Will Gregory et al. introduce FloeNet , a machine-learning emulator trained on the GFDL global sea ice model (SIS2) to reproduce key sea-ice and snow-on-sea-ice processes while conserving mass. The model emulates 6-hour tendencies related to ice and snow growth, melt, and advection. Trained on reanalysis-forced simulations, FloeNet was tested across different climate states, including pre-industrial and an increasing CO₂ scenario. It accurately reproduces sea-ice mean state, trends, and interannual variability, outperforming non-conservative approaches. FloeNet also captures the balance between thermodynamic and dynamic responses to forcing and reproduces coupling-related variables such as ice-surface temperature and ocean salt fluxes. These results suggest strong potential for improving the representation of polar processes in climate emulators.