Volga Sezen, Joint Temperature-Luminosity Classification of Stellar Spectra with a Distance-Aware Loss Function

M.S. Candidate: Volga Sezen
Program: Data Informatics
Date: 28.08.2025 / 13:30
Place: 
B-223

Abstract: Modern sky surveys continue to produce large datasets, making automated classification essential for population studies and new discoveries. In this work, a stellar spectral database was assembled from five public libraries and re-labelled via SIMBAD. A 1-D CNN with three branches of differing kernel sizes was initiated to focus on different features while predicting a star’s MK temperature and luminosity class in one pass. A distance-aware loss was used, coupling cross-entropy with mean-squared error on the 2-D MK grid so physically further misclassifications incur stronger penalties. Alongside the three branch CNN, other architectures were trained and evaluated using several metrics on a held-out test set. Under circularly shifted inputs, performance remained robust up to 3-pixel shifts. An ensemble of the three branch CNN with a custom ResNet50 achieved macro F1 of 67.6% and kappa of 98.3 and 88.4 across two axis, with statistically significant gains over each counterpart under paired bootstrap testing. Guided GradCAM indicates a non-linear relation between branch kernel size and identified features, with some features overlapping known molecular bands. All code, trained weights, curated dataset, and additional results are available at GitHub (github.com/volgasezen/StellarClassification).