THREE-DIMENSIONAL RADIATIVE TRANSFER WITH MACHINE LEARNING: EMULATION AND INSIGHTS FROM AEROSOL OBSERVATIONS
August 11, 2025
Kevin Yang
Committee: Christine Chiu (Advisor); Christian Kummerow; Steven Miller; Imme Ebert-Uphoff (Electrical and Computer Engineering, CIRA)
Abstract
Radiation plays a central role in the Earth system, governing the distribution of energy and driving key processes in weather and climate. Atmospheric radiative transfer (RT), which describes the physical processes governing the propagation and interaction of radiation in the atmosphere, can be modeled with high accuracy using sophisticated mathematical formulations under well-defined assumptions. However, such modeling is computationally intensive due to the multidimensional nature of the problem (e.g., spatial and angular dependencies) and the complexity of the underlying physics (e.g., multiple scattering, spectral absorption, and emission). Consequently, RT calculations remain a major computational bottleneck in atmospheric modeling, limiting the use of more advanced and physically realistic radiative schemes. Machine learning, which can efficiently approximate complex, nonlinear relationships in high-dimensional spaces without explicitly solving the governing equations, offers a promising pathway to overcoming this limitation.In this dissertation, we leverage machine learning techniques to advance development with topics related to three-dimensional (3D) radiative transfer. In the first part, we demonstrate the use of machine learning to emulate shortwave radiation, predicting both surface fluxes and atmospheric heating rates at a horizontal resolution of 100 m in a fully three-dimensional environment. The focus is on shallow cumulus regimes, where cloud fields are highly heterogeneous and 3D radiative effects are significant. The emulator design is physically guided and informed by established principles of radiative transfer. The emulators' performance in terms of accuracy and efficiency will be discussed, as well as their potential applications. In the second part, we apply a machine learning–based aerosol retrieval method for passive satellite observations, developed by Yang et al. (2022), which accounts for 3D cloud radiative effects to enable accurate near-cloud aerosol retrievals, to investigate aerosol–cloud–radiation interactions. Specifically, we examine variations in near-cloud aerosol properties and their associated shortwave direct radiative effects (DRE) across four distinct cloud organizations—Sugar, Gravel, Fish, and Flowers—common in the trade-wind regimes. Differences in hydration-induced enhancement of the aerosol DRE among these organizations are quantified, and their interpretation will be discussed in the context of cloud field characteristics.