This work stands apart from its neighbors in the archive: nearby are studies that use classical relations or machine learning to estimate metallicity, but not to predict luminosity directly. Some colleagues refine distances using Cepheids and red giants, others map the Milky Way, and still others study the chemical evolution of galaxies. Our paper fills a niche: it shows that a neural network can pull absolute brightness out of the shape of a light curve, bypassing the intermediate step with metallicity. This is a different method from what the neighbors use, and it promises greater accuracy at lower observational cost.
For training, they took 63,043 RRab stars and 16,602 RRc stars from the Gaia DR3 catalog. Light curves in the G band were phase-folded with periods from Gaia, aligned by the first Fourier harmonic, and normalized. The model's input consisted of irregular "phase–brightness" sequences, padded or truncated to a fixed length. The architecture is a bidirectional LSTM with two layers and dropout regularization, totaling 134,273 parameters. Training used 80% of the data, validation used 20%, with five-fold cross-validation. For comparison, they ran linear regression, random forest, CNN, RNN, and Transformer. Absolute magnitude labels came from the photometric catalog of Li et al. (2023), but the model never saw metallicity directly, only the shape of the curve.
The model for RRab yielded a root-mean-square error of 0.053 magnitudes, and for RRc, 0.036, which corresponds to distance uncertainties of 2.5% and 1.7% for individual stars. Across 22 globular clusters with RRab and 16 with RRc, the median distance accuracy was 1.0% and 1.7%. For the Large Magellanic Cloud, the distance modulus is 18.498±0.001(stat.)±0.018(syst.), and for the Sculptor dwarf spheroidal galaxy, 19.564±0.003(stat.)±0.019(syst.). Both values agree with previous measurements within 0.7 sigma and improve accuracy by an average factor of 1.8. An ablation analysis showed that the model relies on physically meaningful portions of the curve: for RRab, the rising branch and the minimum matter; for RRc, the rising branch and regions near maximum.
Unlike neighboring studies that either use traditional relations or predict metallicity as an intermediate step, here the neural network links the shape of the light curve directly to luminosity. This eliminates systematics tied to metallicity estimates and opens the door to mass distance determination without spectroscopy. The method can be transferred to other types of variable stars and to data from future sky surveys such as LSST. For science, this is a step toward a unified distance scale that does not depend on the chemical composition of stars.
The next stage is training on multi-band light curves, which will make it possible to account for interstellar reddening and suppress systematic errors even further. With the arrival of LSST and SiTian data, the number of known RR Lyrae stars will grow by orders of magnitude, and the model can be fine-tuned for more distant systems. A separate task is to disentangle Blazhko-type variability, which adds noise to the curves. In the long run, such models will become a standard tool for calibrating the distance scale of the Universe.
The method will influence galactic archaeology, refinements of the Hubble constant, and studies of the structure of the Local Group. It is also useful for planning observations on new telescopes, where a quick distance estimate from photometry is needed.
The immediate steps are applying the model to ZTF data and future Gaia releases, as well as adapting it to infrared light curves. In parallel, it is worth checking how well the model performs on stars with strong Blazhko variability.
For astrophysics, the key unsolved problem is the precise calibration of the distance scale, on which the value of the Hubble constant and the age of the Universe depend. Our work shows that machine learning can reduce systematic errors tied to metallicity and provide an independent check of the Cepheid ladder. This is a direct contribution to resolving the so-called "Hubble tension": the discrepancy between local and cosmological measurements of the expansion rate.
RR Lyrae stars pulsate with periods from 0.2 to 1.2 days, and their brightness varies by 0.2–2 magnitudes. In globular clusters they are so numerous that they are called "beacons of metal-poor old systems."
In Isaac Asimov's novel The Gods Themselves, an alien civilization uses pulsating energy sources for interstellar communication. Our RR Lyrae stars are natural pulsars in the optical range, and their rhythm really could be turned into a distance code.