Stability of NN Emulations of Long- and Short-wave Radiation Parameterizations in a GCM

Abstract:

Stability of the coupling between deterministic and statistical components is one of the major challenges in development of ML/AI-based parameterizations for multi-dimensional non-linear environmental models. In theory, stable generalization to out-of-sample data is not guaranteed by nonlinear ML tools like NNs because nonlinear extrapolation is an ill-posed problem. Thus, in practice, previously unseen inputs may lead to unphysical outputs of the NN-based parameterization, often destabilizing the hybrid model. In our experiments, we demonstrate that an NN emulator thoroughly trained using a sufficiently representative training set can be integrated in the model without becoming unstable and is robust with respect to significant changes in the model. An NN-based emulation of radiative transfer parameterizations for a state-of-the-art GCM (NCEP Global Forecast System based on the FV3 dynamical core) remains stable after significant structural and parametric changes to the host deterministic model (including change of dynamical core , resolution, multiple atmospheric physics parameterizations etc.), and allows for production of plausible medium-range weather forecasts and seasonal predictions.