Applying Denoising Diffusion Probabilistic Models and Convolutional Neural Networks to Study Magnetic Fields in Molecular Clouds

Machine learning methods such as Denoising Diffusion Probabilistic Models (DDPMs) and Convolutional Neural Networks (CNNs) are powerful tools that allow us to reveal complex structures and generate predictions on large datasets. We have built such models to allow us to study the relationship between polarization and magnetic fields in molecular clouds. An overview of these models is presented in the context of their application to synthetic dust observations and real dust polarization observations. We have trained a CNN to classify polarization data based on alignment mechanism with over 98% validation accuracy. We have also trained an auto-encoder CNN to make pixel-by-pixel predictions of magnetic field strength and angle. Having a similar goal to the auto-encoder CNN, we have built DDPMs to generate images of the magnetic field strength and angle. Finally, we have applied the models to real dust emission observations in order to make predictions about the magnetic fields in actual molecular clouds. Results from the CNNs and DDPMs are presented and a comparison of the models is made.