The stars a galaxy acquired from elsewhere

A galaxy can contain stars born elsewhere. ERGO-ML estimates their share of its stellar mass, returning distributions of possible histories. The January author manuscript applies this approach to more than 750,000 Subaru Hyper Suprime-Cam galaxies.[1]

First make the artificial galaxies look observable

The project’s earlier image-comparison study addressed a necessary bridge. A simulation records how a galaxy assembled, while a telescope records light. The team rendered simulated galaxies as survey images, including observational effects, then trained a network jointly on those images and real observations.[2]

Its contrastive training brings altered views of a galaxy together in a mathematical representation and separates dissimilar objects. Rotations, reflections and modest changes to image framing help it learn structure that survives those changes. Similar-looking simulated and observed galaxies can then occupy nearby positions.[2]

That earlier study also identified simulated galaxies whose images had poor observational counterparts. Such a check matters before transferring a hidden history: visual similarity provides a bridge between the datasets, while a mismatch identifies where the bridge is weak.[2]

A second, invertible network learns simulated histories conditional on the image representation, producing probability distributions. More massive, extended galaxies were inferred to contain larger fractions of externally formed stars.[1]

The history model also takes redshift, luminosity and radius, restoring context lost when images are normalised. In simulations, peak mass-fraction estimates were within ten percentage points for 80% of test galaxies. Last-merger dates were much less reliable. Real galaxies have no equivalent known histories for validation.[1]

Report an error or suggest a correction