Retrieving Ice Particle Properties from in Situ Imaging Probes Using a Machine Learning Algorithm

Microphysical model simulations and multi-instrument remote sensing retrieval algorithms both rely on a priori assumptions about ice particle properties. Specifically, relations between the mass, maximum dimension, effective density, and aspect ratio are used to characterize ice particle models within microphysical growth equations and forward operators that simulate radiative properties. However, these relations are highly uncertain owing to the difficulty in measuring ice particles within precipitation systems. In situ ice particle probes provide some information about particle shapes and sizes from several view directions. However, these view directions are not necessarily optimally chosen and introduce some biases into the retrieved ice particle properties. Additionally, retrievals of mass and effective density are under constrained since current imaging probes cannot interrogate the three-dimensional particle structure. To address these biases, we examine the accuracy of various configurations of probe view angles in retrieving known physical properties from a synthetic database of ice particle shapes. The projections of the synthetic particles are simulated for each probe view, and these simulated probe observations are used within a convolutional neural network to predict the ice particle properties. We find that this machine-learning algorithm produces estimates of mass, effective density, and maximum dimension with reduced errors and biases relative to conventional retrieval techniques. In addition, we will discuss the ability for these algorithms to determine the optimal placement of probe sensors for the purposes of further reducing errors in the retrieved article properties.