Three-Dimensional Radiative Transfer Emulation Using Machine Learning Techniques

Clouds play an important role in determining the energy flows in the Earth's system. An accurate representation of cloud-radiation interactions is thus critical for better prediction of weather and climate, quantification of cloud radiative forcing and feedback, and understanding of the role of radiation in cloud evolutions. Even though clouds are finite and clearly a three-dimensional (3D) object, one-dimensional (1D) radiative transfer that ignores cloud inhomogeneity and horizontal photon transport has been widely used due to computational considerations. While errors due to 1D simplification may be less severe compared to other parameterized processes in some cases, the use of 1D radiative transfer has proven insufficient for interpreting remote sensing observation and for simulating cloud organization, highlighting the need for a fast and accurate 3D radiative transfer scheme. Motivated by the work of Prof. Liou on 3D radiative transfer, we will demonstrate how to capitalize on machine-learning techniques to predict shortwave radiation and heating rates at 100 m horizontal resolution in a 3D environment. We will focus on radiation emulations for shallow cumulus regimes where clouds are highly heterogeneous and 3D effects are significant. While some might think embracing a black box like machine learning means abandoning rigorous physical principles of radiative transfer, we will instead show that any emulator should be configured based on our knowledge of radiative transfer; in other words, we should train the emulator, not vice versa.