ARPGEM-ANEMONE: a physical atmosphere model coupled to a Graph Neural Network-based surface ocean emulator

Atmosphere-only climate simulations require sea surface temperature (SST) and sea ice concentrations (SIC) as lower boundary conditions. These are typically prescribed from observations, which is cheap but prevents atmosphere-ocean interactions, or provided by a physical ocean model, which preserves interactions but it comes with a significant computational cost. We present ANEMONE, a Graph Neural Network emulator trained on ERA-5 to predict daily SST and SIC given atmospheric forcings. Evaluated offline over twelve monthly-initialized one-year rollouts, ANEMONE outperforms persistence beyond approximately 5 days and produces bounded errors that saturate near the climatological baseline. For sea ice extent, ANEMONE outperforms climatology at all lead times in both hemispheres. We couple ANEMONE online to the physical atmosphere model ARP-GEM, forming a closed feedback loop: ARP-GEM provides daily accumulated forcings, ANEMONE returns updated SST and SIC. The coupled system runs stably for at least one year at $\sim$100 km resolution with negligible computational overhead. Compared with prescribed SST and SIC configurations, the coupled model exhibits additional biases, especially over the Pacific, consistent with the emulator's surface-only ocean representation. This demonstrates that a Neural Network-based ocean-surface emulator can serve as an interactive boundary condition generator inside a physical atmosphere model, opening a pathway between prescribed forcing and full ocean coupling.