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

Abstract

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 two seven month-long experiments 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.

Improved Weather Forecasting Using Neural Network Emulation for Radiation Parameterization

“…The development of a universal radiation emulator is a challenging topic because radiation emulators have fundamental uncertainty in relation to their dependency on training sets (Belochitski & Krasnopolsky, 2021;Boukabara et al, 2021). This study further investigates the possible impact of training sets on the current radiation emulator results.…”

“…We can see that the use of deep hidden layers or more complex structures (Liu et al, 2020;Pal et al, 2019) may not always produce better performance compared with an NN with a single hidden layer (in terms of speedup), although it offers a variety of possibilities for optimization. Belochitski and Krasnopolsky (2021) also noted that the risks of developing the DNN emulator associated with the control of complexity and nonlinearity.…”

“…For a clear sky, the number of input variables was decreased to 161, as cloud fraction profiles were excluded. Geographical data, such as longitude, latitude, and elevation, were not included as input variables in the RRTMG-K but were added to realistically reflect similar characteristics between horizontally and vertically adjacent grids, consistent with the previous studies (Belochitski & Krasnopolsky, 2021;Krasnopolsky et al, 2010. Surface elevation is an important variable that affects the actual vertical heights in relation to the terrain-following hybrid sigma pressure vertical coordinate of the WRF model (Skamarock et al, 2019).…”

“…The NN software based on a single hidden layer (Krasnopolsky, 2014) was utilized to develop the radiation emulator in this study. Previous radiation emulators (e.g., Belochitski & Krasnopolsky, 2021;Belochitski et al, 2011;Krasnopolsky et al, 2005Krasnopolsky et al, , 2010 have been developed using this software (or a similar version in older studies). In the present study, no tuning for hyperparameters (e.g., batch size, learning rate, activation function, regularization, normalization for inputs and outputs, and weight initialization) was performed, except for the number of neurons.…”

Effects of Cloud Microphysics on the Universal Performance of Neural Network Radiation Scheme

“…Despite the slowdown, accuracy improvement may not be sufficient enough because the uncertainty in microphysics variables can influence the stability of the emulator. Belochitski and Krasnopolsky (2021) (hereafter referred to as BK21) examined the robustness of a radiation emulator in the Global Forecast System (GFS) by applying training results based on the Climate Forecast System (CFS). They achieved stable results following the use of the radiation emulator after making numerous changes in the dynamical core, physics grids, boundary‐layer scheme, radiation scheme's version, trace gases, and CO 2 concentration of the GFS.…”