Letter to the Editor

A neural network approach to determining photometric metallicities of M-type dwarf stars

¹ Departamento de Física de la Tierra y Astrofísica & IPARCOS-UCM (Instituto de Física de Partículas y del Cosmos de la UCM), Facultad de Ciencias Físicas, Universidad Complutense de Madrid, 28040 Madrid, Spain
² Institut de Ciències de l’Espai (ICE, CSIC), Campus UAB, c/ de Can Magrans s/n, 08193 Cerdanyola del Vallès, Barcelona, Spain
³ Institut d’Estudis Espacials de Catalunya (IEEC), C/Esteve Terradas, 1, Edifici RDIT, Campus PMT-UPC, 08860 Castelldefels, (Barcelona), Spain
⁴ Centro de Astrobiología (CSIC-INTA), ESAC Campus, Camino bajo del castillo s/n, 28692 Villanueva de la Cañada, Madrid, Spain

^⋆ Corresponding author: chrduque@ucm.es

Received: 24 March 2025
Accepted: 18 May 2025

Abstract

Context. M dwarfs are the most abundant stars in the Galaxy and serve as key targets for stellar and exoplanetary studies. It is particularly challenging to determine their metallicities because their spectra are complex. For this reason, several authors have focused on photometric estimates of the M-dwarf metallicity. Although artificial neural networks have been used in the framework of modern astrophysics, their application to a photometric metallicity estimate for M dwarfs remains unexplored.

Aims. We develop an accurate method for estimating the photometric metallicities of M dwarfs using artificial neural networks to address the limitations of traditional empirical approaches.

Methods. We trained a neural network on a dataset of M dwarfs with spectroscopically derived metallicities. We used eight absolute magnitudes in the visible and infrared from Gaia, 2MASS, and WISE as input features. Batch normalization and dropout regularization stabilized the training and prevented overfitting. We applied the Monte Carlo dropout technique to obtain more robust predictions.

Results. The neural network demonstrated a strong performance in estimating photometric metallicities for M dwarfs in the range of −0.45 ≤ [Fe/H] ≤ +0.45 dex and for spectral types as late as M5.0 V. On the test sample, the predictions showed uncertainties down to 0.08 dex. This surpasses the accuracy of previous methods. We further validated our results using an additional sample of 46 M dwarfs in wide binary systems with FGK-type primary stars with well-defined metallicities and achieved an excellent predictive performance that surpassed the 0.1 dex error threshold.

Conclusions. This study introduces a machine-learning-based framework for estimating the photometric metallicities of M dwarfs and provides a scalable data-driven solution for analyzing large photometric surveys. The results outline the potential of artificial neural networks to enhance the determination of stellar parameters, and they offer promising prospects for future applications.