The Cluster's Crucible: Machine Learning Spots Backsplash Galaxies ✛
Context
In the constellation Virgo, one and a half million light-years from the cluster’s searing heart, galaxy NGC 4254 drags behind it a long plume of cooling hydrogen. Such tails are a telltale sign that the galaxy is ramming full-tilt into the hot gas and losing its fuel. But what came before? Has it already plowed through the core and is now, singed, flying back out? Or is this its first encounter with the merciless medium? Galaxies don’t live in isolation: their shape, color, and ability to form stars are governed by their environment. In the centers of galaxy clusters — the megacities of the universe — gas-poor elliptical systems are commonplace, while blue spirals can still be found on the outskirts. But how exactly does the environment reshape galaxies? Some have undergone ‘pre-processing’ in groups or filaments before falling into the cluster. Yet among the apparent ‘newcomers’ lurk a different breed — backsplash galaxies, that have already tasted the crucible and been ejected to the periphery. Without 3D orbits, distinguishing them is impossible, and they contaminate samples, masking the true role of cluster processes. Reliable separation of those entering the cluster for the first time from those already burned has become a pressing need.
Methods
To tackle this, astrophysicists turned to machine learning. A random forest classifier was trained on 324 detailed hydrodynamic simulations of massive clusters (The Three Hundred project). Each galaxy in a projected annulus from 1 to 2 R200 was described by 14 observable parameters: from the obvious — distance to the cluster center, line-of-sight velocity (from spectroscopy), the magnitude gap between the two brightest galaxies (a proxy for dynamical state) — to more nuanced ones, such as the tangential velocity direction (determined from bent radio jets or swept-back gas tails) and the distance to the fourth nearest neighbor. The model was trained on 80% of the data at z=0 and tested on the remaining 20%. A key advantage is that predictions don’t require all 14 parameters: the user uploads only what’s available, and the algorithm builds a new forest on the fly, optimized for that subset. The classification threshold can be tuned, trading off completeness for purity or vice versa.
Results
Trained on the full feature set, the forest delivers 84% accuracy — much like an expert coffee taster who flawlessly tells apart two similar blends. It identifies backsplash galaxies with 75% purity and completeness, and their counterparts with 88% each. If only the three most basic quantities are kept (distance, radial velocity, and magnitude gap), accuracy drops to 74%, but this still markedly outperforms traditional methods that use sharp cuts in phase space: those achieve 61% purity with a mere 40% completeness. Adding the sky-projected velocity direction boosts accuracy to 81%; information about neighbors provides a modest improvement. Morphological parameters in current simulations are noisy and contribute little, but that will change with higher-resolution runs. The real breakthrough is the soft decision boundary. Unlike previous approaches, the model yields a continuous probability. By shifting the threshold, one can obtain a high-purity (75%) sample of backsplash galaxies at the cost of dropping completeness to 18%, or an almost complete (85%) sample, albeit with contamination. Astronomers choose the trade-off depending on the science goal. A striking example is the Virgo cluster. Ten galaxies with torn HI tails in its outer corona, for which the direction of motion could be determined, are classified as first-timers. This is direct evidence that the hot intracluster medium strips cold gas almost immediately upon entry, and after one passage little remains.
Implications
The tool radically changes the landscape of galaxy evolution studies. No longer must we guess what caused quenching of star formation in the outskirts — pre-processing in a group or a scorching in the cluster core. By comparing clean populations, we can directly work out which processes and on what scales turn spirals into lenticulars. NGC 4254 and its companions, then, are not singed wanderers but debutantes whose first rendezvous with the crucible is only just beginning.
Future development
The next goal is to extend the method to more distant clusters (z > 0) and to other environments, including filaments and groups where pre-processing occurs. Models are already being prepared that, from environmental density and kinematics, will recognize not only cluster backsplash galaxies but also galaxies that have felt the influence of more modest companions. In parallel, new simulations with higher resolution and alternative physics prescriptions are anticipated, which will allow colors, metallicities, and gas masses to be included in the training, making the classifier robust to the details of baryonic physics.
Impact
The method will be in demand for next-generation spectroscopic surveys (WEAVE, 4MOST, Euclid), where every correctly classified galaxy counts. It directly impacts dark matter physics (backsplash galaxies lose their dark halo) and the understanding of star formation mechanisms in extreme environments.
Next steps
The immediate task is to apply the classifier to large spectroscopic catalogs and compare star formation rates, morphology, and gas content in backsplash and first-infall galaxies. In parallel, data from other hydrodynamic codes (not just Gizmo-Simba) will be incorporated to guarantee independence from numerical recipes.
Key open problems
The work touches on several fundamental questions: how does the environment suppress star formation? What are the timescales for cold gas removal? At what point does a galaxy lose its ability to form stars — already in a group, immediately after entering a cluster, or on a second pass? The answers feed into the broader problem of cluster assembly and the role of the cosmic web in galaxy evolution.
The knowledge machine's view
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
The study trained a random forest classifier to distinguish galaxies that are falling into a cluster for the first time from those that have already passed through the center. The model uses up to 14 observable parameters, achieving 84% accuracy, and can make predictions even with incomplete feature sets. Applying it to galaxies with gas tails confirmed that these galaxies are indeed entering the cluster for the first time.
Where its strength is
The strength of the work lies in its flexible use of incomplete data: the algorithm builds a forest tailored to the specific set of available parameters, making the method applicable to heterogeneous observational samples. This paves the way for large-scale classification without needing all 14 measurements.
Where it could go next
- The method can be extended to other indicators of environmental interaction: in our archive, there is a nearby work that reconstructs a galaxy's history from scars on its disk. Combining morphological and kinematic features could increase the purity of returner selection.based on: How Scars on the Disk Reveal a Galaxy's Past arXiv:2607.19537
- It would be interesting to see how the classifier performs on galaxies with suppressed star formation: one neighboring work shows that a central bar can quench stars. Testing the hypothesis that some 'newcomers' with gas deficiency have actually already passed through the core but were mistakenly classified as first-infall could refine the model.based on: A Spinning Top at the Galaxy’s Core Snuffs Out Stars arXiv:2601.08791v1
- The approach of training on simulations and predicting with incomplete data resonates with a work where a neural network searches for anomalous quasars. One could adapt the anomaly detection idea: train the model on typical trajectories and identify objects that fit neither the newcomer nor the returner class.based on: The Puzzle That Doesn't Fit: Neural Network Searches for Anomalous Quasars arXiv:2607.17728v1
Worth trying
- Test adding the curvature of the gas tail as a feature in the classifier, serving as a proxy for the direction of motion in the plane of the sky.
- Check whether information about the local environment, from the work on contour maps of star clusters, improves the classification.
- Build predicted-class maps for known clusters and compare them with the distribution of galaxies with suppressed star formation.
Why this may matter
Reliably separating first-infall galaxies from returners is crucial for understanding how quickly the cluster environment quenches star formation and removes gas. The proposed method, which does not require complete data, can be applied to large surveys where some parameters are missing, and could help refine the timescales of galaxy evolution in dense environments.
Nearby works in our archive
- Contour Map of Stellar Nurseries: Computer Vision Finds Clusters in the Fireworks Galaxysimilarity 0.66
- How Scars on the Disk Reveal a Galaxy's Pastsimilarity 0.61
- A Galaxy That Slept for Half a Billion Yearssimilarity 0.60
- The Galactic Family Tree: How Star Chemistry Tells the Storysimilarity 0.60
- Invisible Conductors of Galactic Diskssimilarity 0.59
- A Spinning Top at the Galaxy’s Core Snuffs Out Starssimilarity 0.59
- The Puzzle That Doesn't Fit: Neural Network Searches for Anomalous Quasarssimilarity 0.59
The closest work in the archive sits at 0.66 by meaning — the usual level for our corpus.
Original abstract
as written by the authors arXiv:2607.26136v1
The galaxy population in the outskirts of a cluster contains members that have been pre-processed in groups and filaments, as well as backsplash galaxies -- those that have recently passed through the cluster's center. However, disentangling these two pathways is challenging observationally. In this work, we present a machine-learning-powered model, trained on simulations of galaxy clusters from The Three Hundred suite of simulations, which can identify individual backsplash galaxies in astronomical observations. This model can build samples of backsplash galaxies with a purity and completeness of up to ~70%, and galaxies on their first infall with a purity and completeness of over 80%. It can be tuned to optimise either of these two metrics, and can be used with any combination of a set of observable quantities. We have also applied this model to galaxies with asymmetric HI distributions in the Virgo Cluster, and have demonstrated that these galaxies are all likely approaching the cluster for the first time. This supports the idea that cold gas is removed from these galaxies soon after entering a cluster, and demonstrates how this classifier can provide a better understanding of which properties of galaxies are caused by a previous passage through a cluster. We have made this model publicly available in the form of a web app, with a link in the Conclusions of this paper.
Contrary to expectations, 91% of backsplash galaxies in the simulations passed right through the very cluster center — inside 0.7 R200 — before bouncing back. Not all of them manage to escape far: only 10% ‘splash’ beyond 2 R200.
Key numbers
- Model accuracy (14 parameters): 84%
- Backsplash purity/completeness (baseline version): 59% / 61%
- First-infall purity/completeness (baseline version): 81% / 80%
- Fraction of backsplash with pericenter < 0.7 R200: 91%
- Galaxies with HI tails in Virgo classified as first-timers: 10