Machine Learning Method Developed to Infer Physical Properties of Colliding-Wind Binary Stars from Image Data

A new computational method has been developed to solve a long-standing challenge in astrophysics: inferring the physical properties of colliding-wind binaries (CWBs) from observational data. Colliding-wind binaries are systems of two massive stars whose supersonic stellar winds collide and create bow shocks, producing detectable emission across multiple wavelengths including H-alpha, X-ray, and radio. The inverse problem of extracting parameters like mass-loss rates, terminal wind velocities, and orbital elements from short time-series observations has been difficult because the forward hydrodynamic simulations are computationally expensive and the likelihood function is intractable. The researchers developed a factorized spatio-temporal neural network architecture that separates spatial encoding from temporal aggregation, aligning with the underlying physics of local morphology and global dynamical evolution. Combined with a neural spline flow conditioned on these embeddings, the complete pipeline successfully infers seven physical parameters from synthetic H-alpha photon-count observations under realistic detector noise. The method demonstrates well-calibrated posteriors and robustly recovers orbital parameters and mass-loss rates, with the approach naturally expanding uncertainty estimates in information-poor regimes.