Machine learning holds promise for improving the representation of sub-grid processes in weather and climate models, but this also faces many challenges such as generalization, stability and interpretability. Among specific sub-grid processes, emulation of atmospheric radiation codes has perhaps the longest history, motivated by the computational expense of radiation parameterizations. Radiation is also physically well-understood and therefore a good testbed to e.g. compare algorithms for physics emulation. Past studies on this topic have generally used dense neural networks (DNN) to emulate an entire radiation scheme. This has yielded impressive speed-ups, but at what cost? To date, no operational model uses an ML-based radiation scheme.

In this talk I present results from a recent study, showing that at least for shortwave radiation, recurrent neural networks (RNN) gives far better accuracy at a much lower model complexity than DNNs. The RNN approach resembles physical radiative transfer codes, which compute radiative flows through an atmospheric column sequentially layer by layer. Unlike DNNs, and like a physical parameterization, it can in principle be used with various vertical grids once trained.

A more targeted ML approach for radiation, the prediction of optical properties, is also briefly discussed. Similarly to RNN method of predicting radiative fluxes, it has the advantage of resulting in a a problem with much smaller dimensionality than predicting vertical columns of outputs. This in turn makes it much easier to generate representative training data for typical weather and climate modeling applications.