Evaluating the Sensitivity of the Age Inferences of Red Giant Stars to Machine Learning Methodology

Abstract

Stellar ages are vital for understanding the formation of our galaxy, but they are among the most challenging parameters to measure. Many authors address this by using machine learning models trained on stars of known age. Here we used data for 351,995 stars from Milky Way Mapper Data Release 19 to explore the sensitivity of the inferred ages to 1) neural network hyperparameters, 2) machine learning architecture, and 3) training set. We find that the resulting ages are generally insensitive to the neural network hyperparameters or the machine learning architecture, but are somewhat sensitive to the training set chosen. We also find that ages for the oldest, coolest, and lowest metallicity stars in the sample are most sensitive to the methodology used and the training set chosen. In general, our analysis suggests that even simple neural network models are sufficient for accurate age inference, but future work expanding the available training sets will be an important component of predicting reliable ages for the full galactic population.