Abstract
Sensitivity analysis is a cornerstone of climate science, essential forunderstanding phenomena ranging from storm intensity to long-term climatefeedbacks. However, computing these sensitivities using traditional physicalmodels is often prohibitively expensive in terms of both computation anddevelopment time. While modern AI-based generative models are orders ofmagnitude faster to evaluate, computing sensitivities with them remains asignificant bottleneck. This work addresses this challenge by applying theadjoint state method for calculating gradients in generative flow models. Weapply this method to the cBottle generative model, trained on ERA5 and ICONdata, to perform sensitivity analysis of any atmospheric variable with respectto sea surface temperatures. We quantitatively validate the computedsensitivities against the model's own outputs. Our results provide initialevidence that this approach can produce reliable gradients, reducing thecomputational cost of sensitivity analysis from weeks on a supercomputer with aphysical model to hours on a GPU, thereby simplifying a critical workflow inclimate science. The code can be found athttps://github.com/Kwartzl8/cbottle_adjoint_sensitivity.