Using Machine Learning to Identify Brown Dwarfs in Deep JWST Surveys

The James Webb Space Telescope (JWST) has dramatically enhanced our ability to observe faint sources in deep-field surveys, revealing a distant and ancient population of brown dwarfs likely originating from the Milky Way's metal poor thick disk and halo populations. Identifying these rare objects requires analyzing increasingly large spectral datasets with high efficiency and accuracy. To address this, we trained and evaluated several machine learning (ML) classifiers, including k-nearest neighbors, logistic regression, random forests, and support vector machine on JWST/NIRSpec Prism spectroscopic data to isolate brown dwarfs from predominantly extragalactic sources. We explored performance between the different ML algorithms, variations in the balance between brown dwarfs and non-brown dwarfs of the training set, and the use of principal component analysis (PCA) as an initial data filter. The balanced random forest model achieved the best performance, with an accuracy of 98%, brown dwarf recall of 100%, and non-brown dwarf recall of 98%. When applied to NEXUS survey data, the model identified 95 brown dwarf candidates out of 3697 spectra, recovering all 37 sources previously confirmed by visual inspection. These results demonstrate the potential of machine learning to accelerate brown dwarf discovery in JWST deep-field surveys, though further refinement is needed to reduce contamination from non-brown dwarf sources.