Updates to LUCI: A New Fitting Paradigm Using Mixture Density Networks

TL;DR

This paper enhances the LUCI spectral line-fitting pipeline by integrating mixture density networks to output Gaussian posteriors, enabling Bayesian inference and improving fit initialization accuracy.

Contribution

It introduces a novel approach where the CNN outputs Gaussian posteriors for fit parameters, facilitating Bayesian inference in spectral line fitting.

Findings

Reduced computation time for spectral line fitting

Lowered risk of local minima in optimization

Improved accuracy of fit parameter estimates

Abstract

LUCI is an general-purpose spectral line-fitting pipeline which natively integrates machine learning algorithms to initialize fit functions. LUCI currently uses point-estimates obtained from a convolutional neural network (CNN) to inform optimization algorithms; this methodology has shown great promise by reducing computation time and reducing the chance of falling into a local minimum using convex optimization methods. In this update to LUCI, we expand upon the CNN developed in Rhea et al. 2020 so that it outputs Gaussian posterior distributions of the fit parameters of interest (the velocity and broadening) rather than simple point-estimates. Moreover, these posteriors are then used to inform the priors in a Bayesian inference scheme, either emcee or dynesty. The code is publicly available at https://github.com/crhea93/LUCI.