Abstract We present a machine learning approach to estimate stellar atmospheric parameters (Teff, log g, and Fe/H) for stars in the J-PLUS DR3 footprint using photometric and astrometric data from J-PLUS, Gaia, and CatWISE. Our models are based on the gradient boosting algorithm LightGBM and were trained using extinction-corrected photometry and absolute magnitudes derived from Gaia parallaxes. Through a systematic feature selection process, we identified the most informative features for each parameter, showing that using the full set of selected features yields the best overall performance. Evaluation against spectroscopic values from LAMOST DR8 reveals that our models achieve competitive mean absolute errors of 42 K, 0.06 dex, and 0.06 dex for Teff, log g, and Fe/H, respectively. We applied the trained models to a curated sample of 154 stars belonging to open clusters and moving groups. While our metallicity estimates are mildly but systematically underestimated by −0.13 dex, we attribute this to the compression of training data in a compact and non-linear region of the feature space. Despite this bias, the relative metallicity patterns among clusters remain consistent with literature values, demonstrating the robustness of our predictions for comparative studies. Our method provides a reliable alternative for estimating stellar parameters using multi-band photometry combined with astrometric information, offering further applications to surveys such as S-PLUS and J-PAS. Future improvements for parameter predictions will focus on additional magnitude combinations, predicting other parameters such as α/Fe, and techniques for cluster member selection.
Machado-Pereira et al. (2026) studied this question.