GyroNet: Well-calibrated Rotation-based Stellar Ages from a Machine Learning and Bayesian Framework with Gaia DR3 Features

Abstract

Stellar age is among the most difficult fundamental parameters to measure, yet it anchors the interpretation of astrophysical systems across many contexts. Gyrochronology provides an empirically grounded route to stellar ages via rotation, leveraging photometric rotation periods routinely measured in time-domain surveys. Existing tools map stellar rotation to age, but differ in calibration, input features, and uncertainty treatment. We present GyroNet, a machine learning and Bayesian framework that combines rotation periods with Gaia DR3 photometric observables to infer stellar ages and return full posterior distributions with quantified uncertainties. Trained on ∼8000 open-cluster stars from the ChronoFlow compilation, GyroNet couples neural spline flows with Bayesian reweighting over auxiliary Gaia DR3 features to sharpen posteriors and suppress unreliable parameter space. An adaptive reliability mask suppresses out-of-distribution predictions, mitigating cases where proximity-driven photometric artifacts mimic youth. We benchmark GyroNet against gyro-interp and ChronoFlow from 2 Myr to 3.87 Gyr using the continuous ranked probability score and Wasserstein distance, which jointly assess point accuracy and posterior shape. GyroNet achieves the lowest scores on both metrics across most of this range with gyro-interp performing best at ≲100 Myr and ChronoFlow between 700 Myr and 1 Gyr, where benchmark clusters are sparse. GyroNet is publicly released as a pip-installable Python package that supports three usage modes: full auxiliary-feature input, automated Gaia DR3 archive fetch via DR3 source ID, and a baseline rotation-period-plus-color mode.