Snow Particle Characterization. Part B: Morphology Dependent Study of Snow Crystal 3D Properties Using a Convolutional Neural Network (CNN)

Abstract

<div class="section abstract"><div class="htmlview paragraph">This study presents the results of the ICE GENESIS 2021 Swiss Jura Flight Campaign in a way that is readily usable for ice accretion modelling and aims at improving the description of snow particles for model inputs. 2D images from two OAP probes, namely 2D-S and PIP, have been used to extract 3D shape parameters in the oblate spheroid assumption, as there are the diameter of the sphere of equivalent volume as ellipsoid, sphericity, orthogonal sphericity, and an estimation of bulk density of individual ice crystals through a mass-geometry parametrization. Innovative shape recognition algorithm, based on Convolutional Neural Network, has been used to identify ice crystal shapes based on these images and produce shape-specific mass particle size distributions to describe cloud ice content quantitatively in details. 3D shape descriptors and bulk density have been extracted for all the data collected in cloud environments described in the regulation as icing conditions. They are presented under the form of composite size distributions and gathered in size classes, representative of fixed portions of the total mass encountered during the field campaign. The examination of the data shows high discrepancies between crystals of identical size. To solve this issue shape parameters are combined with the morphological analysis to provide comprehensive explanations for the observed snow descriptor variabilities. Finally, the results are summarized under the form of simple habit-specific parametrizations for 3D shape descriptors and bulk density, as functions of crystal size.</div></div>