A061-0008
Using Neural Networks and Code Optimization to Accelerate Radiation Computations for Dynamical Models
Using Neural Networks and Code Optimization to Accelerate Radiation Computations for Dynamical Models
Wednesday, 9 December 2020
Poster
Abstract:
Top-down machine learning approaches to physical parameterizations, such as the emulation of the entire radiation scheme of dynamical models by using a neural network, have met challenges related to physical realism, numerical stability and generalization. In this work we explore a more targeted approach, where neural networks are used to replace the computations of the optical properties of gaseous clear-sky atmosphere, but the radiative transfer is computed using traditional methods based on physical equations.
To generate training data, we collected input data spanning a wide range of atmospheric conditions and greenhouse concentrations and then fed these to a recently developed scheme, RRTM for GCM applications - Parallel (RRTMGP). Neural networks were trained to predict the long-wave optical depth and Planck emission and then implemented in the original Fortran codebase of RRTMGP. Depending on the software and hardware platform, we found a speed-up of 1-4 by using neural networks instead of the original interpolation routine in RRTMGP. To exploit the faster gas optics computations, we also refactored the radiative solver RTE (Radiative Transfer for Energetics) which can be used together with RRTMGP to compute radiative fluxes and heating rates for dynamical models or other applications. As a result, the computation of long-wave fluxes is 2-3 times faster when additionally using neural networks.
A thorough evaluation of our scheme, RRTMGP-NN, show that the errors in fluxes and heating rates are very similar to the original scheme with respect to accurate line-by-line radiation computations. This is the case for both present-day and future climate scenarios, indicating that a targeted machine learning approach can result in a highly accurate and generalizable scheme at the cost of some speed-up compared with top-down approaches.
To generate training data, we collected input data spanning a wide range of atmospheric conditions and greenhouse concentrations and then fed these to a recently developed scheme, RRTM for GCM applications - Parallel (RRTMGP). Neural networks were trained to predict the long-wave optical depth and Planck emission and then implemented in the original Fortran codebase of RRTMGP. Depending on the software and hardware platform, we found a speed-up of 1-4 by using neural networks instead of the original interpolation routine in RRTMGP. To exploit the faster gas optics computations, we also refactored the radiative solver RTE (Radiative Transfer for Energetics) which can be used together with RRTMGP to compute radiative fluxes and heating rates for dynamical models or other applications. As a result, the computation of long-wave fluxes is 2-3 times faster when additionally using neural networks.
A thorough evaluation of our scheme, RRTMGP-NN, show that the errors in fluxes and heating rates are very similar to the original scheme with respect to accurate line-by-line radiation computations. This is the case for both present-day and future climate scenarios, indicating that a targeted machine learning approach can result in a highly accurate and generalizable scheme at the cost of some speed-up compared with top-down approaches.