Mining complex hyperspectral ALMA cubes for structure with neural machine learning

Abstract

Astronomy is producing the largest “Big Data” sets today and in the near future, with instruments such as the Atacama Large Millimeter and sub-millimeter Array (ALMA), the Large Synoptic Survey Telescope (LSST), and the Square Kilometer Array (SKA). These observations afford a deeper, wider, and more dynamic glimpse into the structure and composition of the universe than ever before. However, in addition to unprecedented volume, the data also exhibit unprecedented complexity, mandating new approaches for extracting and summarizing relevant information. ALMA data, in particular, challenges with very high dimensionality (measurements in a large number of spectral channels) where the dimensions represent both compositional information and velocities, and the high spectral resolution allows detailed interpretation of the kinematic structure of sources such as molecular clouds or protoplanetary disks. Traditional tools like moment maps can no longer fully exploit and visualize the rich information in these data. We present a neural map-based clustering approach that can utilize all spectral channels simultaneously and is capable of finding clusters of widely varying statistical properties, which are expected in these complex data sets. Many clustering methods, including modern graph segmentation algorithms, run into limitations when encountering such data. We demonstrate our tools, collectively named “NeuroScope”, through structure mining from an ALMA image of the protoplanetary disk HD142527. We highlight the advantages for both the emerging details and visualization. In addition, we explore an augmentation of leading graph segmentation algorithms with NeuroScope products, which can lead to efficient full automation of our clustering process for fast distillation of large data sets on-board or in archives.