Speeding up radiative transfer solvers with machine learning

Speeding up radiative transfer solvers with machine learning

Machine learning provides great opportunities for accelerating the representation of physical processes in weather and climate models. Radiative transfer solvers, which deal with the transport of solar and thermal radiation, are often the most costly parametrization in models. Speeding them up is one of the main ambitions of SLOCS.

In the very first machine learning paper of our group (study led by Menno Veerman, preprint at arXiv.org) we present our work on speeding up the RTE+RRTMGP radiative transfer solver. This project was a collaboration between our group, the SURF Open Innovation Lab, and the RTE+RRTMGP developers.

Roughly speaking, the computation of radiative fluxes can be split in two comparably expensive parts. First, the optical properties of the atmosphere (optical depth, single scattering albedo, and asymmetry factor) need to be computed as a function of the thermodynamics properties of the atmosphere and the concentration of gases, aerosols, and clouds particles. Subsequently, the radiative fluxes are computed for shortwave and longwave radiation.

Mean absolute error in a) radiative heating rate, b) top-of-atmosphere upwelling flux, and c) surface downwelling flux plotted against speedup for different network sizes. Blue is shortwave, red is longwave.

In our approach, we replaced the optical properties solver with a neural network-based solver. By doing so, the neural network does only local computations, and is therefore trainable with a limited set of data and is portable among grid configurations. In addition, by combining our method with the original RTE flux solver, we respect the radiative transfer equations exactly.

Often, we know a priori the range of atmospheric conditions that can be expected as model input. The width of this range determines the size and depth of the network that is needed to have errors sufficiently small. We trained two types of networks: one tailored for global numerical weather prediction, covering a wide range of conditions, and another for limited-area large-eddy simulation (tested for shallow convection over The Netherlands and for deep convection over the tropical ocean). All our networks are basic feedforward neural networks.

The figure shows the errors of the network plotted against speedup (of the optical properties solver) for the numerical weather prediction-tailored network. For most practical purposes the errors associated with the 5x speedup are acceptable. For the large-eddy simulation-tailored networks even greater speedups can be achieved in the optical properties solver, as the errors are smaller (see manuscript). For the complete radiative transfer computation, this means that we are approaching a 2x speedup.

Veerman, M.A., R. Pincus, R. Stoffer, C. van Leeuwen, D. Podareanu, and C.C. van Heerwaarden. Submitted to Philosophical Transactions A, https://arxiv.org/abs/2005.02265.