The question of how dark matter is distributed in the Galactic halo is far more than an exercise in curve-fitting. The shape of the halo determines the local dark matter density, which in turn sets the expected signal rate for direct detection experiments searching for weakly interacting massive particles deep underground. It also governs how astronomers model the orbits of stellar streams, the dynamics of satellite galaxies, and the total mass of the Milky Way out to its farthest reaches. For decades, simulations of cold dark matter have suggested that halos should follow the NFW profile, whose density diverges as a power law toward the galactic center, producing a so-called “cuspy” core. Yet observations of many galaxies, particularly dwarf spirals, have long hinted that their dark matter is more centrally diffuse, or “cored,” a tension that ranks among the most persistent small-scale challenges to the standard cosmological paradigm.

Singh and Desai approached the problem with the full machinery of modern Bayesian statistics. Rather than simply asking which model fits the data best in a least-squares sense, they computed the Bayesian evidence for each model, a quantity that balances goodness of fit against model complexity and naturally penalizes models that add parameters without earning them. The computations were performed using DYNESTY, a dynamic nested sampling package that efficiently estimates both posterior distributions and evidences for high-dimensional models. The team combined three different baryonic mass models, describing the visible matter in the Galactic bulge, disk, and gas, with seven dark matter halo profiles, including the NFW profile, the Einasto profile, the Burkert profile, the Plummer profile, and several cored variants such as the pseudo-isothermal halo and the core-modified profile of Lazar and collaborators.

On the data side, the authors exploited four recent rotation curve compilations built on Gaia Data Release 3, the third and most precise catalog of positions, distances, and proper motions for nearly two billion stars delivered by the Gaia spacecraft. These include the circular velocity curve of the Milky Way derived from luminous red giant branch stars by Zhou and colleagues, the kinematic mapping of the Galactic disk to roughly 30 kiloparsecs by Wang and collaborators, and related datasets that trace the Galaxy’s rotation from about 5 to 25 kiloparsecs from the Galactic center. Each dataset encodes the same fundamental observable: the orbital speed of stars and gas as a function of distance from the Galactic center. Because the visible matter alone cannot explain why orbital speeds remain roughly flat, or in some recent analyses even decline in Keplerian fashion, at large radii, the shape of the rotation curve at those distances is a direct probe of the invisible halo.

The results carry implications that ripple across dark matter physics. In nearly every dataset-baryon combination the authors tested, the Einasto profile emerged as the preferred phenomenological description, beating the NFW profile by decisive margins in the Bayesian evidence. The Einasto profile, first introduced by Jaan Einasto in 1965 in a completely different context as a model for galactic light distributions, describes the halo density as a slowly varying power of the radius, falling more gradually near the center than NFW’s sharp cusp. Its victory here suggests that the Milky Way’s inner halo is smoother and less centrally concentrated than the classic cold dark matter prediction, at least as probed by Gaia-era kinematics.

Equally significant was the performance of the cored profiles. Models in which the dark matter density flattens to a finite central value, such as the Burkert profile originally proposed to explain dwarf galaxy halos, were systematically preferred over the cuspy NFW model. This finding aligns the Milky Way with a broader pattern seen in external galaxies and lends indirect support to scenarios in which baryonic feedback, the energetic outflows from star formation that push gas and dark matter around, or even self-interacting dark matter, reshapes the inner halo. The authors caution, however, that their conclusions apply within the adopted modeling framework and the specific Gaia-based rotation curve datasets used; the comparison is phenomenological rather than a falsification of any particular particle physics model.

The study also delivered a pointed verdict on modified gravity. Modified Newtonian Dynamics, or MOND, proposed by Mordehai Milgrom in 1983, dispenses with particle dark matter altogether by altering the law of gravity at accelerations below a characteristic scale of roughly 10^-10 meters per second squared. Singh and Desai tested MOND using three different interpolating functions, the mathematical bridges that connect the Newtonian and deep-MOND regimes, coupled to the same baryonic models. Within the implementations they considered, all three MOND variants provided poorer fits to the Gaia-based Milky Way rotation curve than the best dark matter profiles, and the Bayesian evidence firmly ranked them below the Einasto and cored halo models. While the authors are careful to note that this does not exhaust every possible MOND formulation, particularly fully relativistic extensions, it adds the Milky Way’s own kinematics to the list of arenas where modified gravity struggles to compete on equal statistical footing.

One perhaps counterintuitive outcome of the analysis is what did not matter: the baryonic models. The team’s three descriptions of the Galaxy’s luminous content, differing in how they treat the stellar disk, thick disk, and gas distribution, produced no decisive winner. None of the baryonic models was consistently favored over the others across the dark matter profiles and datasets. This suggests that, at the level of precision currently offered by the rotation curve data, systematic uncertainties in the baryonic component are not the dominant driver of model preference; the dark matter profile itself is doing the discriminating work. It also underscores how much room remains for improvement, since future data with tighter errors at large Galactic radii could well sharpen the sensitivity to the visible mass distribution.

The timing of this work is notable. Recent analyses of Gaia DR3 data, including the detection of a Keplerian decline in the Milky Way rotation curve beyond the solar circle by Jiao and collaborators, have reignited debate about the Galaxy’s total mass and halo structure, with some authors arguing that a lighter, faster-declining Milky Way carries cosmological consequences. Assessing the robustness of rotation curves inferred through the Jeans equations against cosmological simulations, as Koop and colleagues have done, remains an active concern, since systematics in distance estimates and stellar selection can bias the inferred velocities. By folding multiple independent rotation curve compilations into a single Bayesian comparison framework, Singh and Desai have provided a way to see which conclusions survive the choice of dataset, a robustness check that single-dataset analyses cannot offer.

The broader stakes extend beyond Galactic astronomy. Dark matter remains one of the most profound unsolved problems in physics, evidenced across galaxy rotation curves, gravitational lensing, and the cosmic microwave background, yet still escaping direct detection after decades of increasingly sensitive experiments. Pinning down the precise shape of the Milky Way’s halo matters for interpreting those experiments, because the local density and velocity distribution of dark matter set the expected event rates in detectors. A smoother, cored, or Einasto-like halo changes the translation between detector limits and particle properties. As Gaia continues to refine our map of the Galaxy and future surveys extend rotation measurements to ever larger radii, the Bayesian machinery demonstrated in this study offers a template: rather than assuming a halo model, let the data adjudicate among them. For now, the Milky Way’s dark matter appears to prefer the gentle mathematics of Einasto over the sharp cusp of NFW, and to leave little statistical room for gravity itself to do the dark matter’s work.

As part of their collaboration with NASA, CAII will play a pivotal role in developing an open-source deep learning framework aimed at processing images captured by the Euclid spacecraft. This undertaking is spearheaded by Principal Investigator Xin Liu, who emphasizes the critical nature of this research. A primary hurdle in the analysis of data generated by Euclid is the phenomenon of blended galaxies. Overlapping images can obscure distinct sources of galactic information, making it challenging to derive accurate astrophysical measurements. The blending of galaxies is particularly problematic in areas such as photometry, redshift estimation, and galaxy morphology, as it introduces biases that can lead to significant inaccuracies in scientific interpretations.

Deep learning, specifically through a tool known as Detection, Instance Segmentation and Classification with Deep Learning (DeepDISC), is revolutionizing the approach to identifying celestial objects. This innovative tool harnesses machine learning techniques to enable a more accurate detection of stars and galaxies. Liu and his team will leverage DeepDISC within the framework of the Euclid mission to not only enhance the accuracy of galaxy identification but also to quantify the uncertainties associated with their predictions. This dual approach promises to significantly improve the reliability of data analysis, ensuring that researchers can draw trusted conclusions from their findings.

Liu articulates the importance of addressing blended sources in Euclid’s data analysis, citing that overcoming this challenge is essential for the integrity of the mission’s scientific outputs. The implications of this research extend beyond Euclid itself; the techniques developed can be adapted for other ambitious space exploration projects, most notably the Vera C. Rubin Observatory. With its anticipated first light occurring later this year, the observatory stands to benefit from advancements in deblending images both from ground-based telescopes and those deployed in space.

The research team is bolstered by a collaborative spirit, with critical contributions from co-principal investigators such as Vlad Kindratenko, Director of CAII, along with Astronomy Professor Yue Shen and Computer Science Professor Yuxiong Wang. Their distinct expertise in computational infrastructure, data analysis methodologies, and machine learning act as a strong foundation for this interdisciplinary project. Each member brings a wealth of knowledge that enhances the scientific rigor and innovative spirit of the mission.

Wang emphasizes that the advancements in computer vision and artificial intelligence are leading to foundational models that redefine our understanding of visual information through natural images. The current phase in AI research represents a thrilling opportunity to adapt these powerful models toward unraveling the mysteries of the universe. In an age where AI is becoming increasingly capable, its integration into fields like astrophysics showcases the potential for transformative discoveries and enhances our collective understanding of fundamental cosmic principles.

The CAII’s initiative reflects broader trends in interdisciplinary collaboration, showcasing how artificial intelligence can intersect with scientific inquiry in profound ways. As AI methodologies evolve and improve, their application to space exploration promises unprecedented advancements in our quest for knowledge. The focus on deep learning and its ability to sort through complex datasets will be vital in maximizing the scientific return of missions like Euclid, ensuring that the wealth of information collected translates into significant scientific insights.

Moreover, the interdisciplinary nature of the project heralds a new era of scientific research where collaboration across different fields is not just beneficial but necessary. The integration of knowledge from AI, computer science, and astrophysics underscores the importance of blending disciplines to address the multifaceted challenges presented by complex datasets in space exploration.

As the Euclid mission prepares to launch, the work being done by CAII serves as a beacon of innovation, particularly in the realm of AI applications in astrophysics. The developments achieved through deep learning frameworks have the potential to redefine how scientists engage with astronomical data, paving the way for future missions and research endeavors in understanding the dark universe.

In conclusion, the collaboration between CAII and NASA represents an exciting juncture in the intersection of artificial intelligence and astrophysics. Through substantial investment, innovative methodologies, and a commitment to scientific excellence, this initiative will not only enhance the accuracy of data analysis in the Euclid mission but also contribute to the broader dialogue about dark matter and dark energy, ultimately aiming to unlock new dimensions of understanding within the universe’s vast expanse.

Subject of Research: Analysis of dark matter and dark energy using AI in the Euclid mission
Article Title: CAII Harnesses AI for the Euclid Mission to Explore the Dark Universe
News Publication Date: [Insert publication date]
Web References: [Provide relevant URLs]
References: [List citations if necessary]
Image Credits: [Provide credit if applicable]

Keywords

AI, dark matter, dark energy, Euclid mission, deep learning, NCSA, CAII, astrophysics, Vera C. Rubin Observatory, machine learning, blended galaxies, data analysis.