Robustness of neural network emulations of radiative transfer parameterizations in a state-of-the-art general circulation model

Abstract. The ability of machine-learning-based (ML-based) model components to generalize the previously unseen inputs and its impact on stability models that use these have been receiving a lot recent attention, especially in context ML-based parameterizations. At same time, emulators existing physically based parameterizations can be stable, accurate, fast when used they were specifically designed for. In this work we show shallow-neural-network-based radiative transfer developed almost decade ago for state-of-the-art general circulation (GCM) are robust with respect substantial structural parametric change host model: two 7-month-long experiments new GCM, remain stable generate realistic output. We concentrate aspect emulators' performance discuss features neural network architecture training set design potentially contributing robustness components.