Thesis Identificación de subclases de enanas blancas en datos espectroescópicos mediante espacios latentes y clustering
Loading...
Date
2026-09
Authors
Journal Title
Journal ISSN
Volume Title
Program
Ingeniería Civil Electrónica
Departament
Campus
Campus Casa Central Valparaíso
Abstract
Las enanas blancas son remanentes estelares compactos cuya clasificación espectral actual (clases como WDA, WDB, entre otras) ha permitido grandes avances en su estudio. Sin embargo, existen indicios de que dentro de estas clases principales podrían existir subestructuras no reconocidas, originadas por diferencias en la composición química atmosférica, fenómenos de microfísica del plasma, magnetismo o acreción, que modulan la forma e intensidad de las líneas espectrales sin cambiar la etiqueta química gruesa. El presente trabajo propone y construye un marco metodológico basado en aprendizaje no supervisado para descubrir y validar subclases de enanas blancas a partir de sus espectros. La metodología utiliza una red neuronal profunda preentrenada como extractor de características: se extraen representaciones intermedias (espacios latentes) de cuatro capas de la red, se reducen mediante Análisis de Componentes Principales (PCA) y se aplican sobre ellas algoritmos de clustering de distinta naturaleza (K-Means, Spectral Clustering y DBSCAN). La calidad de los agrupamientos se evalúa mediante métricas internas (Silhouette, Davies–Bouldin, Calinski–Harabasz y DBCV), su estructura se examina mediante matrices de migración entre configuraciones con distinto número de clusters y, para cada grupo, se generan espectros representativos que se contrastan con la fenomenología astrofísica de las enanas blancas. Los experimentos se realizan sobre espectros del catálogo Sloan Digital Sky Survey (SDSS) para las clases WDA y WDB. Los resultados muestran que la capa de la que se extrae el embedding no determina únicamente la calidad de los agrupamientos, sino el tipo de estructura que puede recuperarse: (...).
White dwarfs are compact stellar remnants whose current spectral classification (classes such as WDA, WDB, among others) has enabled major advances in their study. Nevertheless, there is evidence that unrecognized substructures may exist within these main classes, arising from differences in atmospheric chemical composition, plasma microphysics, magnetism or accretion, which modulate the shape and intensity of spectral lines without changing the coarse chemical label. This work proposes and builds a methodological framework based on unsupervised learning to discover and validate white dwarf subclasses from their spectra. The methodology uses a pretrained deep neural network as a feature extractor: intermediate representations (latent spaces) are extracted from four network layers, reduced via Principal Component Analysis (PCA), and clustered using algorithms of different nature (K-Means, Spectral Clustering and DBSCAN). Cluster quality is assessed through internal metrics (Silhouette, Davies–Bouldin, Calinski–Harabasz and DBCV), their structure is examined by means of migration matrices between configurations with different numbers of clusters, and representative spectra are synthesized for each group and compared against the astrophysical phenomenology of white dwarfs. Experiments are carried out on spectra from the Sloan Digital Sky Survey (SDSS) for the WDA and WDB classes. Results show that the layer from which the embedding is extracted does not merely affect cluster quality, but determines the kind of structure that can be recovered: (...).
White dwarfs are compact stellar remnants whose current spectral classification (classes such as WDA, WDB, among others) has enabled major advances in their study. Nevertheless, there is evidence that unrecognized substructures may exist within these main classes, arising from differences in atmospheric chemical composition, plasma microphysics, magnetism or accretion, which modulate the shape and intensity of spectral lines without changing the coarse chemical label. This work proposes and builds a methodological framework based on unsupervised learning to discover and validate white dwarf subclasses from their spectra. The methodology uses a pretrained deep neural network as a feature extractor: intermediate representations (latent spaces) are extracted from four network layers, reduced via Principal Component Analysis (PCA), and clustered using algorithms of different nature (K-Means, Spectral Clustering and DBSCAN). Cluster quality is assessed through internal metrics (Silhouette, Davies–Bouldin, Calinski–Harabasz and DBCV), their structure is examined by means of migration matrices between configurations with different numbers of clusters, and representative spectra are synthesized for each group and compared against the astrophysical phenomenology of white dwarfs. Experiments are carried out on spectra from the Sloan Digital Sky Survey (SDSS) for the WDA and WDB classes. Results show that the layer from which the embedding is extracted does not merely affect cluster quality, but determines the kind of structure that can be recovered: (...).
Description
Keywords
Enanas blancas, Clasificación espectral, Aprendizaje no supervisado, Espacios latentes, Espectros astronómicos, Agrupamiento (Clustering)
