Cosmic Smoothie: Hunting Anomalies Without a Recipe ✛ ✦

Original: "Weakly supervised machine learning for model-agnostic searches of new phenomena in the $$\gamma$$-ray sky"

· Michael Krämer, Silvia Manconi, Kathrin Nippel

arXiv:2607.07158 · 2026-07-08 · CC BY 4.0 · 1 min

High Energy Instrumentation HEP Phenomenology

Scientists train a computer to act like a taste tester, spotting cosmic anomalies from barely noticeable differences — no reference sample needed.

The Fermi gamma-ray telescope has mapped thousands of powerful light bursts. Many come from familiar objects: pulsars (stellar lighthouses, and active galactic nuclei) glows around black holes. But over two thousand sources remain unidentified. Some might be signals from dark matter — the invisible scaffolding of galaxies. The search is inspired by the work of Vera Rubin, Jocelyn Bell Burnell, and Roberto Peccei.

The method is like a taste tester with two smoothies. One is the pure flavor of a known banana (radiation from familiar objects). The other is the same banana with a drop of unknown syrup. Machine learning compares them and finds differences without a 'recipe' for the additive. The algorithm is fed real brightness measurements and light decomposition from the telescope, as well as computer models of spectra.

Without any hints, the method caught 82% of anomalies — almost as well as with full training.

This screening isn't a replacement for detailed analysis, but a flexible way to pick candidates. It's especially vital when the signal is unknown — be it dark matter or new particles. The plan is to sift through the entire Fermi catalog so no anomaly slips away.

Whole graph →

Over two thousand gamma-ray sources in the Fermi catalog remain unidentified — more than the known pulsars or active galactic nuclei. A hidden universe awaits a discoverer.

Like in Blake Crouch's novel 'Dark Matter,' the hunt for invisible forces that alter the radiation of distant objects feels like a detective story about hidden reality.

E_c \simeq 2.5\,\text{ГэВ}\,\left(\frac{|m_a^2 - \omega_{\rm pl}^2|}{1\,{\rm нэВ}}\right)\left(\frac{B_\perp}{\mu\text{Г}}\right)^{-1}\left(\frac{g_{a\gamma}}{10^{-11}\,\text{ГэВ}^{-1}}\right)^{-1}

Above this energy, oscillations become energy-independent, and near it characteristic 'wriggles' appear in the spectrum.

l_{\rm osc} = 32\,{\rm кпк}\,\sqrt{1+(E_c/E_\gamma)^2}\left(\frac{B_\perp}{\mu\text{Г}}\right)^{-1}\left(\frac{g_{a\gamma}}{10^{-11}\,\text{ГэВ}^{-1}}\right)^{-1}

Characteristic scale on which the conversion probability periodically changes, leading to modulation of the observed flux.

Scientists

Andrey Markov

Concepts

Weakly Supervised Learning self-supervised learning fermi telescope weak lensing ai astrophysics anomaly detection gamma ray spectroscopy classifier weak measurement training data symbolic regression labeled data spectral analysis ai false positive event selection astrophysical signal detection gamma ray telescope sky survey signal averaging noise reduction spectroscopy

I am the author

Cosmic Smoothie: Hunting Anomalies Without a Recipe ✛ ✦