Author(s) :

Jayveer Kochhar .

Volume/Issue :

Abstract :

Measurement of astronomical distances, understanding of stellar evolution, and improvement of our knowledge of galactic structure depend on accurate classification of variable stars. Conventional classification techniques find it difficult to handle the rising volume of data produced by massive sky surveys such as Large Synoptic Survey Telescope (LSST), All Sky Automated Survey for Supernovae (ASAS-SN), and Optical Gravitational Lensing Experiment (OGLE). This work offers a hybrid deep learning method using Multiple-Input Neural Networks (MINN) to improve classification accuracy by means of image-based light curve analysis combined with astrophysical parameters derived via Fourier decomposition and skewness analysis. Problems including class imbalance, phase misalignment, and subtype distinction are addressed in the suggested approach. Minima Phase Standardization aligns phase-folded light curves for uniformity; the Fourier Best Fit model extracts important coefficients reflecting light curve shape. A Variable Star Light Curve Simulator creates synthetic data for underrepresented classes, especially ACeps and Type II Cepheids, to reduce dataset imbalance, therefore guaranteeing a more balanced training dataset. In ten epochs, the hybrid model attained an overall classification accuracy of 89.8%; considerable gains for rare classes were obtained. For common classes, the convolutional neural network (CNN) alone achieved 98.1% accuracy. This work emphasizes, especially for rare and underrepresented variable star types, the need of integrating deep learning with astrophysical insights to increase classification accuracy. By laying the groundwork for automated, large-scale classification of variable stars, the proposed framework accelerates the study of huge astronomical datasets and improves our knowledge of the structure and development of the universe.