Beyond Track and Intensity: Scale-Dependent Diagnostics of a Tropical Cyclone Forecast by a Probabilistic AI Foundation Model, AERIS

Abstract

AI weather prediction models now rival physics-based systems on standard forecast scores, but whether their forecasts are physically consistent across scales remains unclear. We assess Argonne's Earth system foundation model (AERIS), a generative diffusion transformer, using Hurricane Laura (2020). Decomposing the storm-relative surface wind into azimuthal wavenumbers, we compare kinetic energy spectra and ensemble error growth against ERA5, the training dataset. AERIS reproduces the ERA5 spectrum across the resolvable range without spurious small-scale energy from the diffusion process. Ensemble uncertainty saturates first at the smallest scales while the axisymmetric vortex stays predictable through five days, as reported for conventional ensembles. Individual members retain realistic small-scale energetics even when their large-scale vortex departs from ERA5. AERIS, thus, appears to have learned behavior that is physically consistent across scales, rather than merely optimizing forecast scores.