Do AI weather models learn atmospheric perturbation growth characteristics? A FuXi case study based on Breeding of Growing Modes method

Abstract To assess whether artificial intelligence (AI) weather models learn atmospheric perturbation growth dynamics, we compare the data-driven FuXi model with the dynamical Model for Prediction Across Scales (MPAS) using the Breeding of Growing Modes (BGM) method. MPAS rapidly captures baroclinic and convective instabilities, with perturbations saturating after approximately three days. In contrast, FuXi exhibits a pronounced "delayed growth" feature, with extremely slow growth during the initial two days. Although the saturated perturbations are broadly consistent with synoptic-scale systems in the mid-to-high latitudes, substantial non-physical noise persists in the low latitudes. Sensitivity experiments reveal that initial amplitude linearly determines saturation magnitude, while morphology governs spatial distribution, lacking the nonlinear coupling of numerical models. Perturbation kinetic energy spectra exhibit a "flat-tail" at high wavenumbers, indicating absent energy cascade and weak dissipation. Rotational/divergent decomposition confirms dominance of synoptic-scale rotational energy, failing to capture divergent motions. These findings suggest FuXi learns only partial statistical relationships from coarse-grained data rather than complete multiscale physical evolution, offering new insights for assessing AI weather model predictability and ensemble forecasting.