Steering Tropical Cyclones Using Small Perturbations in an AI Weather Model

Tropical cyclone (TC) trajectories are governed by large-scale steering flows with sensitive dependence on initial conditions, raising the question of whether targeted perturbations can induce track deviations. We present a case study within the Weather Jiu-Jitsu framework, which exploits dynamical sensitivity at critical junctures to nudge extreme events, by combining finite-time Lyapunov exponent (FTLE) diagnostics with idealized thermodynamic perturbations in an AI weather forecast model. Applied to Hurricane Sandy (2012) six days before landfall, FTLE-guided perturbations in the steering environment produce a track displacement of 322.9 km, 3.3x larger than randomly placed perturbations of equal seeding energy (mean 97.6 km). FTLE-guided sites in Sandy's deep-Caribbean steering environment generate a persistent positional offset of 36-50 km during the first five forecast days, which is amplified when Sandy recurves into the mid-latitude baroclinic zone at day 5. Three control experiments-random placement (97.6 km), backward FTLE targeting of the subtropical ridge (91.5 km), and forward FTLE computed at 700 hPa within the TC warm core (55.0 km)-suggest that the larger response is associated with targeting 500 hPa environmental steering-flow boundaries using forward FTLE guidance. A supplementary experiment applying FTLE-guided perturbations in the North Pacific 1 day before Sandy's recurvature produces 616.6 km, nearly double the Caribbean result, indicating that trajectory sensitivity concentrates near the recurvature gate rather than across the 7-day forecast window. The required perturbation magnitude exceeds current engineering capabilities, so the results represent a theoretical sensitivity analysis. These findings suggest that FTLE diagnostics may help identify steering-flow boundaries associated with enhanced downstream trajectory sensitivity.