New AI Study Reveals How Galaxies Reveal Their Violent Histories

Galaxies constantly evolve in response to their surrounding environments, changing their shapes as they accrete gas, interact with cosmic neighbors, and plunge through dense clusters. Tracing these morphological transformations provides astronomers with a vital window into the physical processes governing galaxy formation and cosmic history. However, separating the diverse mechanisms that disrupt a galaxy remains a major challenge in modern observational astronomy, as different physical triggers can leave remarkably similar footprints on starlight.

Gravitational interactions and major mergers violently perturb both stellar disks and gaseous reservoirs, generating complex tidal tails, bridges, and asymmetric features. Conversely, ram-pressure stripping acts primarily on a galaxy’s interstellar medium when it plows headfirst through dense intra-cluster or intra-group plasma. This hydrodynamic pressure strips gas away to form one-sided tails and extraplanar star-forming knots while often leaving the underlying stellar disk relatively undisturbed. Distinguishing these phenomena across vast imaging archives is critical for mapping cosmic evolution, though optical morphology alone often highlights that a galaxy is disturbed without immediately revealing why.

To tackle this problem on a massive statistical scale, an international team of researchers analyzed data published in arXiv, examining wide-field optical images drawn from Data Release 10 of the DESI Legacy Imaging Surveys. The study focuses on a parent sample of more than 1.7 million galaxies out to a redshift of $z < 0.15$, utilizing rigorous magnitude and angular-size thresholds to minimize resolution-driven biases and signal-to-noise limitations.

Decoding Starlight with Non-Parametric Indices

Instead of relying on rigid analytical models, the authors utilized the Galaxy Morphology Extractor package to measure ten quantitative, non-parametric indices for reference populations of regular spirals, regular ellipticals, galaxy mergers, and optical jellyfish candidates. These indices evaluate core structural properties including concentration, multiple forms of asymmetry, pixel-wise smoothness, second-order moments of light, Shannon entropy, the Gini coefficient, and gradient pattern asymmetries.

The research team ranked all possible two-dimensional combinations of these indices using a cross-validated Bayesian kernel-density classifier. Their analysis revealed that diagnostic planes combining Shannon entropy with the spatial distribution metric $M_{20}$ or shape asymmetry indices perform exceptionally well at separating regular galaxies from disturbed systems, achieving balanced accuracies of up to 0.77. Using these findings, the team established calibrated diagnostic boundaries that incorporate explicit uncertain regions to account for natural overlap between populations.

The Structural Divergence of Mergers and Stripped Systems

A central breakthrough of the new work involves comparing galaxies undergoing ram-pressure stripping directly against major mergers. While both populations populate disturbed regions of structural parameter space, they exhibit fundamentally different levels of morphological coherence. The data show that jellyfish candidates occupy a relatively compact, well-defined zone characterized by lower concentration, high asymmetry, and elevated entropy. Their consistent appearance stems from the directional nature of ram-pressure stripping, which displaces gas and newborn stars to one side while preserving a recognizable host disk.

Mergers, by contrast, spread broadly across multiple regions of parameter space, frequently overlapping with regular spirals and ellipticals. Because a merger represents an extended evolutionary sequence dependent on mass ratios, orbital geometries, gas fractions, and viewing angles, it produces no single compact morphological signature. Consequently, while automated tools can flag merging activity with enhanced probability, isolating pure merger samples remains inherently difficult based on projected broad-band morphology alone.

Multivariate Machine Learning at Scale

To leverage the full dimensionality of the data, the team trained calibrated LightGBM classifiers that simultaneously process all ten non-parametric indices. In a three-class experiment separating regular ellipticals, regular spirals, and a combined perturbed population, the multi-index model achieved an overall accuracy of 0.92 and a balanced accuracy of 0.89. When expanding to a four-class model that treats mergers and jellyfish candidates separately, the classifier attained an overall accuracy of 0.91.

The resulting probability catalogs provide the astronomical community with powerful statistical tools. By delivering robust, survey-specific morphological probabilities for nearly 1.8 million individual sources, the work sets a new benchmark for large-scale structural analysis, paving the way for future multi-wavelength studies combining wide-field imaging with deep spectroscopic surveys.

The study was posted to the arXiv preprint server on October 23, 2026.