Five codes measured galaxy shapes in simulated Euclid images, one using neural networks

The work

Technical · from the record

The Euclid Morphology Challenge compared five surface-brightness-fitting codes (DeepLeGATo, Galapagos-2, Morfometryka, ProFit and SourceXtractor++) on about 1.5 million simulated Euclid VIS galaxies, of which about 350 000 lie above 5σ. The simulated fields comprised analytic single-Sérsic and double-Sérsic profiles rendered with Galsim and a third set of non-analytic galaxies produced by a variational auto-encoder combined with a normalising flow trained on HST images. Measured against the known input parameters, all codes recovered single-Sérsic radius, axis ratio and Sérsic index with bias and dispersion below 10% down to about IE=23, bulge-disc decompositions were less stable, and results on the non-analytic galaxies were typically degraded by a factor of about 3. The authors report that all codes underestimate their own measurement uncertainties by a factor of around 2, and attribute much of the faint-end difference between codes to the priors or initial guesses each adopts.

How AI was used

Two learned components were used. First, a deep generative model built from a variational auto-encoder and a normalising flow (the FVAE) was trained on HST galaxy images to learn a latent representation of real galaxy light profiles, with the latent distribution conditioned on structural parameters and magnitude; it was then run over a parameter catalogue to render a field of non-analytic galaxy images, which were given the same PSF convolution and noise post-processing as the analytic Galsim fields. Second, one of the five participating morphology codes, DeepLeGATo, estimates structural parameters with convolutional neural networks rather than by fitting an analytic model; it was trained on simulated stamps with known true parameters, with separate weights fitted for objects brighter and fainter than magnitude 24.5, and then run over the challenge fields using fixed 64x64 pixel stamps. The remaining four codes are parametric or non-learned fitters. Predictions from all codes were pooled into a common catalogue and compared with the simulation inputs using completeness, bias, dispersion, outlier fraction and a weighted global score computed in bins of apparent magnitude; uncertainty calibration was assessed by checking how often the true value fell inside each code's reported 1σ interval.