Defense
Student: Eduardo Galvani Massino
Program: Astronomy
Title: "Estimation of emission lines in galaxy spectra using symbolic regression"
Advisor: Prof. Dr. Laerte Sodré Júnior
Judging Comitee:
- Prof. Dr. Laerte Sodré Júnior (President and Advisor) - IAG
- Profa. Dra. Paula Rodrigues Teixeira Coelho - IAG
- Dr. Clécio Roque de Bom - CBPF
- Prof. Dr. Fabrício Olivetti de França - UFABC
Other Members:
- Prof. Dr. Eduardo Serra Cypriano - IAG
- Prof. Dr. Roberto Cid Fernandes Junior - UFSC
- Profa. Dra. Emille Eugenia de Oliveira Ishida - Université Clermont-Auvergne
- Profa. Dra. Ângela Cristina Krabbe - IAG
Abstract:
Semi-analytic models of galaxy formation, such as L-GALAXIES, can predict the spectral continuum of galaxies but do not incorporate methods to estimate the equivalent widths (EWs) of nebular emission lines. This dissertation proposes a machine-learning approach to construct interpretable analytical models that explicitly express the dependence of EWs on stellar population synthesis parameters, flux- and mass-weighted ages and metallicities, stellar mass, and dust extinction, derived by the STARLIGHT code from SDSS DR8 spectra combined with GALEX GR6 ultraviolet photometry. The target lines are [NII]6584, [OIII]5007, Ha and Hb. Three algorithms were compared, Random Forest, Operon, and PySR, with symbolic regression, particularly Operon, achieving performance comparable to Random Forest while yielding explicit and readable equations, that is, revealing analytical relationships among the variables. A statistical framework based on multivariate normal sampling was developed, in which Operon models the means, standard deviations, and covariances of the EWs, capturing physical correlations among emission lines. The [OIII]5007 line proved the most challenging to model across all evaluated scenarios. The resulting models represent a concrete step toward incorporating emission lines into semi-analytic frameworks based on stellar population parameters.
Keywords: galaxies, spectroscopy, optical emission lines, equivalent width, machine learning, symbolic regression