AI-Guided Steering Diagnostics for Atmospheric Rivers and Tropical Cyclones

Could we steer the atmosphere's initial conditions to blunt an extreme event before it forms, and could the same sensitivity information tell us where to observe the atmosphere to forecast it better? Extreme atmospheric rivers and tropical cyclones cause some of the costliest weather disasters while their impact is sensitive to small changes in their track or landfall moisture. AI weather models now match operational numerical weather prediction on forecast accuracy at a small fraction of the computational cost, making it newly practical to test both questions. Using Aurora, we identify dynamically sensitive upstream regions with flow diagnostics and run experiments with localized thermodynamic perturbations at those sites. In two proof-of-concept case studies, a December 2022 California atmospheric river and 2012 Hurricane Sandy, targeted perturbations produce coherent downstream responses: landfall integrated vapor transport falls by up to 5.2%, and Sandy's simulated track shifts by 314—571km. Because the same diagnostics locate where a forecast is most sensitive to the atmospheric state, they point toward two climate-adaptation pathways: candidate sites for adaptive observations that sharpen landfall forecasts, and a first empirical step toward instability-aware intervention to steer weather extremes.