Prospect of Deriving Galaxy Properties through Machine Learning: Application to Medium-Band Data from the 7DT

초록

Galaxy evolution studies require the measurement of the physical properties of galaxies at different redshifts. In this work, we build supervised machine learning models to predict the redshift and physical properties (gas-phase metallicity, stellar mass, and star formation rate) of star-forming galaxies from the broad-band and medium-band photometry covering optical to near-infrared wavelengths, and present an evaluation of the model performance. Using 55 magnitudes and colors as input features, the optimized model can predict the galaxy redshift with an accuracy of σ_{(Δ<i>z</i>/1+<i>z</i>)} = 0.008 for a redshift range of <i>z</i> < 0.4. The gas-phase metallicity [12+log(O/H)], stellar mass [log(<i>M</i>_star)], and star formation rate [log(SFR)] can be predicted with the accuracies of σ_NMAD = 0.081, 0.068, and 0.19 dex, respectively. When magnitude errors are included, the scatter in the predicted values increases, and the range of predicted values decreases, leading to biased predictions. Near-infrared magnitudes and colors (<i>H</i>, <i>K</i>, and <i>H</i>–<i>K</i>), along with optical colors in the blue wavelengths (m425–m450), are found to play important roles in the parameter prediction. Additionally, the number of input features is critical for ensuring good performance of the machine learning model. These results align with the underlying scaling relations between physical parameters for star-forming galaxies, demonstrating the potential of using medium-band surveys to study galaxy scaling relations with large sample of galaxies.

키워드