Fifteen million, one hundred seventy-five thousand, one hundred and seven stars — that's exactly how many suns received individual age estimates in the new catalog. But no number is free from a haze of uncertainty: a star's age is not a point on a timeline, but a probability cloud, a fog where the true value hides behind blurred edges. For galactic archaeology, this timestamp is priceless. Without it, the Galaxy's map stays flat. As early as the beginning of the 20th century, Cecilia Payne-Gaposchkin showed that stars are mostly hydrogen, and Hans Bethe explained how nuclear reactions govern their life cycle. Yet unlike position in the sky or brightness, age cannot be measured directly. It is inferred from stellar evolution models or — for a select few — from seismic oscillations (stellar evolution), as done by the William Borucki Kepler and TESS telescopes. But such precise ages are a drop in the ocean compared to the millions of stars for which we only have photometry and spectroscopy data. The European Space Agency's Gaia mission provided low-resolution spectra (XP) for 220 million sources, from which we can extract metallicity and atmospheric parameters. It is these quantities — temperature, surface gravity, iron, carbon, nitrogen, and alpha-element abundances — that carry the imprint of age. The question is how to reconstruct a reliable chronology from these indirect clues.
The researchers assembled an ensemble of 135 569 reference stars with ages known from the APOGEE project. These stars served as a training set for the Machine Learning algorithm NGBoost — a probabilistic cousin of XGBoost that outputs an entire probability distribution instead of a single number. Standard NGBoost, however, assumes training labels are perfectly accurate, which is false: the reference ages themselves have uncertainties. So the loss function was modified to account for label uncertainty by adding the squared label error to the model variance. Moreover, the input parameters (temperature, element abundances) are also known imprecisely. To gauge how this uncertainty seeps into the final age estimate, they used a numerical simulation Monte Carlo method: for each star, 200 perturbed versions of the input data were generated within the error bars, and the predictions were averaged. Thus was born a model that not only names an age but also honestly reports: "plus or minus this much".
On the test set, the predicted ages fell along the line of equality with the reference values, showing a systematic offset of just –0.119 Gyr and a scatter of 1.200 Gyr. Stars with larger uncertainties were generally predicted with larger errors, meaning the model adequately estimates its own confidence. The algorithm was then applied to 15 175 107 stars from the Gaia XP catalog whose parameters fell within acceptable ranges (temperature 3700–6032 K, surface gravity from 0.66 to 4.83 in log scale, metallicity from –0.76 to 0.53). For each star, we now know not only the age but also three uncertainty components: model uncertainty, propagated input error, and total. For instance, a typical total uncertainty is about 1.4 Gyr — roughly a quarter of Earth's age. Comparison with independent catalogs obtained via stellar evolution and spectroscopy methods confirmed agreement: discrepancies fall within 1.5–1.8 Gyr, with the main systematic offsets inherited from the training set's age scale.
The main achievement is not just a number, but an honest error map. Now researchers can select stars with acceptable age precision to study galactic evolution. The probabilistic approach allows us not to mask uncertainty but to work with it: for example, building age distributions in different regions of the Milky Way, keeping in mind that each age is not a pinpoint but a probability cloud. This is especially valuable for star formation history and the chemical evolution of the disk.
The next step is to include additional age indicators in the model, such as lithium abundance or stellar rotation. Moreover, the method can be extended to data from other large surveys, for instance, the Chinese telescope LAMOST or the upcoming 4MOST spectrograph. Refining the age scale with new asteroseismic data will also thin out the probability fog.
The catalog will be useful in any task requiring a time stamp: from searching for traces of swallowed dwarf galaxies to calibrating stellar evolution models. Moreover, the method can be adapted to estimate other hard-to-measure quantities, such as distances or masses.
Plans include integrating stellar kinematics data into the model and, possibly, using more complex distributions for predictions — not just Gaussians, but mixtures of distributions to capture the multimodality of ages on the red giant branch.
And so, peering through the probability fog, astronomers begin to discern the outlines of the Milky Way's history. The work directly tackles the fundamental problem of determining stellar ages and, consequently, reconstructing galactic evolution. Age inaccuracy remains a major obstacle, and honest accounting of uncertainties is a way not to get lost in this fog, but to chart a reliable course through it.
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
This study builds a probabilistic model (NGBoost with a modified loss function) to estimate the ages of 15 million stars using Gaia XP and APOGEE data, accounting for uncertainties in both training labels and input parameters via a Monte Carlo approach.
Where its strength is
The strength of the approach lies in its honest treatment of uncertainties at every stage, from modifying the loss function to propagating errors through Monte Carlo, yielding not just a point estimate but an age distribution with realistic uncertainties for each star.
Where it could go next
Worth trying
Why this may matter
If the method proves reliable for the oldest stars and in the presence of hidden parameters, it could become a standard tool for constructing a temporal map of the Galaxy with honest error bars, which is crucial for testing cosmological scenarios and models of the Milky Way's evolution.
Nearby works in our archive
The archive is crowded around this work: 3 reviews nearby, closest at 0.71. This ground is well walked — and there is plenty to compare against.
The age of the oldest known star in the Milky Way (HD 140283, "Methuselah") was initially estimated at 16 billion years — older than the Universe! The error was later corrected, but the case shows how hard it is to measure age. The new method aims to reduce such mishaps by honestly pointing out the blurriness of the estimate.
The idea of reading a star's age from its spectrum is reminiscent of the tricorder from "Star Trek," capable of determining the composition and age of any object in a second. Reality is still far from such precision, but a catalog of 15 million stars is a step toward a "galactic tricorder."