Application of Convolutional Neural Networks to Predict Magnetic Fields’ Directions in Turbulent Clouds

We adopt the deep learning method casi-3d (convolutional approach to structure identification-3D) to infer the orientation of magnetic fields in sub-/trans-Alfv\ue9nic turbulent clouds from molecular line emission. We carry out magnetohydrodynamic simulations with different magnetic field strengths and use these to generate synthetic observations. We apply the 3D radiation transfer code radmc-3d to model 12CO and 13CO (J = 1−0) line emission from the simulated clouds and then train a casi-3d model on these line emission data cubes to predict magnetic field morphology at the pixel level. The trained casi-3d model is able to infer magnetic field directions with a low error (≲10\ub0 for sub-Alfv\ue9nic samples and ≲30\ub0 for trans-Alfv\ue9nic samples). We further test the performance of casi-3d on a real sub-/trans- Alfv\ue9nic region in Taurus. The casi-3d prediction is consistent with the magnetic field direction inferred from Planck dust polarization measurements. We use our developed methods to produce a new magnetic field map of Taurus that has a three times higher angular resolution than the Planck map