🔭 astrophysics

Estimating stellar metallicities from Gaia DR3 XP data using LAMOST DR10

This paper presents a highly accurate XGBoost model trained on LAMOST DR10 spectroscopic data to estimate stellar metallicities from Gaia DR3 XP spectra, significantly outperforming existing Gaia-derived values and enabling precise chemical mapping of the Milky Way's radial metallicity gradient.

Original authors: Divyansh Srivastava, Andrzej Niedzielski, Rodolfo Smiljanic

Published 2026-07-22

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Divyansh Srivastava, Andrzej Niedzielski, Rodolfo Smiljanic

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the Milky Way not as a static picture, but as a bustling, ancient city that has been expanding and evolving for billions of years. To understand how this city was built, astronomers act like cosmic archaeologists, digging through layers of stars to find clues about their origins. One of the most important clues is a star's "metallicity." In astronomy, "metals" don't just mean iron or gold; they refer to any element heavier than hydrogen and helium. Think of these metals as the "dirt" or "ash" left over from previous generations of stars that exploded. A star with high metallicity is like a house built with recycled bricks from older buildings, while a metal-poor star is like a house built from fresh, raw materials. By measuring how much metal is in a star, scientists can trace where it was born, how old it is, and how the galaxy mixed its ingredients over time.

For a long time, measuring this metallicity was like trying to guess the ingredients of a soup by looking at a blurry photo of the pot. You needed a high-resolution telescope to get a clear "taste" (a detailed spectrum) of the star, which was slow and expensive. However, a space telescope called Gaia has been taking a census of over a billion stars, capturing their light in a way that is fast and covers the whole sky, but with a resolution that is a bit fuzzy. The challenge has been: how do we turn these fuzzy, low-resolution snapshots into accurate chemical recipes for hundreds of millions of stars?

This paper presents a clever solution to that puzzle. The authors, D. Srivastava, A. Niedzielski, and R. Smiljanic, built a sophisticated computer brain (a machine learning model) to act as a translator. They taught this brain using a massive library of "perfect" recipes from the LAMOST survey, which has high-quality data for about 1.2 million stars. Once the computer learned the patterns, they let it look at the fuzzy Gaia data and predict the metallicity for stars that LAMOST never saw.

The results are a game-changer. When the team tested their new predictions against stars they hadn't shown the computer before, the model was incredibly accurate, with an error margin of just 0.052 dex. To put that in perspective, the standard method used by Gaia itself (called GSP-Phot) was about five times less accurate, with an error of 0.242 dex. It's like upgrading from a blurry guess to a sharp, high-definition measurement. The model works so well that when they applied it to open clusters—groups of stars born together that should all have the exact same metallicity—their predictions were much closer to the true values than previous methods.

However, the authors are careful to note that their tool isn't magic. They found a quirky "hump" in the data: for stars in a cluster, the predicted metallicity would wiggle slightly depending on the star's color, even though the real metallicity should be flat. They traced this back to the limitations of the low-resolution Gaia data itself, suggesting it's a fundamental limit of the technology rather than a mistake in their code. They also warn that for the most metal-poor stars (the "raw material" ones), the model might underestimate how poor they are, so those results should be treated as a lower limit rather than a precise number.

Despite these small quirks, the model successfully mapped the chemical landscape of our galaxy's disk. It confirmed that the inner part of the galaxy has a different chemical gradient than the outer part, with a "break" in the pattern occurring around 5.9 kiloparsecs from the center. This matches what other, more expensive studies have found, proving that this new, fast method is reliable. The team has made their model and the resulting catalog of metallicity estimates available to everyone, giving astronomers a powerful new tool to study the history of the Milky Way without needing to wait for a slow, expensive telescope to point at every single star.

Technical Summary: Estimating Stellar Metallicities from Gaia DR3 XP Data using LAMOST DR10

Problem Statement
The Gaia mission has provided astrophysical parameters for over a billion stars, yet the metallicity estimates ([M/H]) derived from the Gaia DR3 GSP-Phot module suffer from significant systematic biases dependent on color, magnitude, and extinction. While previous efforts have attempted to re-derive metallicities using Gaia's low-resolution Blue and Red Photometer (BP/RP or "XP") spectra anchored to spectroscopic surveys like APOGEE, there remains a need for a homogeneous, high-accuracy metallicity catalog for the vast population of AFGK stars, particularly one that leverages the extensive optical coverage of LAMOST to better constrain the metal-poor tail and dusty disk regions.

Methodology
The authors present a machine learning approach to estimate stellar iron abundances ([Fe/H]) using only Gaia DR3 inputs, trained on high-quality spectroscopic labels from LAMOST DR10.

Training Data: The study utilizes a curated subset of the LAMOST DR10 low-resolution AFGK catalogue. Quality cuts were applied to ensure high signal-to-noise ratios and well-constrained atmospheric parameters, resulting in a training set of approximately 1.2 million stars. The labels were inverse-variance weighted means of repeat measurements, restricted to the range $-2.5 \le [Fe/H] \le 1.0$ .
Input Features: The model ingests 139 features derived exclusively from Gaia DR3. These include:
- 110 XP spectral coefficients (55 BP and 55 RP), rescaled to a standard flux level ( $G=15$ ) to isolate spectral shape from brightness.
- Broadband photometry and colors ( $G, G_{BP}, G_{RP}$ ).
- Astrometry and quality indicators (parallax, proper motions, RUWE, flux excess).
- Luminosity and kinematic proxies (pseudo-absolute magnitudes, reduced proper motion).
- Extinction terms (GSP-Phot dust estimates $A_G$ and $E(BP-RP)$) to correct for interstellar reddening, explicitly excluding GSP-Phot atmospheric parameters to avoid injecting priors.
- Galactic latitude ( $b$ ) as a weak positional indicator.
Model Architecture: A gradient-boosted decision-tree regressor (XGBoost) was employed. The training utilized a squared-error objective with early stopping based on validation set performance.
Bias Mitigation: To address the "color-metallicity" prior inherent in magnitude-limited field star samples (which causes spurious metallicity trends in chemically homogeneous systems like open clusters), the authors implemented a sample weighting scheme. This weights stars within specific color-magnitude diagram (CMD) cells to balance the representation of metal-rich and metal-poor populations, thereby weakening the learned correlation between CMD position and metallicity.
Uncertainty Estimation: Per-star uncertainties are derived from the standard deviation of an ensemble of five XGBoost models trained on random resamples of the data.

Key Results
The model was validated against held-out LAMOST data, external surveys (GALAH DR4 and APOGEE DR17), and 46 open clusters.

Performance on Held-Out Data: On the LAMOST test set, the model achieves a Mean Absolute Error (MAE) of 0.052 dex and an $R^2$ of 0.94, with negligible bias. This represents a roughly five-fold improvement over the native Gaia DR3 GSP-Phot metallicities (MAE = 0.242 dex) on the same stars.
External Validation: The estimates transfer well to independent surveys:
- GALAH DR4: MAE = 0.066 dex.
- APOGEE DR17: MAE = 0.068 dex.
- The residuals show no strong trends with effective temperature, surface gravity, or color, though a slight "shrinkage" toward the mean is observed for extremely metal-poor stars ($[Fe/H] < -1.0$) in the GALAH comparison, suggesting predictions in this regime should be treated as lower bounds.
Open Clusters: For 26 open clusters with high-resolution reference metallicities, the median difference between the model's estimates and spectroscopic values is 0.041 dex. This outperforms both the GSP-Phot estimates (0.248 dex) and a previous APOGEE-trained XGBoost model (0.067 dex).
Galactic Structure:
- Vertical Gradient: The model recovers a vertical metallicity gradient of $-0.237$ dex kpc $^{-1}$ , consistent with previous G-dwarf studies.
- Radial Gradient: Applied to a kinematically selected thin-disk sample, the model reveals a broken radial gradient: an inner slope of $+0.119 \pm 0.018$ dex kpc $^{-1}$ and an outer slope of $-0.058 \pm 0.002$ dex kpc $^{-1}$ , with a break near 5.9 kpc. This aligns with recent high-resolution spectroscopic findings.
- Cluster Gradient: The open-cluster radial gradient is measured at $-0.066$ dex kpc $^{-1}$ , consistent with APOGEE-based measurements.

Limitations and Systematics
The authors identify a systematic "hump" in the predicted metallicity as a function of color within open clusters (peaking near $G_{BP}-G_{RP} \approx 1.0$ ). This feature, with a median amplitude of $\sim0.07$ dex, is present even in chemically homogeneous clusters and is attributed to the intrinsic limitations of inferring metallicity from low-resolution ( $R \sim 50$ ) spectrophotometry where temperature and metallicity are degenerate. While the weighting scheme reduced this effect, it was not eliminated. Additionally, predictions for stars beyond 1 kpc in open clusters show a systematic underestimation of metallicity.

Significance
The paper demonstrates that Gaia DR3 XP data, when anchored to the extensive LAMOST DR10 spectroscopic labels via XGBoost, can produce stellar metallicities with an accuracy of 0.05 to 0.07 dex for AFGK stars. This accuracy is sufficient for chemical studies of the Milky Way's disk, including the recovery of complex radial gradients and the characterization of open clusters. The authors provide a public catalogue and trained model, offering a significant improvement over the native Gaia GSP-Phot products and serving as a competitive alternative to other XP-based metallicity estimates until the release of Gaia DR4. The work highlights that while the model is robust for field stars and broad population studies, users must account for the residual color-dependent systematics when analyzing chemically homogeneous populations with narrow color ranges.

More like this