Joint MK classification of stellar spectra with distance-aware CNNs

Large surveys now produce millions of stellar spectra, demanding fully automatic and physically consistent classification. We train a three-branch 1-D CNN with a distance-aware loss that predicts a star’s Morgan-Keenan (MK) temperature and luminosity label in one pass, penalising larger grid errors more strongly. Trained on spectra gathered from five public libraries and re-labelled via SIMBAD, an ensemble of this network with a ResNet50 baseline reaches macro F1 = 0.676 and Quadratic-Weighted Kappa for T, L = (0.983, 0.884) on the test set, with gains verified by paired bootstrap tests. Guided Grad-CAM highlights features tied to specific spectral types that overlap molecular bands and possible mislabelling can be flagged with careful cross comparisons. All code, trained weights, curated dataset and further results are available at our GitHub repository.