Clustering the Latent Dimensions of a Variational Autoencoder to Identify Representative Trajectories of Ice Crystals in Cirrus Clouds Modeled By Lagrangian Microphysics

Within cirrus clouds, ice crystals have diverse trajectories as they change with time. Particle-based Lagrangian cloud models, which use a smaller set of supercrystals to represent a population of cloud particles with weighted multiplicities, can account for a wide range of environmental histories that affect these particles. Recently, this Lagrangian method has been incorporated with the large eddy simulation (LES) in CM1 to model a cirrus cloud case from the Ice Cryo-Encapsulation Balloon (ICEBall) Campaign. An intercomparison of different mechanisms for ice crystal depositional growth found that the individual thermodynamic histories of the ice crystals are the primary factor in their microphysical variability, rather than their current microphysical properties or environmental conditions [Chandrakar et al. 2024, GRL]. This particle-based modeling approach allows us to identify trends in the supersaturation history, which is a good predictor of ice microphysical evolution.

By passing the supercrystals supersaturation histories into a Variational Autoencoder, we compress the 35 dimensional time series into two dimensions. After applying K-means clustering and connecting the generated cluster labels to the original uncompressed supercrystal trajectories, we observe that generating the clusters within this latent space generates more disparate clusters. Additionally, after compressing the trajectories we can generate representative reconstructions from the deciles across the latent space. From observing the variances in these reconstructions, it is evident that the vast majority of the variances are contained within one of the latent dimensions, suggesting that a 1-d reduced-order representation of thermodynamic history is possible. An independent approach using the Uniform Manifold Approximation and Projection (UMAP) algorithm shows that the clusters lie on separate sections of an approximately 1-d manifold, strengthening the evidence for a single latent variable that can unify the diverse trajectories. This suggests that complex ice particle histories in cirrus clouds can be represented with much greater simplicity.