MUSHY: Multimodal Flow-Based Amortized Bayesian Inference for Spectroscopic Data Fusion
TL;DR
MUSHY is an amortized Bayesian inference method, using a rectified flow model with multimodal inputs to learn a posterior distribution over physical parameters, that can precisely infer physical parameters with well-calibrated posteriors from any combination of surveys.
Abstract
We are in an era of unprecedented spectroscopic data availability, with surveys such as APOGEE, GALAH, and DESI observing millions of stars. However, data are heterogeneous and to date, efforts to exploit synergies between surveys have been limited. In particular, there is a lack of principled methods that combine spectroscopic data from different surveys to improve inference of physical parameters. In this work we present a new method, MUSHY, for spectroscopic data fusion. MUSHY is an amortized Bayesian inference method, using a rectified flow model with multimodal inputs to learn a posterior distribution over physical parameters. Using synthetic spectra generated via ATLAS12/SYNTHE, including realistic instrumental effects and noise, we demonstrate that MUSHY can precisely infer physical parameters with well-calibrated posteriors from any combination of surveys, with multimodal inputs providing stronger constraints on the posterior than single-survey data alone. This work paves the way for principled combination of spectroscopic data from heterogeneous sources.