D3D: Deep 3D scattering of solar radiation in the atmosphere due to clouds

The cloud radiative effect, the strong interaction between cloud and radiation, is a leading physical mechanism in climate. However, due to the uncertainties of this effect, caused by three dimensional (3D) structures of clouds, it is difficult to simulate the cloud system and its effects in climate models. For this reason, 3D radiative transfer models have been developed for the simulation of radiation absorption/scattering and considering the complex dependencies of the amount, shapes and microphysical properties of cloud field. During the last decade, technological improvements resulted in the use of all-sky imaging (ASI) systems for the detection/characterization of clouds as well as to have observational information of the 3D cloud structure.  Moreover, deep learning neural networks, the new and very promising generation of neural models that can achieve state-of-the-art accuracy, are proved, very recently, to be able to accurately predict the distribution of radiance in artificial clouds.

Under this framework, the D3D project aims to conduct innovative research in the study of interactions between clouds and solar radiation by combining a) 3D reconstructed cloud from ASIs, b) full 3D radiative transfer simulations and c) deep learning neural networks. ASIs will provide the necessary observations of the 3D cloud structure to quantify the 3D radiation effects and the linkages to aerosol and cloud optical properties. A 3D radiative transfer model will be used for simulations and the reconstruction of the radiance field for a variety of atmospheric conditions, including different types of aerosols and clouds. The 3D cloud information from the ASIs for selected atmospheric conditions will be the principal input parameter in the 3D model. The derived polar radiances will be validated against the estimated ones by a radiometer installed proximal to the ASIs. The deep-learning neural networks will be trained, validated, and tested against model-derived datasets for various artificial atmospheric conditions. At the final stage, the model outputs and measured 3D radiances for cloudy conditions will be used as model input to render radiances at fast rates.