Targeted Observing for Long-term Tropical Cyclone Forecasts: An Observing System Simulation Experiments within a Data-driven Weather Model

Targeted observation is a cost-effective strategy that involves the optimal deployment of additional observations, which is particularly efficient for predictions of weather phenomena with intense spatiotemporal variation such as tropical cyclones (TCs). Traditionally, targeted observations rely on numerical weather prediction (NWP) models for identifying sensitive areas, ingesting observations, and conducting forward prediction. However, their limited forecast accuracy and high computational costs greatly hinder the timeliness of targeted observations, particularly for forecasts beyond 2 days. The recent development of fast and accurate artificial intelligence (AI) models offers a new possibility for long-term targeted observation. In this study, we calculated the sensitive areas for targeted observation of 5-day TC track forecasts using an AI-based conditional nonlinear optimal perturbation (CNOP) approach. Observation system simulation experiments (OSSE) are conducted to evaluate the impact of CNOP-based targeted observations in both AI and NWP frameworks. The analysis for 12 TCs demonstrates that CNOP-based sensitive areas effectively encompass the highly flow-dependent, dynamically crucial systems critical for long-term TC forecasting. AI-based OSSEs indicate that targeted observations within sensitive areas improve 5-day TC track forecast skills by 25%, exceeding those from non-sensitive areas. These targeted observations, when applied within NWP-based OSSEs, also demonstrate a 22% improvement in forecast skill. These results exhibit the benefit of targeted observations for long-term forecasts, and the cross-model validity of AI-based targeted observation strategy. Our study reveals a physical consistency in perturbation growth between AI and NWP models, providing the theoretical basis for applying AI models in long-term targeted observations, even with NWP systems.