Shedding new light on radiative transfer: Deep learning for multispectral emulation of the RRTM

Radiative transfer (RT) is a computationally expensive part of numerical weather prediction (NWP). This has driven the development of parameterizations such as the Rapid Radiative Transfer Model (RRTM; Mlawer et al. 1997), which is a hybrid physical/statistical approximation of the full line-by-line models, and the RRTM for General Circulation Models (RRTMG; Pincus and Stevens 2013), which is a lightweight version of the RRTM. However, even with these relatively efficient parameterizations, RT accounts for ~20% of the computing in a modern NWP model. This has driven efforts to emulate parameterizations such as the RRTM(G) with machine learning – especially neural networks (NN; see review by Krasnopolsky and Belochitski 2020) – which are yet more computationally efficient, often by a factor of 1000 or more. However, NN emulation has traditionally focused on broadband quantities – where the electromagnetic spectrum is partitioned into a shortwave band (wavelengths <~ 4 μm) and a longwave band (>~ 4 μm) and the emulator predicts one set of target variables (heating rates and fluxes) for each. Modern NWP models, such as version 16 of the Global Forecast System (GFSv16), require a multispectral RT solution, with the spectrum divided into 30 wavelengths instead of two wide bands. We have developed a multispectral NN-based emulator of the RRTM, building on the broadband emulator presented in Lagerquist et al. (2023, henceforth L23). This represents one of the first efforts at multispectral RT emulation.

The evolution from a broadband emulator to a multispectral one involved several challenges; we will discuss these challenges and our strategies to overcome them. First, the target space has 3870 quantities to predict (heating rates at 30 wavelengths x 127 vertical levels, surface downwelling flux at 30 wavelengths, and top-of-atmosphere upwelling flux at 30 wavelengths) – and some of these, which we call “trivial quantities,” are always zero. To prevent the NN from trying to learn complex relationships for trivial quantities, we applied a mask inside the NN architecture. Without masking, the NN struggled to predict exactly zero for trivial quantities; furthermore, the NN’s attempt to learn complex relationships for trivial quantities diverted “attention” away from non-trivial quantities, leading to an overall degradation of performance. Second, quantities in the target space have vastly different ranges – e.g., fluxes at the 1.27-μm wavelength can reach 300 W m^-2, while fluxes at 55.56 μm barely reach 1 W m^-2. To ensure skillful predictions for all target variables, we used a custom loss function that includes a weighting scheme to emphasize those variables with a smaller range. Third, in progressing from a broadband emulator to a multispectral one, we did not want broadband performance to suffer. In other words, we wanted the multispectral emulator to predict broadband quantities as skillfully as the emulator in L23. To address this challenge, we included a custom term in the loss function that focuses on shortwave quantities (summed over the 14 shortest wavelengths) and one for longwave quantities (summed over the 16 longest wavelengths).

When evaluated on shortwave quantities, the multispectral NN performs nearly as well as the broadband NN of L23. And interestingly, when evaluated on longwave quantities, the multispectral NN solidly outperforms the broadband NN of L23. We will discuss possible reasons for this interesting result and show that, furthermore, the multispectral NN performs skillfully for nearly all of the 3870 target variables.

Krasnopolsky, V., and Belochitski, A., 2020: “Using machine learning for model physics: An overview.” arXiv e-prints, 2002 (00416), URL https://arxiv.org/abs/2002.00416.

Lagerquist, R., D. Turner, I. Ebert-Uphoff, and J. Stewart, 2023: “Estimating full longwave and shortwave radiative transfer with neural networks of varying complexity.” Journal of Atmospheric and Oceanic Technology, 40 (11), 1407-1432, URL https://doi.org/10.1175/JTECH-D-23-0012.1.

Mlawer, E., S. Taubman, P. Brown, M. Iacono, and S. Clough, 1997: “Radiative transfer for inhomogeneous atmospheres: RRTM, a validated correlated-k model for the longwave.” Journal of Geophysical Research: Atmospheres, 102 (D14), 16 663–16 682, URL https://doi.org/10.1029/97JD00237.

Pincus, R., and B. Stevens, 2013: “Paths to accuracy for radiation parameterizations in atmospheric models.” Journal of Advances in Modeling Earth Systems, 5 (2), 225–233, URL https://doi.org/10.1002/jame.20027.