EGU General Assembly 2020
© Author(s) 2020. This work is distributed under
the Creative Commons Attribution 4.0 License.

Predicting atmospheric optical properties for radiative transfer computations using neural networks

Menno Veerman1, Robert Pincus2,3, Caspar van Leeuwen4, Damian Podareanu4, Robin Stoffer1, and Chiel van Heerwaarden1
Menno Veerman et al.
  • 1Wageningen University and Research, Meteorology and Air Quality Group, Wageningen, The Netherlands
  • 2NOAA/Earth System Research Lab, Physical Sciences Division, Boulder, Colorado, USA
  • 3Cooperative Institute for Environmental Studies, University of Colorado, Boulder, Colorado, USA
  • 4SURFsara, Amsterdam, the Netherlands

A fast and accurate treatment of radiation in meteorological models is essential for high quality simulations of the atmosphere. Despite our good understanding of the processes governing the transfer of radiation, full radiative transfer solvers are computationally extremely expensive. In this study, we use machine learning to accelerate the optical properties calculations of the Rapid Radiative Transfer Models for General circulation model applications - Parallel (RRTMGP). These optical properties control the absorption, scattering and emission of radiation within each grid cell. We train multiple neural networks that get as input the pressure, temperature and concentrations of water vapour and ozone of each grid cell and together predict all 224 or 256 quadrature points of each optical property. All networks are multilayer perceptrons and we test various network sizes to assess the trade-off between the accuracy of a neural network and its computational costs. We train two different sets of neural networks. The first set (generic) is trained for a wide range of atmospheric conditions, based on the profiles chosen by the Radiative Forcing Model Intercomparison Project (RFMIP). The second set (case-specific) is trained only for the range in temperature, pressure and moisture found in one large-eddy simulation based on a case with shallow convection over a vegetated surface. This case-specific set is used to explore the possible performance gains of case-specific tuning.

Most neural networks are able to predict the optical properties with high accuracy. Using a network with 2 hidden layers of 64 neurons, predicted optical depths in the longwave spectrum are highly accurate (R2 > 0.99). Similar accuracies are achieved for the other optical properties. Subsequently, we take a set of 100 atmospheric profiles and calculate profiles of longwave and shortwave radiative fluxes based on the optical properties predicted by the neural networks. Compared to fluxes based on the optical properties computed by RRTMGP, the downwelling longwave fluxes have errors within 0.5 W m-2 (<1%) and an average error of -0.011 W m-2 at the surface. The downwelling shortwave fluxes have an average error of -0.0013 W m-2 at the surface. Using Intel’s Math Kernel Library’s (MKL) BLAS routines to accelerate matrix multiplications, our implementation of the neural networks in RRTMGP is about 4 times faster than the original optical properties calculations. It can thus be concluded that neural networks are able to emulate the calculation of optical properties with high accuracy and computational speed.

How to cite: Veerman, M., Pincus, R., van Leeuwen, C., Podareanu, D., Stoffer, R., and van Heerwaarden, C.: Predicting atmospheric optical properties for radiative transfer computations using neural networks, EGU General Assembly 2020, Online, 4–8 May 2020, EGU2020-5574,, 2020

Comments on the presentation

AC: Author Comment | CC: Community Comment | Report abuse

Presentation version 2 – uploaded on 06 May 2020 , no comments
Added link to preprint of corresponding submitted paper
Presentation version 1 – uploaded on 04 May 2020
  • CC1: Comment on EGU2020-5574, Andrew Gettelman, 06 May 2020

    Very interesting presentation. Two questions:

    1. Do you think your results that the networks are better for the cases they are trained on is a general result? I.e. better training yields more accurate networks which can be faster? 

    2. Is the speedup on the whole cost of RRTMGP or just the optics part?

    • AC1: Reply to CC1, menno veerman, 06 May 2020

      Dear Andrew,

      Thank you for the questions. 
      1. We think this is a general result. By training for a single case we can reduce the complexity of the problem for the neural networks, so the networks will be more accurate. Consequently, we can use smaller networks, which are faster 
      2. This is the speedup of the gas optics and Planck source function computations. 

      If you would like to read further, we just published a preprint of the paper corresponding to this paper: