Random Forest classification of Gaia DR3 white dwarf-main sequence spectra: a feasibility study

TL;DR

This study assesses the feasibility of classifying Gaia DR3 spectra of white dwarf-main sequence binaries using a Random Forest algorithm trained on synthetic data, achieving high accuracy especially at higher resolving powers.

Contribution

It extends a known Random Forest classification method to Gaia WDMS spectra, demonstrating its effectiveness and robustness in a new, large-scale astronomical dataset.

Findings

Classification accuracy depends on SNR and resolving power.

Algorithm achieves nearly 80% accuracy on synthetic Gaia spectra.

High accuracy (60%) even for spectra dominated by MS stars.

Abstract

The third Gaia data release provides low-resolution spectra for around 200 million sources. It is expected that a sizeable fraction of them contain a white dwarf (WD), either isolated, or in a binary system with a main-sequence (MS) companion, i.e. a white dwarf-main sequence (WDMS) binary. Taking advantage of a consolidated Random Forest algorithm used in the classification of WDs, we extend it to study the feasibility of classifying Gaia WDMS binary spectra. The Random Forest algorithm is first trained with a set of synthetic spectra generated by combining individual WD and MS spectra for the full range of effective temperatures and surface gravities. Moreover, with the aid of a detailed population synthesis code, we simulate the Gaia spectra for the above mentioned populations. For evaluating the performance of the models, a set of metrics are applied to our classifications. Our results show that for resolving powers above ~300 the accuracy of the classification depends exclusively on the SNR of the spectra, while below that value the SNR should be increased as the resolving power is reduced to maintain a certain accuracy. The algorithm is then applied to the already classified SDSS WDMS catalogue, revealing that the automated classification exhibits an accuracy comparable (or even higher) to previous classification methods. Finally, we simulate the {\it Gaia} spectra, showing that our algorithm is able to correctly classify nearly 80% the synthetic WDMS spectra. Our algorithm represents a useful tool in the analysis and classification of real Gaia WDMS spectra. Even for those spectra dominated by the flux of the MS stars, the algorithm reaches a high degree of accuracy (60%).