学术文献库

We explore the capability of deep learning to classify cosmic structures. In cosmological simulations, cosmic volumes are segmented into voids, sheets, filaments and knots, according to the distribution and kinematics of dark matter (DM), and galaxies are also classified according to the segmentation. However, observational studies cannot adopt this classification method using DM. In this study, we demonstrate that deep learning can bridge the gap between simulations and observations. Our models are based on three-dimensional convolutional neural networks and trained with data of the distribution of galaxies in a simulation to deduce the structure classes from the galaxies rather than DM. Our model can predict the class labels as accurate as a previous study using DM distribution for the training and prediction. This means that galaxy distribution can be a substitution for DM for the cosmic-structure classification, and our models using galaxies can be directly applied to wide-field survey observations. When observational restrictions are ignored, our model can classify simulated galaxies into the four classes with an accuracy (macro-averaged $F_{\rm 1}$-score) of 64 per cent. If restrictions such as limiting magnitude are considered, our model can classify SDSS galaxies at $\sim100~{\rm Mpc}$ with an accuracy of 60 per cent. In the binary classification distinguishing void galaxies from the others, our model can achieve an accuracy of 88 per cent.