. Using machine learning to identify mixed-phase conditions in cloud radar Doppler spectra

Abstract
Mixed-phase conditions impact cloud radiative effects, precipitation processes, and cloud lifetimes. Despite their importance, mixed-phase cloud effects are difficult to correctly simulate due to a lack of understanding of the microphysical processes involved. To improve our understanding of mixed-processes we need additional observations to characterize when mixed-phase conditions exist and how they relate to atmospheric state. This work explores the potential of machine learning to identify cloud phase in vertically pointing active sensors at the US Department of Energy Atmospheric Radiation Measurement (ARM) sites. The ability to detect mixed-phase conditions in the stratiform regions of deep convection and in convective clouds with weak updrafts is investigated using k-means clustering of Ka band radar Doppler spectra moments. Doppler spectra data is harder to interpret in deep convective clouds due to high turbulence and precipitation which broaden and attenuate the signal respectively. Despite these challenges, a mixed-phase signature can be seen in the clouds. The phase identification from the Ka-Band measurements is compared to signatures in the Raman Lidar and hydrometeor identification in lower frequency precipitation radars to validate the results. An additional method of using a Bayesian classifier with lidar and radar data together is introduced as a means of estimating the probabilities of a given phase classification. While this supervised-learning method has the potential to merge multiple measurements with an estimate of the uncertainty in a given context, developing a robust operational algorithm requires having sufficient representative training data which can be challenging. Some preliminary results are shown from the M-PACE field campaign in Barrow, Alaska.