Neural networks weigh cluster X-ray maps to estimate galaxy cluster masses

The work

Technical · from the record

Convolutional neural networks were trained to predict galaxy cluster masses (M500c) from mock eROSITA X-ray photon maps built on the Magneticum hydrodynamical simulation, which include background emission, instrument response, point spread function and AGN contamination. Single-band bolometric photon maps gave a predictive mass scatter of 17.8%, compared with about 26% for the idealised bolometric X-ray luminosity proxy and about 34% for the best realistic scalar proxy, the photon count within R500c; splitting the maps into soft, medium and hard energy bands reduced the scatter to 16.2%, and adding mock spectroscopic galaxy dynamics gave 15.9%. A gradient-based saliency analysis found that the networks assign much lower per-pixel importance to cluster centres than the photon-count scaling relation does, with the median proxy saliency about ten times the CNN saliency at 0.07 R500c, and assign AGN-dominated pixels lower median saliency than ICM-dominated pixels.

How AI was used

Mock eROSITA observations of 3285 Magneticum clusters were produced with PHOX for intra-cluster medium emission, a projection of simulated AGN sources, and SIXTE for background, instrument response and point spread function. Photon lists were binned into 128x128 maps over a 2.3 h-1 Mpc aperture, normalised for redshift and luminosity distance and log-scaled, in both a single 0.5-10 keV band and three separate soft, medium and hard bands; matching mock spectroscopic galaxy catalogues were smoothed with Gaussian kernel density estimators and sampled onto 64x64x64 phase-space grids. A feed-forward convolutional network of four convolutional layers, two max-pooling layers and four dense layers was trained from scratch with a mean squared error loss on logarithmic M500c, using the Adam optimiser at learning rate 1e-3, ReLU activations, L2 regularisation, a 90/10 training-validation split and early stopping after 20 epochs without improvement. The same architecture was applied to three-channel multi-band inputs, and a joint model combined a 2D convolutional X-ray extractor with a 3D convolutional dynamical extractor feeding a shared dense network trained end to end. Cross-validation folds were assigned by splitting the simulation volume into eight equal quadrants to avoid shared formation environments. Scalar proxy baselines were fitted with a power-law multi-property scaling and covariance model, with an exhaustive search over observable combinations. Trained networks were then interrogated with absolute-gradient saliency maps, compared against a closed-form saliency derived for the photon-count proxy, binned radially, and averaged over binary masks separating ICM-dominated from AGN-dominated pixels.