A machine learning approach to infer the accreted stellar mass fractions of central galaxies in the TNG100 simulation
We propose a random forest (RF) machine learning approach to determine the accreted stellar mass fractions ($f_\mathrm{acc}$) of central galaxies, based on various dark matter halo and galaxy features. The RF is trained and tested using 2,710 galaxies with stellar mass $\log_{10}M_\ast/M_\odot>10.16$ from the TNG100 simulation. Galaxy size is the most important individual feature when calculated in 3-dimensions, which becomes less important after accounting for observational effects. For smaller galaxies, the rankings for features related to merger histories increase. When an entire set of halo and galaxy features are used, the prediction is almost unbiased, with root-mean-square error (RMSE) of $\sim$0.068. A combination of up to three features with different types (galaxy size, merger history and morphology) already saturates the power of prediction. If using observable features, the RMSE increases to $\sim$0.104, and a combined usage of stellar mass, galaxy size plus galaxy concentration achieves similar predictions. Lastly, when using galaxy density, velocity and velocity dispersion profiles as features, which approximately represent the maximum amount of information extracted from galaxy images and velocity maps, the prediction is not improved much. Hence the limiting precision of predicting $f_\mathrm{acc}$ is $\sim$0.1 with observables, and the multi-component decomposition of galaxy images should have similar or larger uncertainties. If the central black hole mass and the spin parameter of galaxies can be accurately measured in future observations, the RMSE is promising to be further decreased by $\sim$20%.