Stellar Insights at Scale with Likelihood-Free Inference

Eclipsing binary stars (EBs) offer excellent laboratories to test and constrain stellar evolution, study accretion physics, and calibrate cosmic distances. Traditionally, detailed models of EB light curves along with phase-resolved radial velocity (RV) measurements are used to determine the full orbit solution as well as the stellar parameters of the constituents. However, although we have light curves for hundreds of thousands of EBs, we typically lack corresponding RV measurements, which are heavily resource consuming in that they require multiple moderate- to high-resolution spectra. As a result of this issue, coupled with the computational expense of modeling EBs, the vast majority of EB systems have not been studied in detail. With the coming of Rubin Observatory and other powerful instruments that will collect light curve and color information on millions of EBs, an urgency arises in our ability to study these systems at scale with little to no RV information. Here we present preliminary results of applying novel likelihood-free inference (LFI) to infer the fundamental parameters of ASAS and ZTF EBs. LFI, originally developed outside of astronomy for parameter inference, employs a neural density estimator (NDE) to determine the likelihood function based on a finite number of light curve simulations drawn from realistic priors. Therefore, after training the NDE on these light curves, which we simulate using the EB modeling software PHOEBE, no further use of the expensive forward model is required; that is, the training computation is amortized and reused for rapid and accurate inference, which in turn permits the application of a more robust forward model.