CREDIT-MESACLIP: stable emulation of high resolution coupled models

High-resolution output from MESACLIP highlights the importance of resolving ocean mesoscale behavior for process understanding and prediction of subseasonal-to-decadal variability. Here we train WxFormer, a state-of-the-art vision transformer available through the NSF NCAR CREDIT framework, to emulate native-resolution output from MESACLIP. We begin training with daily atmosphere and ocean surface data with monthly full depth ocean data, emulating high resolution 3D coupled ocean-atmosphere interactions. We further extend emulation to a 5-daily, uncoupled ocean on the same native MESACLIP grid and including the full 3-dimensional ocean. In both cases we achieve stable rollout. We discuss the role of normalization for emulating coupled systems with a range of timescales and report spectral statistics to evaluate the performance of our MESACLIP emulator to reproduce variability in the ocean and atmosphere across a range of scales.