Classifying ALMA continuum data using machine learning

The volume of data from telescopes stored in archives has increased rapidly over time. The ALMA archive, for instance, holds such an immense amount of information that astronomers cannot effectively access it without search tools specifically designed for this archive. These tools function by assigning each observation with keywords, allowing users to search based on these keywords. However, this system has its shortcomings. Firstly, the astronomer assigning an observation might not recognize or include all relevant keywords. Secondly, given the vast and varied phenomena in our universe, creating a comprehensive keyword list becomes a near-impossible task.

Machine learning offers enormous potential for both finding objects of interest and then interpret their morphology in terms of physically meaningful properties. In this project, we trained a convolutional  neural network to detect gaussian disks with extensions in the ALMA-archive. Extensions in the dust emission perpendicular to disks could be a sign of dust leaving the disk via winds, but to date only a handful of such observations have been acquired. We want to know whether this is a common feature for young disks or not. Our data set originates from five unique observations and are used to train a neural network.

By using linear and non-linear augmentation techniques (artificially creating data), we expanded the data set and trained a neural network. Preliminary results show that this methodology can be used to find observations in the archive, that resembles the input training data. In other words, Gaussian disks with extensions can be identified. A future, and long-term, goal is to refine the approach in such a way that the code can handle data and perform morphological classification for various archives.