Neural networks estimate dark matter halo masses from a galaxy's neighbours

The work

Technical · from the record

Fully-connected neural networks were trained on simulated galaxy catalogues to predict the peak mass of a galaxy's host dark matter halo from observationally accessible information: the galaxy's own stellar mass plus two standard environment measures, projected distances to its fifty nearest neighbours and counts of neighbours in fixed cylindrical apertures. Networks were trained on the SMDPL simulation box and tested on the separate Bolshoi-Planck box, with a baseline built by interpolating the average stellar mass - halo mass relation. The best-performing network had an overall root-mean-squared error of 0.20 dex, and 0.17 dex when applied to only the central halos in the test data. Retraining with inputs masked indicated that, beyond the galaxy's stellar mass, the halo mass information was concentrated in the innermost region (about 1 Mpc) around the halo centre, and the nearest-neighbour and cylinder-count networks gave similar errors.

How AI was used

Dark matter halos from the SMDPL and Bolshoi-Planck simulations were populated with galaxies using the UniverseMachine empirical model, and galaxies with stellar mass above 10^9 M_sun were selected. For each galaxy, feature vectors were built from the projected distances and number-density-rescaled stellar masses of its fifty nearest neighbours within a 1000 km/s redshift offset, and from counts of neighbours in cylinders of radius 0.5, 1, 2 and 5 h^-1 Mpc binned by redshift separation in 250 km/s intervals up to 2000 km/s; the target galaxy's own stellar mass, converted to a cumulative number density, was appended as an input. Features and labels were individually standardised, with log scaling for masses, and a weighting scheme derived from the binned halo mass function was tested alongside unweighted training. Three fully-connected regression networks, taking kNN inputs (length 101), cylinder counts (length 33), or both (length 134), were trained with supervised learning in Keras with a TensorFlow backend to predict peak halo mass, using a mean absolute error loss, ReLU activations, Adam, an initial learning rate of 0.001, batch size 128, and up to 50 epochs with early stopping after a patience of 10 and a ten-epoch warm-up. Depth, layer-narrowing factor, an optional skip connection from the stellar mass input, activation, optimiser, learning rate and batch size were compared on the SMDPL validation split, which was 30% of the box volume, with 70% used for training. Trained networks were then run over the Bolshoi-Planck box, which was never seen during training, and compared with interpolation of the average SMDPL stellar mass - halo mass relation. To attribute the information used, networks were retrained with the same structure and hyperparameters after masking neighbours beyond a given k, the larger cylinder radii, or the larger redshift bins, and separate networks were trained and tested on central halos only.