Robustness of Neural Network Emulations of Radiative Transfer Parameterizations in a State-of-the-Art General Circulation Model

TL;DR

This paper demonstrates that neural network emulators of radiative transfer in a modern climate model remain stable and produce realistic results despite significant model changes, highlighting their robustness and potential for climate modeling.

Contribution

It shows that decade-old neural network emulators are robust to structural and parametric changes in a state-of-the-art GCM, ensuring stability and realism in climate simulations.

Findings

Neural network emulators remain stable in new model configurations.

Emulators produce realistic radiative transfer outputs.

Architectural and training choices influence stability.

Abstract

The ability of Machine-Learning (ML) based model components to generalize to the previously unseen inputs, and the resulting stability of the models that use these components, has been receiving a lot of recent attention, especially when it comes to ML-based parameterizations. At the same time, ML-based emulators of existing parameterizations can be stable, accurate, and fast when used in the model they were specifically designed for. In this work we show that shallow-neural-network-based emulators of radiative transfer parameterizations developed almost a decade ago for a state-of-the-art GCM are robust with respect to the substantial structural and parametric change in the host model: when used in the AMIP-like experiment with the new model, they not only remain stable, but generate realistic output. Aspects of neural network architecture and training set design potentially contributing to stability of ML-based model components are discussed.