An AI-Built Dark Matter Registry Points to Missing Tools, Not Just Missing Theory

Dark matter reveals itself through gravity, but its underlying nature remains unknown. Proposed explanations span particles produced in the early Universe, ultralight fields, interacting dark sectors and primordial black holes. Each comes with different assumptions and predictions. Working out which observations can genuinely test an idea means following a trail through papers, calculations and software—not simply looking up a published limit.

In a new preprint, Ioana A. Zelko, Jo Bovy and Nolan Koblischke present DarkMatterFM-Registry, a computer-readable collection of 6,145 dark-matter model cards. Their central finding is about research infrastructure rather than dark matter itself: within their analysis, many obstacles to testing promising models involve missing or incomplete shared software and statistical tools, rather than missing fundamental theory. The work does not detect dark matter or establish which candidate is correct.

The researchers started with astronomy and particle-physics full-text collections covering 1994–2024. They filtered for papers containing both “dark matter” and “model,” then used a language-model abstract screen on the larger astronomy collection. After removing duplicates, 18,871 candidate papers entered extraction. A language model reading the full texts accepted 4,922 papers as introducing or substantially developing models and produced separate cards for model variants, spanning approximately 636 families.

The cards organize information that ordinary literature searches leave scattered: a model’s ingredients, parameters, production mechanism, assumptions, observable effects and relevant computational tools. They also record warnings about scientifically unsafe shortcuts. For example, two models can both suppress small-scale cosmic structure while producing differently shaped patterns of suppression. Translating both into the same standard warm-dark-matter mass can therefore give misleading comparisons.

Next, the team ranked cards using a human-written rule applied by a language model. The score measures potential value for observational follow-up, including testability, cosmological relevance, computational feasibility and robustness of the extracted information. It is not a probability that a model describes reality. The top 2,004 cards received additional information from primary and related papers, including observational constraints and routes connecting theory to measurements. Such routes can involve a prediction code, a simulation and a likelihood—a statistical tool for comparing predictions with data.

A further language-model pass assessed those computational steps as available, partial, missing or unknown. Across roughly 7,000 steps, nearly half were classified as partial and about one in twelve as missing. The resulting 5,166 capability-gap entries were dominated by likelihoods and software tools. Deterministic keyword grouping then organized gaps into 25 proposed projects. Their ranking combines the follow-up priorities of affected models, how severely a gap obstructs analysis and an estimated effort index. Those effort scores are categories, not delivery schedules.

The highest-impact proposal was a general Boltzmann-code framework for non-cold dark matter, associated with gaps affecting 556 model cards. These codes calculate how cosmic fluctuations evolve, linking microscopic physics to large-scale structure. The best estimated benefit per unit effort came from a library of model files for existing relic-abundance solvers, affecting 660 cards. Those solvers calculate how much dark matter a scenario produces; the proposed work would make more scenarios usable within tools that already exist. Reusable indirect-detection likelihoods were another leading priority.

For structure-formation measurements specifically, the largest identified gap was a non-cold-dark-matter likelihood for the Lyman-alpha forest. This forest consists of hydrogen absorption features in distant quasar spectra, which trace intervening matter. The authors identify 416 cards relying on a thermal warm-dark-matter proxy that the cards themselves flag as unsuitable. A more appropriate comparison could help researchers evaluate those scenarios without forcing them into an unreliable benchmark.

The registry’s reliability remains a substantial qualification. Structural checks passed 97.8% of cards, but passing a data-format check does not establish scientific accuracy. A separate language-model judge assessed sections from 75 sampled cards. The main text reports about 81% correct and 18% partially correct, with three incorrect verdicts and none classified as hallucinated. However, observables and constraints were judged fully correct only 46% and 50% of the time, respectively. Common problems included attributing information from related papers to the primary source and leaving stated values unrecorded. The appendix also contains inconsistent aggregate accuracy percentages that need clarification.

No expert re-grading is reported; the authors identify that as future work. Coverage is also bounded by the source collections, search filters and older literature cutoff. Project priorities depend on automated judgments, chosen weights and keyword grouping, which excludes unmatched gaps. The registry is therefore best understood as an inspectable starting point for expert checking and community planning—not a definitive census or a validated promise that building any one tool will make every associated model fully testable.