A machine learning artificial neural network calibration of the strong-line oxygen abundance

Abstract

The H ii region oxygen abundance is a key observable for studying chemical properties of galaxies. Deriving oxygen abundances using optical spectra often relies on empirical strong-line calibrations calibrated to the direct method. Existing calibrations usually adopt linear or polynomial functions to describe the non-linear relationships between strong-line ratios and Te oxygen abundances. Here, I explore the possibility of using an artificial neural network model to construct a non-linear strong-line calibration. Using about 950 literature H ii region spectra with auroral line detections, I build multilayer perceptron models under the machine learning framework of training and testing. I show that complex models, like the neural network, are preferred at the current sample size and can better predict oxygen abundance than simple linear models. I demonstrate that the new calibration can reproduce metallicity gradients in nearby galaxies and the mass–metallicity relationship. Finally, I discuss the prospects of developing new neural network calibrations using forthcoming large samples of H ii region and also the challenges faced.