and train both together to get the best of worlds: Respecting physical laws and simulation of known dynamics conventionally, but correcting this dynamical core towards data with neural networks. A climate model that can learn from data. Sure, it's not perfect, for example

it is not a coupled system: Sea surface temperatures are boundary conditions and not dynamically simulated. Advancing this methods towards coupled systems would be a next milestone. But it shows how far you can get when your model can learn from data without relearning everything