The Milky Way has been eating its neighbors for billions of years — and getting away with it, because the evidence disappears. When a dwarf galaxy falls into the Milky Way, gravity shreds it across the sky, scattering its stars like a broken necklace until they look indistinguishable from every other halo star. Now, a team from Istanbul University has written the forensic methodology to read those scattered stars — and today their findings are out in a peer-reviewed journal with four previously unknown victims named.
A study published July 29, 2026 in the Publications of the Astronomical Society of the Pacific (PASP) reports the discovery of five new substructures in the Milky Way’s inner stellar halo, using an unsupervised machine learning pipeline applied to data from nearly 1.4 million stars. Four are classified as robust remnants of ancient dwarf galaxies. The fifth is a tentative candidate for something rarer: the scattered debris of a dissolved massive globular cluster — a star city that fell apart so completely it left only a chemical ghost. Together they raise the known census of accreted structures in our galaxy’s halo by roughly 27 percent.
Chemical DNA vs. a Shredded Library
To understand why finding these structures is hard, it helps to understand what happens when a galaxy merger ends. Two galaxies collide in slow motion over hundreds of millions of years. The smaller one — the satellite — gets stretched by tidal forces into ever-thinner threads of stars. Eventually the threads disperse. In the outer halo, where orbits are slow and long, those threads can remain detectable for billions of years as coherent “stellar streams.” In the inner halo — closer to the Galactic center, where orbits are fast and crowded — the shredding accelerates. The process is called phase-mixing: stars from the same progenitor disperse in orbital phase until no spatial or kinematic grouping remains.
For decades, astronomers hunted these structures in “integrals-of-motion” space — plotting orbital energy against angular momentum and looking for clumps. That worked remarkably well in the outer halo. In the inner halo, it fails, because debris from multiple ancient mergers stacks on top of each other in energy-angular momentum space. Stars from entirely different progenitors end up on nearly identical orbits.
The Istanbul team’s solution: stop relying only on orbital information. Use chemistry too.
Stars are born in molecular clouds with a specific chemical fingerprint — specific ratios of iron, magnesium, aluminum, manganese, silicon — and those ratios are largely preserved in the star’s photosphere for its entire life. Two stars that formed in the same dwarf galaxy share a chemical heritage that gravity cannot scramble, even after billions of years. This is the logic of chemical tagging, formalized by Freeman and Bland-Hawthorn in 2002.
How the Pipeline Works: A Twelve-Dimensional Search
PhD candidate Furkan Akbaba and Dr. Olcay Plevne began with spectroscopic data from the SDSS-V Milky Way Mapper’s nineteenth data release (DR19), a Q1-ranked survey that delivered atmospheric parameters and elemental abundances for approximately 1.4 million stars using twin spectrographs mounted on the 2.5-meter (8.2 ft) Apache Point Observatory telescope in New Mexico and the Du Pont telescope at Las Campanas Observatory in Chile. They cross-matched this against the European Space Agency’s Gaia DR3 satellite, which provides positions, proper motions, and parallaxes for approximately 1.5 billion stars.
After applying quality filters — keeping only red giant branch stars with signal-to-noise ratios above 50, reliable distances within about 5 kiloparsecs (16,300 light-years) of the Sun, and clean measurements in all six chemical dimensions — the working sample shrank to 41,928 stars.
Step one was chemical pre-selection. The team used Principal Component Analysis on six abundance ratios — iron, magnesium, aluminum, manganese, nickel, and silicon — to compress the chemical information and then applied K-means clustering. This step isolated 2,185 stars most likely to be “ex-situ” (born outside the Milky Way and later captured): they showed the low-aluminum, high-magnesium-to-manganese ratios characteristic of stars formed in chemically inefficient dwarf galaxy environments, where star formation was slow and supernova enrichment limited, as defined by the Horta et al. chemical selection criteria.
Step two was where the AI earned its paycheck. The team built a twelve-dimensional feature space — six chemical dimensions plus six dynamical ones (total orbital energy, angular momentum, eccentricity, maximum orbital height, and two orbital action variables derived from Gaia astrometry). They compressed this twelve-dimensional cloud into three dimensions using UMAP (Uniform Manifold Approximation and Projection), a manifold-learning algorithm that preserves neighborhood relationships even as it collapses high-dimensional data into something the next algorithm can see. Then HDBSCAN — Hierarchical Density-Based Spatial Clustering of Applications with Noise — automatically identified dense clumps within the resulting three-dimensional map, without being told in advance how many clusters to expect or where to look.
The key advantage of UMAP over older dimensionality reduction methods like t-SNE is that it preserves both local and global structure, meaning that clusters that are genuinely distinct in twelve dimensions appear genuinely distinct in three. HDBSCAN’s advantage over K-means is that it labels outliers as noise rather than forcing them into clusters — avoiding false detections.
“We didn’t tell the AI ‘find this, find that,'” Akbaba told Anadolu Agency. “It searched completely on its own, and after it found the 15 known structures, we saw that four additional structures emerged — ones that didn’t match the literature. That’s when we understood it was a discovery.”
The Known Structures Came Back First
The pipeline’s integrity check was reassuring: it successfully recovered nine kinematic groupings corresponding to seven previously catalogued halo substructures, including Gaia-Enceladus/Sausage (the largest known ancient merger, believed to have collided with the Milky Way eight to eleven billion years ago and responsible for most of the inner halo’s metal-poor stars), the Helmi Streams, Sequoia, Thamnos, Heracles, LMS-1, and Arjuna. The recovered properties closely matched published reference values — confirmation that the twelve-dimensional framework was working as intended.
Beyond those known structures, five HDBSCAN groups matched nothing in the existing catalogue.
Portrait of Four Ghost Galaxies — and One Star City’s Ghost
The Istanbul team named the new structures FO1 through FO5, after their own initials. Statistical validation — including a permutation-based compactness test in the full twelve-dimensional space, silhouette analysis, and thirty full pipeline repetitions with bootstrap resampling — confirmed four of the five as robust.
FO1 is the most metal-rich of the group, with an average iron abundance ([Fe/H]) of −0.83 ± 0.08, considerably higher than most accreted halo populations. It follows a tightly bound, low-inclination prograde orbit and achieved the strongest bootstrap stability in the study at 82.7 percent — meaning that in 82.7 percent of pipeline re-runs on resampled data, the same ten stars clumped together again. Its compact orbit deep in the Galactic potential suggests it may trace debris from a relatively massive ancient merger that settled far into the Milky Way’s interior.
FO3 is the most dynamically unusual. Despite sitting at low orbital energy in the inner halo, it moves on a nearly circular orbit — an eccentricity of just 0.38, against the typical values of 0.7 to 0.9 seen in most accreted populations. This “dynamically cold” orbit, combined with a chemical signature nearly indistinguishable from the Helmi Streams, creates a puzzle: was FO3 originally related to the Helmi progenitor and later circularized by the evolving Galactic potential, or did it arrive along a different, unusually gentle trajectory? Its low orbital height above the Galactic plane — just 1.12 kilometers per second in maximum excursion units (Z_max = 1.12 kpc, meaning it stays within a narrow disk-like plane) — deepens the mystery.
FO4 displays enhanced alpha elements — elevated magnesium ([Mg/Fe] = 0.26 ± 0.03) and silicon ([Si/Fe] = 0.27 ± 0.03) — pointing to a progenitor galaxy where star formation ended early, before slower-burning supernova processes could dilute the alpha-element budget. Its orbit overlaps in some respects with Gaia-Enceladus/Sausage, but its markedly more negative orbital energy places it deeper in the Galactic potential, suggesting it traces a chemically distinct system rather than a simple extension of the dominant merger.
FO5 is perhaps the most scientifically significant of the four robust candidates — because of what it disproves, not just what it finds.
When an AI Corrects Another AI’s Classification: FO5 and Shiva
FO5 occupies a region of orbital energy and angular momentum space that substantially overlaps with the recently proposed Shiva structure, identified in 2024 by Malhan and Rix and thought by some researchers to represent proto-Galactic material — stars formed inside the primordial Milky Way, not captured from outside. Roughly 39 percent of FO5’s member stars fall within the Shiva selection boundary.
But FO5’s chemistry tells a completely different story. Its aluminum abundances ([Al/Fe] = −0.14 ± 0.08) are consistently below the threshold that distinguishes accreted (ex-situ) stars from native (in-situ) populations — and crucially, this is true star-by-star for all eleven FO5 members located inside the Shiva polygon, not just as an average. Shiva, by contrast, shows median aluminum near zero with roughly half its stars above zero — consistent with rapid, high-density enrichment in a proto-Galactic environment. FO5’s depleted manganese ([Mn/Fe] = −0.46 ± 0.13) further signals inefficient chemical enrichment in a low-mass progenitor system, the hallmark of a captured dwarf galaxy rather than native Milky Way material.
The implication extends beyond FO5 itself: if an accreted structure can be dynamically indistinguishable from Shiva in orbital space while chemically confirming an external origin, it raises the question of whether other stars currently classified as proto-Galactic Shiva material have been contaminated by similarly accreted debris. A purely dynamical selection — the kind that was necessary before chemo-dynamical methods were available — could not separate them. Without the chemical dimensions, FO5 would likely have been absorbed into Shiva and its external origin never detected.
Reading a Dissolved Star City: FO2
FO2, the tentative fifth candidate, is the outlier. Its ten member stars show a striking nitrogen enrichment averaging [N/Fe] = +0.83 ± 0.16 — more than twice the nitrogen abundance of typical Gaia-Enceladus stars. This nitrogen spike, paired with enhanced aluminum and depleted carbon, is the chemical calling card of “second-generation” stars born inside massive globular clusters, where earlier stellar generations polluted the surrounding gas with nitrogen-rich ejecta before the next round of star formation began.
Globular clusters are the oldest and densest stellar structures in the galaxy — some harbor millions of stars in a sphere just a few hundred light-years across. Unlike dwarf galaxies, globular clusters have almost no dark matter and are therefore more vulnerable to tidal stripping. When one falls deep into the Milky Way, the tidal forces can shred it completely, scattering its stars. FO2’s chemical signature suggests that is exactly what happened here.
FO2 failed several of the pipeline’s stability tests — its bootstrap co-clustering fraction was just 16.1 percent, compared to 82.7 percent for FO1 and 71.6 percent for FO3. This weakness is itself consistent with the interpretation: a dissolved globular cluster’s debris would form a diffuse, hard-to-pin-down haze rather than a tight clump, precisely the shape of a weak statistical signal in a clustering algorithm tuned to find dense overdensities.
Why the Inner Halo Was the Last Frontier
The five new structures share one defining property: all sit in the deeply bound inner halo, with total orbital energies at or below −1.8 × 10⁵ km² s⁻² — placing them closer to the Galactic center than most known accreted structures. This region is the most crowded and chemically confused part of the Milky Way’s stellar graveyard, which is exactly why it was the last to yield its secrets.
“We’re studying the history of how galaxies form,” Plevne explained to Anadolu Agency. “By looking at how our galaxy formed, we’re investigating how other galaxies in the universe formed. Each new structure we find adds something to our understanding of galaxy evolution.”
Prior to this work, roughly fifteen accreted substructures had been identified in the Galactic halo. The Istanbul team’s four robust new detections bring that total to approximately nineteen, with FO2 as a tentative twentieth — a roughly 27 percent expansion of the known inventory.
The individual membership lists remain small: FO1 and FO2 contain just ten confirmed member stars each. The full spatial extent and detailed chemical gradients of each structure remain unconstrained. But the discovery opens a specific research program: go back through expanded spectroscopic catalogues and find more members.
That expansion is already on the way. SDSS-V will continue releasing data from its Milky Way Mapper program. The 4MOST spectrograph at ESO’s Visible and Infrared Survey Telescope for Astronomy in Chile is designed specifically for follow-up stellar spectroscopy. Extended Gaia radial velocity measurements — currently limited to stars brighter than about magnitude 16 — will cover fainter and more distant halo members. And NASA’s Nancy Grace Roman Space Telescope, scheduled to launch August 30, 2026, carries a 300-megapixel infrared camera with a field of view more than one hundred times larger than Hubble’s — specifically engineered for wide-area surveys of the kind of stellar populations that populate the outer halo.
How Does an AI Actually Find a Ghost Galaxy?
The technical pipeline at the center of this paper answers a question that has troubled galactic archaeologists for years: how do you find a coherent population of stars in a dataset where the stars are not spatially coherent and are not kinematically coherent — but do share a chemical heritage?
The answer is dimensionality. In two-dimensional orbital energy versus angular momentum space, accreted structures overlap badly in the inner halo. In six chemical dimensions alone, you can identify ex-situ stars (those born outside the Milky Way) but not necessarily separate individual progenitors from each other — many dwarf galaxies produce similar broad abundance patterns. Combining both, in twelve dimensions simultaneously, creates a space where the two constraints reinforce each other: a group of stars must be compact in chemistry and in orbital properties to survive the clustering step.
UMAP’s role is to make that twelve-dimensional space legible to HDBSCAN. It works by building a fuzzy topological graph of nearest-neighbor relationships in the high-dimensional space, then finding the low-dimensional layout that best preserves those neighborhood relationships. The result is a three-dimensional map where stars that were close neighbors in twelve dimensions remain close neighbors — even if the three dimensions themselves are not straightforwardly interpretable as “chemistry” or “orbit.”
HDBSCAN then finds density peaks in that map without needing to be told how many peaks to expect. It works hierarchically, running the equivalent of thousands of DBSCAN passes at different density thresholds and selecting the clusters that remain stable across the widest range. Stars that belong to no stable cluster are labeled as noise — a crucial feature for a blind survey, because it means false detections have a natural escape valve.
The fact that this pipeline first recovered nine previously known groupings with properties matching the literature — before turning up five unknowns — is the most important validation step in the paper. An algorithm that finds known structures correctly before finding new ones has cleared the basic scientific bar.
What the Milky Way’s History Is Teaching Us
Every new accreted structure in the Milky Way’s halo is a data point in the larger question of how galaxies assemble in the universe. The standard cosmological model predicts that large galaxies grow by eating smaller ones — hierarchical structure formation — and that the Milky Way should have consumed dozens of dwarf galaxies over the past twelve billion years. Finding the chemical fingerprints of individual consumed galaxies is the closest observational test that cosmological model has at human scales.
Before Gaia, astronomers could identify perhaps five or six accreted structures. The satellite’s combination of positions, motions, and distances for more than a billion stars transformed that to fifteen. Adding chemical abundances from SDSS-V, processed through a twelve-dimensional AI pipeline, now pushes the count further — and the inner halo, previously impenetrable, is no longer a blank spot on the map.
The four ghost galaxies discovered by Akbaba and Plevne — FO1, FO3, FO4, and FO5 — do not yet have names beyond their initials, do not have known progenitor masses, and do not have confirmed infall times. What they have is a chemical identity that no amount of orbital mixing can erase. That is, it turns out, enough.
Frequently Asked Questions
How does machine learning find galaxies that no longer exist?
The trick is that galaxy remnants are not completely gone — their stars are. When the Milky Way absorbs a smaller galaxy, it shreds it, but the individual stars survive and retain the chemical abundances they had when they formed. Machine learning can simultaneously process six chemical measurements and six orbital measurements per star — twelve dimensions of information — and identify groups of stars that cluster together in that high-dimensional space. Stars from the same extinct galaxy form compact clumps because they share both a chemical heritage and a family of similar orbits. A conventional two-dimensional search in orbital space alone would miss these clumps because billions of years of gravitational mixing have scrambled their positions and orbits into indistinguishability; the chemical dimensions are what allow the algorithm to separate them.
Does this discovery change what we know about the Shiva structure in the Milky Way’s inner halo?
Potentially, yes. The new structure FO5 overlaps with Shiva — a substructure proposed in 2024 as native proto-Galactic material — in orbital energy and angular momentum space. But FO5’s chemical abundances, particularly its consistently low aluminum-to-iron ratios, identify it as accreted material from outside the Milky Way rather than native Galactic stars. Every one of FO5’s member stars inside the Shiva selection region shows this accreted chemical signature. This means that purely dynamical searches — which selected the Shiva population without chemical information — may have captured a mixture of proto-Galactic stars and genuine accreted debris. Separating them requires exactly the kind of chemo-dynamical approach this paper demonstrates.
What is a dissolved globular cluster, and why is finding one in the inner halo significant?
Globular clusters are some of the oldest structures in the universe — dense balls of hundreds of thousands to millions of stars that formed very early in cosmic history. Unlike dwarf galaxies, they have almost no dark matter, so they are more vulnerable to being torn apart by the Milky Way’s tidal gravity. When a massive globular cluster is disrupted, its stars scatter through the halo. The structure FO2 shows a striking nitrogen enrichment — a chemical signature that forms only in the dense interiors of massive globular clusters, where early-generation stars pollute later-generation stars with nitrogen-rich material. Finding this signature in the inner halo provides evidence that globular cluster disruption — not just dwarf galaxy mergers — contributes meaningfully to building the Milky Way’s innermost stellar population.
What telescopes and surveys will follow up on these discoveries?
The current member lists are small — FO1 and FO2 contain just ten stars each — because the analysis was limited to stars within about 5 kiloparsecs (16,300 light-years) of the Sun with high-quality spectroscopic data. Three major expansions are underway: continued data releases from the SDSS-V Milky Way Mapper, the upcoming 4MOST spectrograph survey at ESO in Chile (designed for exactly this kind of stellar population follow-up), and extended Gaia radial velocity data covering fainter stars. NASA’s Roman Space Telescope, scheduled to launch August 30, 2026, will add an infrared wide-field imaging capability that will help identify candidate member stars for spectroscopic follow-up in the outer halo where these structures may have additional debris.
