fuses refined satellite observations with the traditional predictors used by the National Hurricane Center. At its core is a multiplicative ensemble of linear and non-linear supervised models that both classifies whether an RI event will occur and quantifies its probability.
"The first application of machine learning to TC rapid-intensification forecasting, redefining the industry standard."
RI is rare, non-RI cases vastly outnumber it, which quietly trains a model to always predict "no ramp" and still look accurate. That's the exact failure that costs lives. I addressed it directly with an oversampling strategy inside the model's cost function, recalibrating the learning toward the minority RI class.
Skill was measured the way the field demands it, probability of detection, false-alarm ratio, the Peirce Skill Score, and the Brier Skill Score, and tuned to maximize genuine forecast skill, not just raw accuracy. A balanced training set turned out to be the single biggest lever on real-world RI performance.
Benchmarked against National Hurricane Center operations from 2009–2014, ML.HIFS was adopted by NASA JPL for operational testing and influenced methodology at the NHC, US Naval Research Lab, NOAA, and international space agencies. The work informed future NASA mission design and was referenced by the White House Office of Science & Technology Policy.