Graphs showing the vertical profiles of the error in shortwave downwelling flux, upwelling flux, and heating rates computed from fluxes.
Vertical profiles of the error in shortwave downwelling flux (left column), upwelling flux (middle column) and heating rates computed from fluxes (right column), based on a year of global reanalysis data that was not used for training. Results are shown for three different emulation methods: replacing the radiation scheme with (a) a dense (feed-forward) neural network (NN) or (b) a bidirectional recurrent NN, or (c) replacing the gas optics with NNs. The solid and dotted lines show the mean error and mean absolute error, respectively, while the shaded area indicates the 5th and 95th percentile of differences. For NN-RadScheme (a) the mean heating rate errors at top-of-atmosphere (0.01 Pa) reach around 3.5 K*day-1 (the x-axis has been cropped). Credit: Ukkonen [2022], Figure 6 (a, b, d)
Editors’ Highlights are summaries of recent papers by AGU’s journal editors.
Source: Journal of Advances in Modeling Earth Systems

The use of machine learning to represent sub-grid processes is increasingly being explored as a way of reducing the uncertainty and computational expense of large-scale models. To shed light on the best approaches, Ukkonen [2022] evaluates different ways of emulating a radiation scheme using machine learning.

Radiation differs from some other sub-grid processes in that it’s well-understood, but its computational expense has motivated attempts to replace the entire radiation code with a neural network. Past studies on emulating radiation, as well as other sub-grid physics, have traditionally taken the vertical profiles of relevant variables and concatenated them into one long input or output vector of a dense neural network. In this approach, the number of inputs and outputs, and hence the vertical resolution, must be fixed. A recurrent neural network (RNN), in contrast, can be used to traverse through an atmospheric column sequentially layer by layer. When applied to shortwave radiation, a method based on bidirectional RNNs, whose structure was inspired by physical radiative transfer equations, improved the accuracy by an order of magnitude compared to a dense network that used an order of magnitude more parameters. If RNNs prove effective for other processes, the smaller dimensionality may be crucial in allowing machine-learned parameterizations to generalize.

Another way of using machine learning for radiation is to keep the radiative transfer equations but replace the gas optics – a more data-driven component of radiation schemes – with a neural network. This approach did not sacrifice accuracy, and still gave a meaningful speedup. The author’s research presents a clear example of how machine learning can be combined with physical modeling and domain knowledge to improve the prediction of sub-grid processes.

Citation: Ukkonen, P. (2022). Exploring pathways to more accurate machine learning emulation of atmospheric radiative transfer. Journal of Advances in Modeling Earth Systems, 14, e2021MS002875.

—Jiwen Fan, Editor, Journal of Advances in Modeling Earth Systems

Text © 2022. The authors. CC BY-NC-ND 3.0
Except where otherwise noted, images are subject to copyright. Any reuse without express permission from the copyright owner is prohibited.