Contributed Talk - Splinter Learning
Wednesday, 23 September 2020, 09:25 (virtual room G)
Emission-line diagnostics of HII regions using conditional Invertible Neural Network
Da Eun Kang, Eric Pellegrini, Ralf Klessen, Lynton Ardizzone
Our understanding of stellar feedback and birthplaces of young massive stars is largely based on the observation of gas and molecular lines. However, the entire forward process from nature to observed emissions is so complicated, degenerated, and non-linear that it is not easy to understand this overall process by ordinary forward modeling methods. To overcome the degeneracy between observation and nature, we present a new machine learning tool that can characterize and analyze HII regions. We apply conditional Invertible Neural Network (cINN), one of deep learning architectures, to link the physical properties of HII regions and optical emission-line luminosities. To train the network, we used a mock database of 505,748 HII regions generated by WARPFIELD-EMP. From the given emission-line luminosity information, our network provides the posterior probability distribution of HII region models composed of essential properties such as mass, star formation efficiency, and age of the star-forming cloud. Using some test observations, we show how accurately and precisely our network predicts by re-confirming the emission-line luminosity of predicted models through WARPFIELD-EMP.