Classification-of-variable-stars

Master's Thesis

Abhina Premachandran Bindu

Mentor: Dr Greg Olmschenk, NASA Postdoc

City College of New York, City University of New York

This is a deep learning project where a convolutional neural network is used to classify and identify variable stars. The CNN model used is a 10 convolutional layered network that is trained on 5 different variable star classes. The latest trained model possess a validation accuracy of 85.07 % and have identified 10,644 new variable stars as well as non-variable stars from 209,660 light curves whose status of variability is unknown. The model is trained and infered and the parameters are adjusted to refine the performance. The data used are the FITS files from Transiting Exoplanet Survey Satellite, an allsky survey satellite deployed in 2018 to find transiting exoplanets by recording the lightcurves of millions of stars. Therefore, our CNN model is a vital tool in classifying and identifying variable stars from millions of TESS light curves. Thereby aiding the study of variable stars, which is a subfield in Astrophysics, crucial in understanding the struture and evolution of stars and our Universe.