Reliable coarse-grained turbulent simulations through combined offline learning and neural emulation

In this preprint, M²LInES postdoc Chris Pedersen, along with Pavel Perezhogin, Joan Bruna, and Laure Zanna, show that the standard offline learning paradigm used to produce ML models of unresolved dynamics can be complemented with some learned time evolution of the system, and produce more stable hybrid physics and ML models.