Neural networks sort 400,000 JWST galaxies to trace when spirals and spheroids appeared

The work

Technical · from the record

Deep learning was used to assign about 400,000 COSMOS-Web galaxies brighter than F150W=27 to four morphological classes (spheroids, bulge-dominated, disk-dominated, peculiar), and a fine-tuned Zoobot model was used to estimate the probability that each galaxy hosts a stellar bar. These classifications were combined with SED-fitted stellar masses to measure stellar mass functions for each morphological type between z~0.2 and z~7, separately for quiescent and star-forming galaxies. The authors report that at z>4.5 the massive population (log M*/M_sun > 10) is dominated by disturbed morphologies (~70%) with the remainder very compact (~30%), that Hubble-type morphologies rise in abundance below z~4 and dominate massive galaxies by z~3, and that massive quenched galaxies are predominantly bulge-dominated from z~4 onward while low-mass quenched galaxies are typically disk-dominated. Using bar frequency as a proxy, they estimate that rotating stellar disks have been common (>50%) among massive star-forming galaxies since z~2-2.5.

How AI was used

Two learned classifiers were applied to COSMOS-Web NIRCam cutouts. For global morphology, a convolutional neural network with adversarial domain adaptation transferred CANDELS/HST morphology labels to JWST imaging; the same architecture, 32x32 pixel input size and normalisation as the source work were kept, but the networks were retrained from scratch with COSMOS-Web images as the target domain. Ten trainings with different initialisations and slightly different training sets were run for each of F150W, F277W and F444W, and the four class probabilities were averaged across the ten networks, with their standard deviation used as a robustness measure; the classification filter was chosen by redshift to keep a similar rest-frame band. For internal structure, the pre-trained Zoobot EfficientNetB0 encoder (5.33M parameters, trained on the GZ Evo dataset of 820k images) was fine-tuned on Galaxy Zoo classifications of JWST CEERS images, with three independent fine-tunings per filter. COSMOS-Web cutouts were sized from each galaxy's effective radius and axis ratio and interpolated to 424x424 pixels to match the model input. The fine-tuned model, which outputs Dirichlet parameters emulating volunteer votes, was sampled 100 times assuming 100 volunteers per sample to derive featured, edge-on and bar probabilities, which were then thresholded to select bar candidates. Non-learned steps followed: a half-light-radius cut against the F444W PSF defined a compact class, an exponential fit to the observed bar fraction versus magnitude supplied debiasing weights, and 1/Vmax stellar mass functions were computed and fitted by MCMC with model choice by corrected Akaike information criterion.