A random forest spectral classification of the Gaia 500 pc white dwarf population

¹ Departament de Física, Universitat Politècnica de Catalunya, c/Esteve Terrades 5, 08860 Castelldefels, Spain
² Institut d’Estudis Espacials de Catalunya, Esteve Terradas, 1, Edifici RDIT, Campus PMT-UPC, 08860 Castelldefels, Spain

^★ Corresponding author: santiago.torres@upc.edu

Received: 7 March 2025
Accepted: 5 May 2025

Abstract

Context. The third Gaia Data Release (Gaia DR3) has provided the astronomical community with astrometric data on more than 1.8 billion sources, along with low-resolution spectra for 220 million of them. Such a large amount of data is difficult to handle by means of visual inspection. In recent years, artificial intelligence and machine learning algorithms have started to be applied in astronomy for data analysis and automatic classification, with excellent results.

Aims. In this work, we present a spectral analysis of the Gaia white dwarf population up to 500 pc from the Sun based on artificial intelligence algorithms to classify the sample into their main spectral types and subtypes.

Methods. In order to classify the sample, which consists of 78 920 white dwarfs with available Gaia spectra, we have applied a random forest (RF) algorithm to the Gaia spectral coefficients. We used the Montreal White Dwarf Database of previously labeled objects as our training sample. We compared this classified sample with other already published catalogs and with our own higher resolution Gran Telescopio Canarias (GTC) spectra. This allowed us to construct a golden sample of well-classified objects.

Results. The RF spectral classification of the 500 pc white dwarf population achieved an excellent global accuracy of 0.91 and an F1-score of 0.88 for the DA classification (i.e., white dwarfs that show Balmer spectral lines) versus the non-DA classification. In addition, we obtained a very high accuracy of 0.76 and a global F1-score of 0.62 for the non-DA subtype classification. In particular, our classification shows an excellent recall for DAs, as well as DBs and DCs (>90%), along with a very good precision (≥80%) for DQs, DZs, and DOs. Unfortunately, our algorithm does not perform as well with respect to correctly classifying subtypes due to the low resolution of the Gaia spectra.

Conclusions. The use of machine learning techniques, in particular, the RF algorithm, has enabled us to spectrally classify 78 920 white dwarfs – an increase of 543.6% over those previously labeled – with reasonable accuracy. Having an estimate of the spectral type for the vast majority of white dwarfs up to 500 pc provides the possibility of making better estimates of cooling ages, star formation rates, and stellar evolution processes, among other fundamental aspects necessary for studying the white dwarf population.