Something like NeuralGCM, which combines a dynamical core with a learned subgrid physics module generalizes pretty well

www.nature.com/articles/s41...

Your intuition is correct: current AI forecast models do not perform well in climates that are different from the ones they were trained on. It would honestly be shocking if you could have a climate model just by training on reanalysis. Figures from Rackow et al. arxiv.org/pdf/2409.18529