Sensitivity toward dark matter annihilation imprints on 21-cm signal with SKA-Low: A convolutional neural network approach

This study investigates the sensitivity of the radio interferometers to identify imprints of spatially inhomogeneous dark matter annihilation signatures in the 21-cm signal during the pre-reionization era. We focus on the upcoming low-mode survey of the Square Kilometre Array (SKA-Low) telescope. Using CNNs, we analyze simulated 3D 21-cm differential brightness temperature maps generated via the DM21cm code, which is based on 21cmFAST and DarkHistory, to distinguish between spatially homogeneous and inhomogeneous energy injection/deposition scenarios arising from dark matter annihilation. The inhomogeneous case accounts for local dark matter density contrasts and gas properties, such as thermal and ionization states, while the homogeneous model assumes uniform energy deposition. Our study focuses on two primary annihilation channels to electron-positron pairs ($e^+e^-$) and photons ($γγ$), exploring dark matter masses from 1 MeV to 100 MeV and a range of annihilation cross-sections. For $γγ$ channel, the distinction across dark matter models is less pronounced due to the larger mean free path of the emitted photons, resulting in a more uniform energy deposition. For $e^+e^-$ channel, the results indicate that the CNNs can effectively differentiate between the inhomogeneous and homogeneous cases. Despite observational challenges, the results demonstrate that these effects remain detectable even after incorporating noise from next-generation radio interferometers, such as the SKA. We find that the inhomogeneous dark matter annihilation models can leave measurable imprints on the 21-cm signal maps distinguishable from the homogeneous scenarios for the dark matter masses $m_{\rm DM}=1$ MeV and the annihilation cross-sections of $\geq 5 \times 10^{-30}{\rm cm^3/sec}$ ($\geq 5 \times 10^{-29}{\rm cm^3/sec}$ for $m_{\rm DM}=100$ MeV) for moderate SKA-Low noise.

💡 Research Summary

This paper investigates the capability of the upcoming Square Kilometre Array low‑frequency instrument (SKA‑Low) to detect spatially inhomogeneous signatures of dark‑matter (DM) annihilation in the redshifted 21‑cm signal from the pre‑reionization era. The authors employ a convolutional neural network (CNN) to distinguish between two classes of DM‑induced energy injection: a homogeneous model, in which the deposited energy is spatially uniform, and an inhomogeneous model, in which the energy deposition follows the local DM density contrast and the evolving gas properties.

The simulated data are generated with the publicly available DM21cm pipeline, which couples the semi‑numerical 21cmFAST code (for density, temperature, ionization, and spin‑temperature fields) with DarkHistory (for the detailed treatment of high‑energy particle cascades). Three‑dimensional differential brightness temperature cubes (ΔTb) are produced on a 64³ grid with a cell size of 48 Mpc, covering redshifts z ≲ 45. Two annihilation channels are considered: electron‑positron pairs (e⁺e⁻) and photons (γγ). Dark‑matter masses span 1 MeV to 100 MeV, and thermally averaged cross‑sections ⟨σv⟩ range from 10⁻³¹ to 10⁻²⁸ cm³ s⁻¹. For each parameter set, dozens of independent realizations are created, and realistic SKA‑Low system‑temperature noise (based on a 1000‑hour integration) is added to the maps.

The energy injection rate per unit volume is modeled as
dE/dVdt = ρ₀,DM² (1+z)⁶ (1+δ)² ⟨σv⟩/mDM,
where δ is the local matter overdensity. In the homogeneous case the (1+δ)² factor is replaced by its volume‑averaged boost, while the inhomogeneous case retains the cell‑by‑cell variation, thereby capturing the “boost factor” arising from structure formation. The e⁺e⁻ channel deposits energy locally because the produced electrons and positrons have short mean free paths; consequently, regions of high DM density experience stronger heating and ionization, producing pronounced hot and cold spots in the ΔTb field. By contrast, the γγ channel’s photons travel large distances before depositing energy, smoothing out spatial variations and making the signal more uniform.

The CNN architecture consists of three 3‑D convolutional layers (32, 64, and 128 filters respectively), each followed by batch normalization, ReLU activation, and max‑pooling, then two fully‑connected layers (256 and 128 neurons) and a sigmoid output that classifies a map as “inhomogeneous” or “homogeneous”. The network is trained with Adam optimizer (learning rate = 1e‑4) on 70 % of the data, validated on 15 %, and tested on the remaining 15 %. Performance metrics include accuracy, precision, recall, and the area under the ROC curve (AUC).

Results show that for the e⁺e⁻ channel the CNN achieves an AUC of ≈0.96 and classification accuracy above 90 % when the annihilation cross‑section exceeds ≈5 × 10⁻³⁰ cm³ s⁻¹ for a 1 MeV DM particle (or ≈5 × 10⁻²⁹ cm³ s⁻¹ for a 100 MeV particle), even after adding realistic SKA‑Low noise. The γγ channel yields an AUC near 0.58, indicating that the network cannot reliably separate the two models because the energy deposition is nearly homogeneous. Traditional power‑spectrum analyses fail to distinguish the models in both channels, underscoring the advantage of deep‑learning methods that can capture non‑linear spatial patterns.

The authors conclude that (i) spatially inhomogeneous DM annihilation leaves detectable imprints on the 21‑cm brightness‑temperature field, (ii) SKA‑Low’s sensitivity is sufficient to observe these imprints for a realistic range of DM masses and cross‑sections, and (iii) convolutional neural networks provide a powerful, model‑independent tool for extracting subtle astrophysical signatures from high‑dimensional data. They suggest future work to explore additional annihilation channels (e.g., μ⁺μ⁻, π⁰), larger parameter spaces, and the application of the method to actual SKA‑Low data pipelines.