Informing Depositional Ice Growth Models Through 3-D Reconstruction of Ice Crystal Images Using Machine Learning

Cirrus clouds play a key role in earth’s energy balance and hydrological cycle, however they contribute to large uncertainties in climate and weather models. These uncertainties are driven by an incomplete physical understanding of microphysical processes, as well as the scale disparities between climate/weather model resolutions and size scales of individual cloud particles. For cirrus clouds, crystal habit (i.e., shape) impacts the physical evolution of particles, which is important for understanding the macroscopic evolution of clouds. In addition, ice crystal habit influences cloud-radiation interactions. Current models of depositional ice growth have limited representations of crystal habit, if at all. In our work, we use single-particle images from cloud particle imagers (CPI) to understand the real-world distribution and diversity of 3-D crystal geometries, and subsequently use this to inform models of depositional growth. Since single-view images do not explicitly capture the 3-D features of crystals (e.g., mass-density relationships) that may be necessary to inform microphysical parameterizations (e.g., depositional growth), we adopt methods from computer vision and 3-D modeling to develop a supervised learning pipeline that infers 3-D features from CPI images, using “synthetically” generated particles as training data. As proof of concept, we computationally generated a prototype dataset of 10,000 geometrically-diverse bullet rosette ice crystals and then trained models using this dataset to infer 3-D features (e.g., effective density) from single-view images. We chose bullet rosettes for our prototype dataset because they are known to be abundant in cirrus clouds, and furthermore, we can connect these geometric representations to recently developed bullet rosette growth models (e.g., Pokrifka et al. 2023). We show that even non-deep-learning regression models are able to infer 3-D features from 2-D images with relatively high accuracy. Furthermore, deep learning approaches allow for explicit 3-D reconstructions of ice crystal geometries from single-view images. We will present preliminary results of explicitly reconstructed 3-D ice crystals, and also discuss how improved habit representations can be incorporated into contemporary ice growth models.