Abstract Details

Name: Nitesh Kumar
Affiliation: Department of Phtsics and Astrophysics, University of Delhi
Conference ID: ASI2021_66
Title : Spectral interpolation using Artificial Neural Network(ANN)
Authors and Co-Authors : Nitesh Kumar (Department of Physics and Astrophysics, University of Delhi, Delhi-110007, India), Philippe Prugneil (Centre de Recherche Astrophysique de Lyon (CRAL, CNRS, UMR 5574), Université Lyon, Lyon, France), Prof. H. P. Singh(Department of Physics and Astrophysics, University of Delhi, Delhi-110007, India)
Abstract Type : Poster
Abstract Category : Instrumentation and Techniques
Abstract : To determine the atmospheric parameters of the star, the observed spectrum of the star is compared to a spectrum interpolated or approximated, over a grid of theoretical model spectra. This interpolation is a critical aspect of the process. We are interpolating the spectra using the state of the art machine learning method, a multi-layer feed-forward neural network, commonly referred as artificial neural network or ANN, and we are focusing our effort at characterizing the biases due to the interpolation. The goal is that the interpolation remains a minor contribution to the total error budget. Our goal is to define mathematical criteria for an interpolator that must be fulfilled in terms of precision. We implement these defined criteria that prevents overfitting while meeting the desired precision using artificial neural networks which takes the known atmospheric parameters( or some function of it) and the corresponding spectra as input. We investigated several different preprocessing schemes, which make the task of interpolation/approximation computationally efficient, such as normalising the spectra to bolometric flux(simplification), normalisation the spectra with a blackbody of corresponding effective temperature(variance reduction) before training the neural network. Once the network is trained, the output from ANN is multiplied by the effective temperature black-body radiation in the post-processing method named as variance restoration. In our work, we are using the Gottingen spectral library as the input grid to train and tune the ANN and the trained ANN can be used as the spectral interpolator and can work with full-spectrum fitting software like ULySS to determine the atmospheric parameters from the observed spectrum of the stars.