A faint glow of gamma rays hangs over the center of the Milky Way, stretching across thousands of light-years and refusing to give up its source. For years, one of the strongest counterarguments to a dark matter explanation was that the signal looked more like a swarm of hidden stars. A new analysis now complicates that case.

The signal is known as the Galactic Center Excess, or GCE, a roughly spherical surplus of gamma rays seen around the heart of our galaxy by NASA’s Fermi Large Area Telescope. Since it emerged in the data more than a decade ago, it has fueled one of astrophysics’ most stubborn debates.

One possibility is that the glow comes from dark matter particles annihilating each other, a long-theorized process that could produce gamma rays. Another is that the excess comes from huge numbers of millisecond pulsars, rapidly spinning neutron stars too dim to be picked out one by one.

“Interpreting the signal is particularly difficult because the Galactic Center is an exceptionally bright and crowded region of the gamma-ray sky,” said Florian List, a study author and researcher at the University of Vienna.

Earlier work favoring the pulsar idea focused on how gamma rays were distributed across the sky. If the excess came from many unresolved point sources rather than a smooth signal, the map should show subtle departures from the random pattern expected from diffuse emission such as dark matter. Those studies argued that the excess had that point-source character.

But they came with tradeoffs. To keep the calculations manageable, they generally ignored two things: correlations between neighboring pixels and the energy carried by individual photons.

The new study tried to move past both limits. The team trained a convolutional neural network on 1 million simulated Fermi gamma-ray maps, including 790,000 for training, 200,000 for validation, and 10,000 for testing. The simulated observations covered 10 logarithmically spaced energy bins from 2 to 20 GeV and were built to reflect the telescope’s changing response across energies.

That mattered because the Galactic Center Excess has a distinct spectrum, and different possible sources can leave different energy signatures. By combining spatial and spectral information for the first time, the network could ask a more demanding question: not just whether the glow looked clumpy, but how bright those clumps would need to be.

The answer pushed the source population to far lower brightness than many earlier studies suggested.

Where previous analyses pointed to unresolved sources just below Fermi’s detection threshold, the new method found that any point sources responsible for the excess would need to be so dim that, collectively, they become almost indistinguishable from smooth emission.

“Our new analysis shows that the sources would have to be so faint that they would be almost indistinguishable from the emission expected from annihilating dark matter,” said Nick Rodd, a study author and scientist at Lawrence Berkeley National Laboratory.

That shift has major consequences for the pulsar interpretation. According to the paper, the median source-count distribution would require roughly 200,000 sources in the Galactic Center region to explain the signal, while even the 90 percent upper quantile still implies about 35,000. That is far above the few hundred to few thousand sources assumed in some earlier scenarios. For comparison, one influential 2016 non-Poissonian template fit implied about 200 sources.

The new model also found that the inferred source population sits almost entirely below the one-photon threshold, meaning the sources are so faint they often would not even contribute a single detected photon on average. At that level, the difference between unresolved point sources and truly smooth emission becomes extremely hard to tell apart.

The team tested that directly. A separate neural network trained to distinguish point-source emission from Poisson-like emission could exclude only 3 percent of the map as inconsistent with Poisson emission at 95 percent confidence. In practical terms, that means the signal looks much more compatible with the kind of smooth gamma-ray glow expected from dark matter than earlier analyses had concluded.

That does not make this a dark matter discovery.

“The origin of the Galactic Center Excess is one of the longest-running debates in astrophysics,” List said. “Our work does not show that dark matter is responsible for the signal. However, it suggests that it is still too early to rule out this possibility.”

That line is important because the new paper is as much about removing an objection as it is about proving a case. For years, one of the strongest arguments against dark matter was the claim that the excess looked statistically like a collection of unresolved point sources. This analysis weakens that argument by showing the supposed sources would need to be so faint that they blend into something very close to smooth emission.

The study drew on 812 weeks of Fermi data collected between August 4, 2008, and February 23, 2024. The researchers examined the Inner Galaxy, defined as the region within 25 degrees of the Galactic Center, while masking parts of the bright Galactic plane and known cataloged sources. They modeled several kinds of gamma-ray emission, including diffuse Galactic backgrounds, isotropic emission, the Fermi bubbles, a disk population of point sources, and the central excess itself.

Notably, adding energy information changed the inferred source-count distribution for the Galactic Center Excess much more than it did for the disk population, where point sources are already expected. That contrast adds weight to the idea that the central excess is a special case rather than a generic failure of the method.

The authors are careful not to oversell the result. The biggest unresolved issue is still background modeling.

The Galactic Center is messy. Diffuse gamma-ray emission from ordinary astrophysical processes is bright, structured, and difficult to model perfectly. The team found that changing the diffuse background model could move the median prediction for the number of sources from roughly 10,000 to roughly 100,000, even if the sources still remained well below the older bright-source picture.

That sensitivity means the core debate is not over. The paper argues that more work is needed on improved diffuse models, broader energy ranges, and machine-learning methods that more directly build the smooth-emission degeneracy into the analysis.

Even so, the result matters because it reopens space for a possibility many researchers had treated as increasingly cornered. The Milky Way’s central glow still might come from an enormous hidden population of millisecond pulsars, but that explanation now looks more demanding than before. Dark matter, meanwhile, remains very much alive in the argument.

The immediate impact is not a dark matter detection, but a reset in how the Galactic Center signal is interpreted. By folding photon energies into the analysis, the study weakens a major statistical case against dark matter and raises the bar for pulsar-based explanations.

That gives future searches a clearer target: better background models, follow-up hunts for millisecond pulsars in radio and other observations, and refined machine-learning tools that can test whether the Milky Way’s central glow is truly smooth or only appears that way.

In a field where dark matter has remained frustratingly out of reach, narrowing the argument around one of the sky’s most debated signals is meaningful progress.