Computational cost and accuracy in calculating three-dimensional
radiative transfer: Results for new implementations of Monte Carlo and SHDOM
Robert Pincus and K. Franklin Evans
Journal of the Atmospheric Sciences Vol. 66, pages 3131-3146. doi:10.1175/2009JAS3137.1
This paper examines the trade-offs between computational cost and accuracy for two new state-of-the-art codes for computing three-dimensional radiative transfer: a community Monte Carlo model and a parallel implementation of the Spherical Harmonics Discrete Ordinates Method (SHDOM). Both codes are described and algorithmic choices are elaborated. Two prototype problems are considered: a domain filled with stratocumulus clouds and another containing scattered shallow cumulus, absorbing aerosols, and molecular scatterers. Calculations are performed for a range of resolutions and the relationships between accuracy and computational cost, measured by time-to-solution and memory use, are compared.
Monte Carlo accuracy depends primarily on the number of trajectories used in the integration. Monte Carlo estimates of intensity are computationally expensive and may be subject to large sampling noise from highly-peaked phase functions. This noise can be decreased using a range of variance reduction techniques but these techniques can compromise the excellent agreement between the true error and estimates obtained from unbiased calculations. SHDOM accuracy is controlled by both spatial and angular resolution; different output fields are sensitive to different aspects of this resolution, so that the optimum accuracy parameters depend on which quantities are desired as well as on the characteristics of the problem being solved. The accuracy of SHDOM must be assessed through convergence tests and all results from un-converged solutions may be biased.
SHDOM is more efficient (i.e. has lower error for a given computational cost) than Monte Carlo when computing pixel-by-pixel upwelling fluxes in the cumulus scene, while Monte Carlo is more efficient in computing flux divergence and downwelling flux in the stratocumulus scene, especially at higher accuracies. The two models are comparable for downwelling flux and flux divergence in cumulus and upwelling flux in stratocumulus. SHDOM is substantially more efficient when computing pixel-by-pixel intensity in multiple directions; the models are comparable when computing domain-average intensities. In some cases memory use, rather than computation time, may limit the resolution of SHDOM calculations.
Download a reprint (PDF format) or read the article on-line at the journal web site.
Copyright 2009 American Meteorological Society. Further reproduction or electronic distribution is not permitted.