Instability-Aware Perturbations of Extreme Events in an AI Weather Foundation Model

Extreme weather events are intensifying under climate change, yet recent advances in weather prediction operate within a forecast-only paradigm that does not directly mitigate impacts once an extreme event is anticipated. Motivated by chaos control theory, we explore whether small, instability-aware perturbations can leverage intrinsic atmospheric sensitivity to influence extreme weather evolution within an AI-based forecasting framework. We use the Aurora foundation model and identify dynamically sensitive perturbation locations using Finite-Time Lyapunov Exponent (FTLE) diagnostics. To implement a physically interpretable intervention compatible with foundation models, we introduce an idealized cloud seeding based perturbation operator that mimics condensation-driven latent heat release applied in the lower–mid troposphere. In a case study, these upstream perturbations induce coherent downstream changes in integrated vapor transport, leading to reduced peak landfall intensity and slower precipitation accumulation. These results demonstrate that instability-aware perturbations within an AI foundation model can induce dynamically meaningful downstream impacts, providing a first step toward bridging chaos control concepts and data-driven weather prediction.