Physics‐informed ensemble forecasts in data‐driven models

Abstract Ensemble forecasts promote the productive shift toward quantifying prediction uncertainties, a trend increasingly embraced by current artificial intelligence (AI) models. However, most skillful data‐driven ensemble forecast models rely on probabilistic neural architectures to directly learn the future state distribution, usually overlooking the role of initial uncertainties. This oversight is partly due to the limited sensitivity of AI models to random perturbations. Recent studies indicate initial perturbations with specific spatial structures can evolve similarly in both AI and dynamic models. Inspired by this, the variational inference–conditional nonlinear optimal perturbation (VI–CNOP) method is applied in the FuXi model to investigate typhoon track ensemble forecasts. This approach involves solving a physics‐informed nonlinear optimization problem. Numerical results show VI–CNOP optimally estimates the distribution of growing‐type errors, enabling efficient sampling of more representative members than other widely used initial perturbation generation methods. Further diagnosis shows the evolutions of initial perturbations derived from VI–CNOP lead to the potential vorticity anomalies, which in turn modify the intensity and spatial extent of the Western Pacific Subtropical High, causing the uncertainties of typhoon trajectories. This investigation illustrates that introducing physics‐informed initial perturbations in AI models not only yields higher forecast performances and meaningful spread–skill relationships, but also enhances the interpretability of AI models and understanding of perturbation evolutions.