Machine Learning for Protostellar Image Fitting. A Convolutional Neural Network Approach

Abstract

In this work we present a Convolutional Neural Network (CNN) architecture that can be utilized to regress two key physical features of massive protostars from images in the 19 µm and the 37 µm bands: the inclination angle with respect to the line of sight (θview) and the protostellar mass (m∗). The network was trained on images generated by Monte Carlo Radiative Transfer simulations following the description of massive star formation from the Turbulent Core Model [1] [2] [3]. We show by testing the network on the synthetic data that it is feasible to regress the values for the aforementioned features solely from the morphology of the protostar displayed in the images, with a mean offset of µδ(θ) = −0.43◦ and standard deviation of σδ(θ) = 4.45◦ for θview and µδ(m) = −0.69 M⊙ and σδ(m) = 2.31 M⊙ for m∗. The network is subsequently re-trained with added negative samples to also discern between inputs that contain protostars and inputs that contain random noise, albeit with a slight increase in the dispersion of the offsets. This new model is then tested on an image of Cepheus A from the SOFIA Massive Star Formation Survey and the estimations for θview and m∗ agree within the error range with other estimations in the literature obtained by different methods.