Neural network turns galaxy shape distortions into maps of invisible cluster mass

The work

Technical · from the record

The authors trained a convolutional neural network to turn noisy weak-lensing shear measurements into projected mass (convergence) maps, using mock catalogs built for the field of view and depth of the Vera C. Rubin Observatory's LSST. On 200 held-out test fields the reconstructed convergence correlated with the truth with a pixel-by-pixel regression slope of about 0.70, compared with about 0.47 for the earlier H21 CNN, and the kappa-kappa auto-correlation function was recovered except below about 2 pixels. Cluster detection from the CNN maps reached about 75 per cent completeness at about 10^14 solar masses, against about 35 per cent for the KS93 method, and all three methods reached 100 per cent completeness above about 3.5x10^14 solar masses. Applied to Subaru/Hyper Suprime-Cam observations of the Coma cluster, the reconstruction gave an NFW fit of M200 = 7.4 (+2.4/-3.1) x10^14 solar masses and showed three filament branches also found by earlier matched-filter and shear-peak analyses.

How AI was used

Publicly available MassiveNuS ray-tracing convergence maps for a source plane at z=2.5, covering a 3.5 by 3.5 degree field at 512x512 resolution, were used to simulate mock weak-lensing shear catalogs with a source density of about 33 per square arcminute, an intrinsic shape dispersion of 0.24 and magnitude-dependent measurement errors. Each mock field was reduced to a three-channel 512x512 input of mean ellipticity components and mean measurement error, with no Gaussian smoothing of the inputs. The 10,000 maps were split into 7,000 training, 2,800 validation and 200 test fields, with rotations and flips expanding the training set. A convolutional network of eight 2D convolutional layers, with differently sized filters in the initial layers concatenated together, skip connections, LReLU activations, dropout and constant 512x512 internal resolution (5.6x10^6 free parameters), was trained for 1,000 epochs with Adam at a learning rate of 1e-5 and gradient clipping, using a weighted mean-squared-error loss modified to raise the weighting of high-convergence pixels. The trained network was then run over the held-out test fields, over 1,000 mock null fields, over inputs masked with an irregular Subaru/HSC footprint, and over a published Subaru/Hyper Suprime-Cam shear catalog of the Coma field; peak finding, projected-mass measurement, TreeCorr correlation functions and DisPerSE filament extraction were applied to the resulting maps as non-learned post-processing, with KS93 and the earlier H21 CNN run as comparisons.