Photometric Classification of Single and Binary Stars with Mock CSST Observations

Abstract

Identifying unresolved binary stars from broadband photometry is important for studies of stellar evolution, binary populations, and Galactic structure, but remains challenging because many single and binary systems occupy overlapping regions of photometric feature space. We use mock seven-band Chinese Space Station Telescope (CSST) observations generated from the TRILEGAL Galactic population model, together with a physically motivated noise prescription, to assess an artificial neural network framework for classifying main-sequence (MS) stars, red giant branch (RGB) stars, white dwarfs (WDs), and representative unresolved binaries. We consider two complementary schemes: a direct seven-class classifier for MS, MS+MS, MS+WD, RGB, RGB+WD, WD, and WD+WD systems, and a hierarchical two-stage framework that first identifies the dominant stellar type and then separates single stars from binaries. Evaluated on nine independent sky fields, the direct classifier achieves an overall accuracy of 69.61%, while the final 16-class output of the hierarchical model reaches 58.73%. WD-containing binaries are recovered with high completeness, with recalls of 90.67% for MS+WD and 98.75% for RGB+WD systems, mainly because hot WD companions produce detectable short-wavelength excesses. However, the limited precision of some rare classes means that these predictions are better interpreted as high-completeness candidate samples than as high-purity catalogs. Strong degeneracies remain between single MS stars and MS+MS binaries, and between single WDs and WD+WD systems. These results quantify both the diagnostic potential and the limitations of CSST broadband photometry for stellar and unresolved binary classification.