Taking the temperature of the ISM with deep learning
Recent, high-resolution surveys of 21cm emission have revealed that neutral hydrogen (HI) in the local interstellar medium (ISM) contains a wonderful wealth of structures which reflect the complex interplay of Galactic dynamics and star formation feedback. Measuring the basic physical properties of these structures is crucial for understanding their origins, and also correcting observations of extragalactic light. However, constraining their temperature and density requires observations of 21cm absorption, which are severely limited. In this talk I will present our recent efforts to measure the temperature of HI across the sky using deep learning. We train a convolutional neural network using synthetic spectra from numerical simulations to predict quantities which formally require 21cm absorption — the true HI column density and the fraction of cold, optically thick HI along the line of sight — from 21cm emission alone. We validate the model using 21cm absorption observations from the literature, finding excellent accuracy. With this model, we construct the highest-resolution, highest-fidelity map of cold HI in the local ISM using 21cm emission data from the GALFA-HI and the HI4Pi surveys. This map characterizes the structure of neutral gas envelopes to molecular clouds with unprecedented resolution, and significantly improves dusty Galactic foreground estimation for extragalactic surveys. Via comparison with tracers of dust reddening (E(B-V)), we demonstrate that E(B-V)/N(HI) increases with increasing cold gas fraction, which will be leveraged to produce high-resolution, high-fidelity E(B-V) map at high latitudes.
Zoom Recording: https://uky.zoom.us/rec/share/fNUjvt8ROxoPA3MEISKc0BQUETJXHEtHJeVQWbdLn3ZDSYHjFfy9Enyzxn8-rHiG.d-eL5anXE5xOT5Qa