THE APPLICATION OF CONVOLUTIONAL NEURAL NETWORK INTO APOGEE SPECTRA TO DETERMINE STELLAR PARAMETERS T_EFF, LOG?(G), DAN [FE/H]

• Main Author: Al Farisi, Salman
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/49992
• Institution: Institut Teknologi Bandung

Advances in artificial intelligence, specifically in machine learning method and the availability of overflowing data in the era of big data astronomy generate plenty of opportunities to explore new knowledge from observational results. In the APOGEE spectroscopic surveys, the determination of stellar astrophysical parameters (T_eff, log?(g), dan [Fe/H]) is a task that in needs of fast automation. Machine learning shows its ability as fast automation for stellar spectra analysis by utilising built models for further prediction. StarNet as an CNN-based spectral analysis shows decent precision over all range of temperature and signal to noise ratio (SNR) in APOGEE DR13, with mean total propagated error of 26.430 K for $T_{eff}$, 0.046 dex for $log(g)$, and 0.017 dex for $[Fe/H]$. Modifications of neural network architecture in StarNet was performed as a hope to better data understanding. Three modified architectures are proposed: 1 dimensional VGG16, StarNet+LSTM, and StarNet+GRU. All of three proposed architecture show smaller mean total propagated error in error propagation test and make it possible to perform an ensemble learning. Beside of those, the experiment of two synthetic spectra, ASSET and APOGEE synthetic spectra was trained to build a pretrained model as a hope for transfer learning in the further observation.