Are AI weather models learning atmospheric physics? A sensitivity analysis of cyclone Xynthia

Abstract

Artificial Intelligence (AI) weather models are explored for initial condition sensitivity studies to analyze the physicality of the relationships learned. Gradients (or sensitivities) of the target metric of interest are computed with respect to the variable fields at initial time by means of the backpropagation algorithm, which does not assume linear perturbation growth. Here, sensitivities from an AI model at 36-h lead time were compared to those produced by an adjoint of a dynamical model for an extreme weather event, cyclone Xynthia, presenting very similar structures and with the evolved perturbations leading to similar impacts. This demonstrates the ability of the AI weather model to learn physically meaningful spatio-temporal links between atmospheric processes. These findings should enable researchers to conduct initial condition studies in minutes, potentially at lead times into the non-linear regime (typically >5 days), with important applications in observing network design and the study of atmospheric dynamics.