Evaluation and Selection of Autoencoders for Expressive Dimensionality Reduction of Spatial Ensembles

This paper evaluates how autoencoder variants with different architectures and parameter settings affect the quality of 2D projections for spatial ensembles, and proposes a guided selection approach based on partially labeled data. Extracting features with autoencoders prior to applying techniques like UMAP substantially enhances the projection results and better conveys spatial structures and spatio-temporal behavior. Our comprehensive study demonstrates substantial impact of different variants, and shows that it is highly data-dependent which ones yield the best possible projection results. We propose to guide the selection of an autoencoder configuration for a specific ensemble based on projection metrics. These metrics are based on labels, which are however prohibitively time-consuming to obtain for the full ensemble. Addressing this, we demonstrate that a small subset of labeled members suffices for choosing an autoencoder configuration. We discuss results featuring various types of autoencoders applied to two fundamentally different ensembles featuring thousands of members: channel structures in soil from Markov chain Monte Carlo and time-dependent experimental data on droplet-film interaction.