Isotropic diffuse gamma-ray flux limits from Carpet-3 experiment data using neural networks with domain adaptation
Diffuse gamma rays above 100 TeV probe PeV cosmic rays and astrophysical neutrino sources. Their strong attenuation on background radiation makes the observed flux sensitive to the location of the emission and to possible modifications of photon propagation. Searches for an isotropic component require rejection of an overwhelming hadronic background, but uncertainties in air-shower physics, cosmic-ray composition and detector response can lead to differences between simulated and measured events. Here we report upper limits on the isotropic diffuse gamma-ray flux from 100 TeV to 10 PeV using 1078 days of Carpet-3 observations collected between 2022 and 2026. We identify photon candidates with a neural network combining the surface-array signals and measurements from the upgraded 410 $\mathrm{m^2}$ muon detector. Unsupervised domain adaptation uses unlabeled experimental events during training to mitigate the impact of simulation--data discrepancies on the classification. The limits are among the strongest available and provide benchmarks for tests of extended Galactic emission and modified photon propagation from extragalactic sources.