Machine Learning Approaches for Detection and Classification of Astrochemical Spectral Lines

摘要

Astronomical spectroscopy is a field that has been growing for a number of years, analyzing the features of molecular spectral lines from astronomical data cubes provides insight to the composition and dynamics of our universe. With the arrival of powerful telescopes like ALMA, the size of the data cubes will be constantly growing. This is why we believe that some automatic analysis methods will be helpful assisting the astrochemists work. We experimented with a method to analyze astronomical data cubes, detect their regions of interest, by using a non supervised clustering algorithm, and then, create a spectrum for each region of interest, and classify the molecular species found in the spectra, by using a supervised training algorithm. The training is done using synthetic spectra, and the validation is done using radio astronomical data cubes from ALMA data. Initial experiments contemplated a basic physical model and two popular Machine Learning methods were tested for the task of classifying molecular spectra, Support Vector Machines and Artificial Neural Networks; experimental results provide class probabilities ranging from 76.9% to 94.09% for Artificial Neural Networks, however, Support Vector Machines only rendered class probabilities slightly better than random, with 75% accuracy in the classification. A new approach using Mixed Membership Models, a technique from the datamining world is being tested. A parallelization test was also performed, obtaining a speedup of 2x in the process of real world data files.