ERGO-ML: Uncovering the assembly histories of HSC galaxies directly from images via simulation-based inference and cosmological simulations

A fundamental prediction of the LambdaCDM cosmology is the hierarchical build-up of structure and therefore the successive merging of galaxies into more massive ones. As one can only observe galaxies at one specific time in cosmic history, this merger history remains in principle unobservable. This study, part of the ERGO-ML series (Extracting Reality from Galaxy Observables with Machine Learning), aims to circumvent this problem by inferring posteriors of merger and assembly related quantities for images of observed galaxies by learning them from survey realistic mock images of cosmological simulations. Our exploration is centered on galaxies within the stellar mass range of 10^9 to 10^12 M⊙ at redshifts z=0.1‑0.4, using the TNG50 and TNG100 cosmological simulations of the IllustrisTNG project (https://www.tng-project.org/) to generate survey-realistic galaxy mock datasets of Hyper Suprime-Cam (HSC) observations (https://www.tng-project.org/explore/gallery/bottrell23/). To ensure comparability between the simulated and observed image domains for this task, we employ contrastive learning techniques, specifically the self-supervised contrastive learning method NNCLR. This leads to a direct comparison of a substantial sample of simulated and observed galaxies based on their stellar-light images in the g, r, and i-bands, leveraging the complete information content of the observations. The outcome is a 256-dimensional representation space that comprehensively encapsulates all relevant observable galaxy properties. Our analysis reveals a notable alignment between a significant majority (≳70 per cent) of TNG galaxies and observed HSC images, validating the realism of these simulations. Intriguingly, a subset of simulated galaxies displays characteristics such as larger sizes, steeper Sersic profiles, smaller Sersic ellipticities, and larger asymmetries, which are identified as potentially unrealistic and can be therefore excluded from the inference. For the next step we use a conditional invertible neural network, a type of normalizing flows. With this we can not only learn point-predictions but arbitrarily shaped posteriors of critical statistics describing the merger and assembly history of galaxies from IllustrisTNG and then infer those for observed HSC galaxies based on the previously learned image representations. Examples of these statistics include the stellar exsitu fraction, as well as details like the mass and time of the last major merger. By directly inferring these quantities from HSC images, our approach serves as a groundbreaking method, bridging the gap between simulated and observed galaxy images and providing a comprehensive understanding of the manifold galaxy populations and its evolutionary history.