Representation of Typical Droplet Shapes in 2D Latent Space

Abstract

In this study, we investigate the clustering of 5000 droplets, each originating from one of five distinct droplet classes, each representing a unique geometry. The shape coordinates of the droplets are mapped to a 2D latent space through a two-step, fully reversible process involving Fourier series and autoencoders. Thus, each droplet contour sampled by 400 points is represented by 2 scalar numbers. We present data mapping 5000 droplets in two-dimensional space, which reveals ringlike clusters. Each cluster corresponds to a distinct droplet class, with the classification based on the droplet morphology, which is known a priori. Further, we find that well-known clustering algorithms like DBSCAN show limited success in identifying these ring-shaped clusters underscoring the need to explore alternative clustering approaches, particularly those tailored to handle ring-shaped distributed data.