Mapping the Metal of the Milky Way: How Gaia’s Spectra Help Us Understand Giant Stars

This study, led by Lin Yang, investigates the chemical compositions of roughly 20 million giant stars in the Milky Way using data from the European Space Agency’s Gaia mission. A star’s metallicity—its content of elements heavier than hydrogen and helium—is a key clue in understanding when and where it formed. The focus on “giant stars,” which are luminous and in later stages of evolution, helps researchers trace the formation history of the galaxy. Especially valuable are metal-poor stars, which are some of the oldest stars in the universe.

Using Gaia XP Spectra to Estimate Metallicity

Gaia’s XP spectra are low-resolution data products that represent how starlight is spread out by wavelength, similar to color measurements. These spectra are not traditionally ideal for measuring metallicity because they lack the fine detail that high-resolution spectroscopy provides. However, recent machine learning techniques have opened new doors. The team introduced an innovative tool called the Uncertainty-Aware Cost-Sensitive Neural Network (UA-CSNet) to estimate metallicities from Gaia XP spectra while accounting for the challenges of noisy data and the rarity of very metal-poor stars.

Preparing and Processing the Data

The model was trained using two trusted spectroscopic catalogs—PASTEL and SAGA—which provide metallicity values for well-studied stars. The researchers carefully selected stars based on their color, brightness, position in the sky, and dust extinction to ensure accurate input data. They corrected the Gaia XP spectra for systematic errors and interstellar dust, and then converted them into a uniform format for input into the neural network. The resulting dataset included over 1,000 stars for training and several hundred more for independent testing.

How the Neural Network Works

UA-CSNet features two main branches: one estimates the metallicity (\[Fe/H]) and the other estimates the uncertainty in that prediction. By accounting for both types of uncertainty—epistemic (from limited data) and aleatory (from measurement noise)—the model improves both the accuracy and trustworthiness of its predictions. The team also used a “cost-sensitive” approach, giving extra importance to rare types of stars like the extremely metal-poor ones to help the model learn more effectively from them.

Testing the Model’s Accuracy

To test how well the model works, the researchers compared their predictions to metallicities from other star catalogs, including SDSS/SEGUE and LAMOST, as well as other Gaia-based studies. The UA-CSNet estimates showed close agreement, especially at the low-metallicity end. The model was also tested on stars from known clusters, where all members are expected to share similar properties, and again performed well. Importantly, their estimates remained accurate even when the stars had high carbon abundances, which tend to confuse other photometric methods.

Building the Galactic Metallicity Map

The final version of the model was applied to over 31 million stars, with a clean, high-quality sample of about 20 million selected for reliability. This catalog includes over 1 million very metal-poor stars and hundreds of thousands of extremely metal-poor stars. The researchers also mapped how these metallicities are distributed in the Galaxy, revealing trends with distance from the center and height above the Galactic plane—offering insights into how different parts of the Milky Way formed and evolved.

Conclusion and Future Applications

The UA-CSNet demonstrates that Gaia’s XP spectra, when combined with advanced machine learning, can be used to accurately estimate stellar metallicities—even in the very metal-poor regime. This is a significant step forward because it allows astronomers to study a much larger sample of stars than traditional spectroscopic surveys alone. The publicly available catalog created by this work will serve as a foundational dataset for many future studies on the structure, formation, and chemical evolution of our galaxy.

Source: Yang