Contact binaries: a neural network generates radial velocity curves in a flash, whereas physical models used to calculate them laboriously

Faced with the influx of data from large-scale sky surveys, a fully connected network and a SIREN model speed up the simultaneous fitting of light curves and radial velocity curves.

Large-scale time series surveys of the sky have enabled the identification of a considerable number of contact binaries – pairs of stars so close together that they share a common envelope – using both light curves and radial velocity measurements. However, modelling them relies on complex physical calculations, which prevent them from being studied on a large scale.

An emulator integrated into the fitting pipeline

A study published on arXiv proposes training a fully connected neural network and a SIREN model (a network with sinusoidal activation functions) to rapidly and accurately generate radial velocity curves. Once trained, these models significantly reduce the computational cost. The authors have integrated them into the light curve calculation framework developed by Li et al. (2025), which allows both types of observations to be fitted simultaneously.

To verify the reliability of the tool, the team constructed sets of synthetic observations, estimated their orbital parameters and quantitatively measured the deviation from known actual values. The aim is to enable large-scale statistical studies of these systems.