Is dark matter collisionless? In CDM the shape of the ICL closely follows that of the dark matter halo, but if dark matter self-interacts the inner halo becomes rounder while the ICL stays elongated. Comparing the two in galaxy clusters therefore offers a test of CDM against SIDM.

Can the ICL map the dark matter in clusters? Stars stripped from satellites end up more bound and more centrally concentrated than the stripped dark matter, by an amount set by the infalling satellite population. The ICL therefore jointly constrains the spatial distribution of the dark matter and the assembly history of the cluster.

Simulations that match the integrated properties of dwarf galaxies still differ sharply in their structure. Comparing simulated and observed dwarfs (paper I) and the coupling between their morphology and star formation history (paper II, in preparation) gives a more discriminating test of galaxy evolution models than integrated properties alone.

Surveys such as LSST will image far more galaxies than can be classified by eye. Unsupervised machine learning sorts galaxies into morphological groups without training sets, and recovers known trends of mass, colour and star formation with morphological type.

Tidal features record the merger histories of galaxies. Mock images from NewHorizon show that the 10-year LSST will recover around 80 per cent of the tidal-feature flux around Milky Way-mass and more massive galaxies, but that their classification depends strongly on depth and viewing angle.

Comparing simulations with observations on equal terms requires putting simulated galaxies through the same observational effects as real data. Realistic mock images are used to test galaxy formation models and to prepare low-surface-brightness science with Rubin/LSST and Euclid.