15 175 107 stars. Each one got an age. But no number is free of uncertainty: a star's age isn't a nail hammered into the timeline: it's a probability cloud, a fog hiding the true value. For galactic evolution, such a timestamp is priceless: without it, the map of the Milky Way remains flat. As early as the twentieth century, Cecilia Payne-Gaposchkin showed that stars are mostly hydrogen, and Hans Bethe explained how nuclear reactions govern their life cycles. Yet unlike brightness or position in the sky, age can't be measured directly. It's inferred from models of stellar evolution or, for a lucky few, from seismic oscillations (asteroseismology, i.e., age measurement via a star's vibrations), as done by the Kepler telescope under William Borucki and TESS. But such precise ages are a drop in the bucket compared to the millions of stars for which we have only photometry (brightness measurements) and spectroscopy (light broken down by wavelength). The European Space Agency's Gaia mission provided low-resolution spectra for 220 million sources. From these we extract temperature, surface gravity, metallicity (the fraction of elements heavier than helium)—parameters that carry the imprint of age. The question is how to reconstruct chronology from these indirect clues.
The researchers trained on an ensemble of 135 569 stars from the APOGEE project, whose ages were known relatively well. These benchmarks were used to train NGBoost — a machine-learning algorithm that outputs not a single number but an entire probability distribution. However, standard NGBoost assumes the labels are perfectly accurate, but in reality, the benchmark ages themselves contain errors. So the loss function was modified: the total model variance now includes the squared error of the training age: \(\sigma_{\text{total}}^2 = \sigma^2 + \sigma_{A,\text{obs}}^2\). In other words, the model's uncertainty is combined with the known inaccuracy of the labels. And the input data (temperature, chemical composition) are also known with errors. To estimate how this uncertainty leaks into the result, they applied a Monte Carlo method (simulation with repeated random sampling): for each star they generated 200 variants of input parameters, slightly perturbed within the error bounds, and averaged the predictions. Thus was born a model that not only gives an age but also honestly reports the spread.
On the test set, the predicted ages fell along the line of perfect agreement, showing a systematic bias of just –0,119 billion years and a scatter of 1,200 billion years. Stars with larger uncertainties also received wider predicted intervals — the model adequately gauged its own confidence. The algorithm was then applied to 15 175 107 Gaia stars with parameters in certain ranges (temperature 3700–6032 K, log surface gravity 0,66–4,83, metallicity from –0,76 to 0,53). For each, we now know not only the age but also three error components: the model's own, the one propagated from the input data, and the total. The typical total uncertainty is about 1,4 billion years — roughly a quarter of the Earth's age.
Comparisons with independent catalogs obtained via asteroseismology and spectroscopy confirmed the agreement: discrepancies fall within 1,5–1,8 billion years, and the main systematic shifts are inherited from the age scale of the training set. The key outcome isn't just the numbers: it's the honest map of errors. Now astronomers can select stars with acceptable age precision for studying galactic evolution. The probabilistic approach doesn't mask uncertainty; it works with it — building age distributions for different regions of the Milky Way while keeping in mind that each age is not a nail but a probability cloud. This is especially valuable for star formation history and the chemical evolution of the disk. Looking ahead, the model could incorporate additional indicators (lithium abundance, stellar rotation) and use data from other large surveys like LAMOST or the future 4MOST spectrograph. Refining the age scale with new asteroseismic data will narrow the probabilistic fog. Thus, peering through it, researchers begin to discern the outlines of the Milky Way's history. Age uncertainties remain the main obstacle, but honestly accounting for uncertainties is a way not to get lost in the fog, but to chart a reliable course through it.
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 mistake was later corrected, but the incident highlights just how hard it is to measure age. The new method aims to reduce such mishaps by honestly pointing out the fuzziness 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 any object's composition and age in a second. Reality is still far from that kind of precision, but a catalog of 15 million stars is a step toward a "galactic tricorder."