Improving Tropical Cyclone Track Forecast Skill Through Assimilating Target Observation Achieved by AI-Based Conditional Nonlinear Optimal Perturbation

Abstract

The artificial intelligence (AI)-based weather forecasting model named FuXi and its data assimilation (DA) system FuXi-En4DVar has been developed for high-efficiently forecasting high-impact weather events such as tropical cyclones (TCs). Besides conventional observations, target observations are essential to further improve initial field accuracy and then increasing high-impact weather event forecasting skills. The identification of the sensitive area, where the additional observations should be deployed, is the key to implementing target observations. In this paper, a sensitive area identification system is established for the FuXi model on the basis of FuXi-En4DVar, based on the fully nonlinear method of conditional nonlinear optimal perturbation (CNOP). The CNOP represents the optimally growing initial perturbation and can be calculated by using the adjoint of numerical models in numerical forecast models, but in the AI-based FuXi model, it is solved by directly using the automatic differential algorithm embedded in the FuXi model. Such an approach of calculating CNOP significantly increases the computational efficiency. Applying this system to the forecasts of 11 TCs demonstrates that the additional target observations can significantly improve TC track forecast skills, as compared with the other additional observations. Moreover, a small number of additional target observations can be expected to achieve the forecast skill comparable to, or even surpassing to, that obtained by tens of times more observations. This validation shows the potential of applying dynamical CNOP to AI-based model for highly effectively identifying the sensitive area for target observations associated with TC forecasting.