Presentation #220.01 in the session “Machine Learning in Astronomy: Measuring the Properties of Stars with Machine Learning (Meeting-in-a-Meeting)”.
Star formation is messy! The process spans many orders of magnitude in spatial scale and involves a variety of physical processes: gravity, magnetic fields, radiation and turbulence. Forming stars announce their presence by emitting radiation and ejecting high-velocity material. However, identifying this stellar feedback is challenging, so feedback signatures have traditionally been identified “by eye” — either by professional astronomers or by citizen scientists. We develop a new supervised convolutional neural network approach, the 3D Convolutional Approach to Structure Identification (CASI-3D), that provides a more reliable, quantitative and faster alternative to visual searches. We train CASI-3D on numerical simulations and apply it to molecular emission in order to to identify shells and outflows created by young stars. CASI-3D confirms most prior visual identifications and identifies instances of new, previously missed feedback. These results suggest the physical impact of stellar winds has been over-estimated by an order of magnitude while the impact of protostellar outflows has been underestimated. CASI is publicly available on Gitlab; it is a fully general method that is suitable for a wide range of physical problems that require structure identification in 2D and 3D data.