Yazan Bakhshwin

Artificial Intelligence Specialization · Cohort 2026

The Iron in Their Rhythm

The metallicity ([Fe/H]) of stars in the Milky Way provides a fossil record of its assembly, and the ancient RR Lyrae variables are among its best tracers. Spectroscopic metallicities exist for only a small fraction of these stars, but since their pulsational properties depend on iron abundance, the light curves collected by photometric surveys offer a path to metallicity at survey scale. Using thousands of RR Lyrae with DESI spectroscopic metallicities cross-matched to ZTF photometry, we train random forest and gradient boosted tree regressors to recover spectroscopic quality metallicities from light curves alone, modeling the two main pulsation subtypes (RRab and RRc) separately. We compare nested feature sets, from pulsation period and amplitude to the added Fourier decomposition of the light curve shape, and benchmark them against a convolutional neural network that reads the phase-folded light curves directly. Fourier shape parameters improved predictions more than any data quality cut we tested, the network performed comparably using only the folded photometry, and combining the two gave our most accurate predictions so far. We also test how well the learned relations carry over to stars from other surveys, where differences between samples remain an open challenge. Our findings suggest that photometry alone could provide homogeneous metallicities for the more than 10⁵ RR Lyrae observed by current and upcoming surveys.

Read the final report (PDF)