Teaching Galaxies When to Arrive: Using Machine Learning to Time the Fall of Milky Way Satellites

Understanding when small dwarf galaxies fell into the Milky Way is crucial for piecing together how galaxies form and evolve. In this paper, Kim et al. focus on predicting the infall time of dwarf satellite galaxies, the moment when a small galaxy first crossed into the gravitational influence of a much larger host like the Milky Way. This timing matters because once a dwarf galaxy falls in, environmental effects such as tidal forces and gas stripping can shut down its star formation, leaving lasting fingerprints in its stellar populations. Traditional methods for estimating infall times rely on reconstructing galaxy orbits, but these approaches are computationally expensive and uncertain. To address this challenge, the authors introduce a faster and more interpretable machine-learning approach.

Challenges with Traditional Approaches

The authors begin by motivating why infall times are so difficult to measure. Dwarf galaxies experience many physical processes, supernova feedback, cosmic reionization, and interactions with their host galaxy, that complicate their histories. Standard orbital calculations often assume simplified models of the Milky Way and struggle to track galaxies accurately more than about 6 billion years into the past. Kim et al. argue that instead of directly modeling orbits, one can take advantage of observable properties that are already linked to infall. In particular, they emphasize the connection between infall time and quenching time, defined here as τ90: the lookback time when a galaxy formed 90% of its stars.

Simulating Satellite Galaxies with A-SLOTH

To build their model, the authors use A-SLOTH, a semi-analytic model that simulates Milky Way–like galaxies and their satellites while incorporating detailed physics such as reionization and stellar feedback. From 30 Milky Way–like host galaxies, they generate over 12,000 satellite galaxies and classify them by stellar mass. For each satellite, they select three input features that are commonly available from observations: τ90, stellar mass (M★), and stellar metallicity ([Fe/H]). The target quantity is the infall time, defined as the first moment the satellite crossed within the host’s virial radius. Importantly, satellites that previously belonged to another group before falling into the Milky Way are treated separately, since this “group preprocessing” strongly affects their evolution.

A Machine-Learning Model for Infall Times

The machine-learning model itself is based on LightGBM, a gradient-boosting decision tree algorithm designed to handle complex, nonlinear relationships efficiently. After training on 80% of the simulated data and testing on the remaining 20%, the model achieves a mean squared error of about 5 Gyr when predicting Milky Way infall times for satellites without prior group membership. By analyzing feature importance, the authors find that τ90 is by far the most influential predictor, contributing much more information than stellar mass or metallicity. This confirms the physical intuition that when a galaxy stopped forming stars is closely linked to when it first encountered a massive host environment.

What the Results Reveal About Different Galaxies

The results also reveal important differences between satellite populations. Intermediate-mass satellites are predicted most accurately, while very low-mass satellites show larger scatter, especially for recent infall times. This is because low-mass galaxies are extremely sensitive to cosmic reionization and supernova feedback, which can quench star formation even before infall. Heavy satellites, on the other hand, are massive enough to shield their gas and continue forming stars despite infall, weakening the correlation between τ90 and infall time. When the authors instead train the model to predict the first infall into any host, including earlier group environments, the accuracy improves dramatically, with the error dropping to about 1.7 Gyr. This highlights the dominant role of the earliest infall event in shutting down star formation.

Applying the Model to the Real Local Group

Finally, Kim et al. apply their model to real satellite galaxies of the Milky Way and Andromeda (M31). For Milky Way satellites, the predicted infall times generally agree with previous observational studies, though some well-known outliers such as CVn II and UMa I remain challenging. For M31 satellites, where no direct infall-time measurements exist, the model predicts a clear trend between τ90 and infall time that closely matches expectations from simulations. Overall, this work demonstrates that machine learning, when guided by physical insight, can provide a powerful and efficient new way to estimate the assembly history of galaxies, using quantities that astronomers can actually observe.

Source: Kim