By combining spectra, light curves and images from VizieR and IRSA, the classifiers identify young stars with over 90 per cent accuracy and rule out outliers with over 99 per cent accuracy.
Large surveys usually identify young stellar objects by their excess of infrared radiation, based on colour combinations specific to each instrument. The result: catalogues built with different criteria that are difficult to compare. The NEMESIS project, on the other hand, is looking for a single method that can be applied to all available data.
The team gathered the spectral energy distributions, light curves and images of a large number of objects of all kinds via the VizieR database in Strasbourg and NASA's IRSA, and then converted them into visual representations. A set of binary classifiers — architectures from the PyTorch library and custom-designed convolutional networks — learn to distinguish young stars from everything else. It recognizes them with greater than 90% accuracy and identifies contaminants with more than 99%.
Applied to many existing catalogues, the tool results in 274,408 unique and reliable candidates. Their position in the color-magnitude diagrams and their distribution in relation to the local structures of the Galaxy confirm their youth.
Read the full article on the original source
Go to the source →