Axion‑Like Particles in the CMB Spectrum

The cosmic microwave background (CMB) is the oldest light we can observe. Its exquisite smoothness and tiny ripples encode the physics of the first 380 000 years of the Universe. Among the many “new‑physics” candidates that could leave a subtle fingerprint on this relic radiation are axion‑like particles (ALPs) – ultralight bosons that couple weakly to photons. By scrutinising the CMB’s spectrum and its anisotropies, cosmologists have turned the sky into a laboratory that can test ALP–photon interactions far beyond the reach of any terrestrial experiment.

In this pillar article we walk through why the CMB is such a powerful probe, how ALPs would modify its temperature and polarization, and what the latest measurements—from the COBE FIRAS spectrometer to the Planck satellite and upcoming missions like PIXIE—actually tell us about the allowed ALP parameter space. Along the way we draw honest parallels to the precision required in bee navigation, and we highlight how AI‑driven data analysis is becoming indispensable for both cosmology and conservation.

1. What are Axion‑Like Particles?

Axion‑like particles are a broad class of pseudo‑scalar bosons that share two defining properties:

The original QCD axion was proposed in the late 1970s to solve the strong‑CP problem. Its mass and coupling are linked by the relation \(m_a \propto 1/f_a\) (with \(f_a\) the Peccei‑Quinn symmetry‑breaking scale). ALPs, by contrast, decouple mass from coupling: string‑theory compactifications, hidden‑sector gauge groups, and many dark‑matter models predict a plenitude of such particles with independent \(m_a\) and \(g_{a\gamma\gamma}\).

Why care about ultralight ALPs?

1. Dark‑matter candidate – If the relic abundance is set by the misalignment mechanism, ALPs with \(m_a\sim10^{-22}\)–\(10^{-20}\,\text{eV}\) could make up all of the dark matter, smoothing structure on kiloparsec scales.
2. Cosmic birefringence – A background ALP field that slowly evolves can rotate the plane of linear polarisation of photons, an effect that accumulates over cosmological distances.
3. Energy‑loss channel – In hot astrophysical plasmas (e.g. supernova cores) ALPs can be produced and escape, providing an extra cooling channel that is constrained by observations of SN 1987A.

The photon coupling is the portal that lets the CMB feel the presence of ALPs. In the early Universe, large‑scale magnetic fields (primordial or generated during structure formation) enable photon–ALP mixing, leading to spectral distortions, altered anisotropy power, and polarisation rotation. The next sections unpack how these effects manifest and how we measure them.

2. The Cosmic Microwave Background: A Pristine Laboratory

The CMB is a nearly perfect black‑body with temperature

\[ T_{\rm CMB}=2.72548\pm0.00057\ {\rm K}, \]

as measured by the FIRAS instrument on the COBE satellite in the early 1990s. Its spectrum follows the Planck law to one part in 10⁵, leaving very little room for any additional energy injection after recombination.

Two complementary observables are used to search for new physics:

The µ‑distortion parameterises energy release at redshifts \(5\times10^4\lesssim z\lesssim2\times10^6\), when Compton scattering efficiently redistributed photon energies but photon‑producing processes (double‑Compton, bremsstrahlung) were too slow to restore a perfect black‑body. The y‑distortion captures heating at later times (\(z\lesssim5\times10^4\)), such as the Sunyaev‑Zel’dovich (SZ) effect from hot gas in galaxy clusters.

Because ALP–photon conversion can inject or remove photons at precisely those epochs, the CMB spectrum is a direct testbed for the ALP parameter space. Moreover, the anisotropy power spectra (especially the damping tail at high \(\ell\)) are exquisitely sensitive to any change in the photon propagation speed or to extra opacity caused by ALPs. The next sections detail the physics of the conversion and the resulting observational signatures.

3. How ALPs Interact with Photons in the Early Universe

3.1 Photon–ALP mixing in a magnetic field

In a homogeneous magnetic field \(\mathbf{B}\) the photon–ALP system can be written as a three‑component vector \((A_{\parallel},A_{\perp},a)\), where \(A_{\parallel}\) is the photon polarisation parallel to \(\mathbf{B}\). The propagation equation in the eikonal approximation reads

\[ i\frac{d}{dz} \begin{pmatrix} A_{\parallel}\\ a \end{pmatrix} = \begin{pmatrix} \Delta_{\gamma} & \Delta_{a\gamma}\\ \Delta_{a\gamma} & \Delta_{a} \end{pmatrix} \begin{pmatrix} A_{\parallel}\\ a \end{pmatrix}, \]

with

• \(\Delta_{a\gamma}= \frac{1}{2}g_{a\gamma\gamma}B_T\) (mixing term, \(B_T\) the transverse field component),
• \(\Delta_{a}= -\frac{m_a^2}{2E}\) (ALP mass term),
• \(\Delta_{\gamma}= -\frac{\omega_{\rm pl}^2}{2E}\) (plasma frequency contribution, \(\omega_{\rm pl}^2 = 4\pi\alpha n_e/m_e\)).

The mixing angle

\[ \theta = \frac{1}{2}\arctan\!\left(\frac{2\Delta_{a\gamma}}{\Delta_{\gamma}-\Delta_{a}}\right) \]

controls the conversion probability after traveling a distance \(L\):

\[ P_{\gamma\to a}= \sin^2(2\theta)\,\sin^2\!\Bigl(\frac{\Delta_{\rm osc}L}{2}\Bigr),\quad \Delta_{\rm osc}\equiv\sqrt{(\Delta_{\gamma}-\Delta_{a})^2+4\Delta_{a\gamma}^2}. \]

Two regimes are especially relevant for the CMB:

3.2 Magnetic fields in the early universe

The strength and coherence length of primordial magnetic fields are uncertain, but several observational anchors exist:

• Upper limits from CMB anisotropies: Planck 2015 constrained a stochastic field with a power‑law spectrum to be \(B_{\rm rms}\lesssim 1\) nG on Mpc scales.
• Faraday rotation of distant quasars suggests intergalactic fields of order \(10^{-9}\)–\(10^{-12}\) G.
• Simulations of structure formation predict amplified fields up to \(\sim\mu\)G in galaxy clusters, with coherence lengths of a few kpc.

For the spectral‑distortion analysis we typically adopt a conservative field of \(B_T=1\) nG coherent over a comoving Mpc, which yields a mixing term \(\Delta_{a\gamma}\approx 1.5\times10^{-28}\,{\rm eV}\,(g_{a\gamma\gamma}/10^{-12}\,{\rm GeV}^{-1})\). Even such a weak field can lead to observable effects when integrated over the Hubble horizon.

4. Spectral‑Distortion Constraints

4.1 Energy exchange from resonant conversion

During the recombination epoch (\(z\approx1100\)), the free‑electron density drops from \(n_e\sim10^{-2}\,\text{cm}^{-3}\) to \(10^{-4}\,\text{cm}^{-3}\) over a few hundred thousand years. The corresponding plasma frequency falls from \(\omega_{\rm pl}\approx 0.6\) eV to \(\approx0.06\) eV. If an ALP mass lies in this narrow window, resonant conversion can remove a fraction \(f_{\rm conv}\) of CMB photons at a specific frequency band.

The resulting µ‑type distortion is roughly

\[ \mu \simeq 1.4\,f_{\rm conv}, \]

because the energy removal is redistributed by Compton scattering into a Bose‑Einstein spectrum. FIRAS measured \(|\mu|<9\times10^{-5}\), translating into a maximum conversion fraction \(f_{\rm conv}<6\times10^{-5}\). Plugging the resonant‑conversion probability yields the bound (see e.g. Mirizzi et al. 2009)

\[ g_{a\gamma\gamma}\ \lesssim\ 2\times10^{-12}\ {\rm GeV}^{-1} \quad\text{for}\quad 10^{-14}\,\text{eV}\lesssim m_a\lesssim10^{-12}\,\text{eV}, \]

assuming a 1 nG field. The exact numbers shift modestly with the assumed field strength; a 0.1 nG field weakens the limit by a factor of ≈3.

4.2 Non‑resonant y‑distortions from late‑time conversion

At lower redshifts (\(z\lesssim5\times10^4\)), the plasma frequency is small and resonant conversion is ineffective. However, non‑resonant mixing can still convert photons into ALPs, effectively removing CMB photons and creating a y‑type distortion:

\[ y \simeq \frac{1}{4}\int\! \frac{d\ln a}{H(a)}\,\Gamma_{\gamma\to a}(a), \]

where \(\Gamma_{\gamma\to a}\) is the conversion rate per unit time. Using the COBE FIRAS limit \(|y|<1.5\times10^{-5}\) and a simple model of tangled fields with coherence length \(\lambda_B=1\) Mpc, one obtains

\[ g_{a\gamma\gamma}\ \lesssim\ 5\times10^{-12}\ {\rm GeV}^{-1} \quad\text{for}\quad m_a\lesssim10^{-14}\,\text{eV}. \]

4.3 Future prospects: PIXIE and beyond

The proposed Primordial Inflation Explorer (PIXIE) aims for a sensitivity of \(|\mu|\sim10^{-8}\) and \(|y|\sim10^{-8}\). If realised, the same calculations would tighten the coupling limit by roughly two orders of magnitude, probing

\[ g_{a\gamma\gamma}\ \lesssim\ 10^{-14}\ {\rm GeV}^{-1} \quad\text{for}\quad 10^{-14}\,\text{eV}\lesssim m_a\lesssim10^{-12}\,\text{eV}. \]

Such a reach would intersect the region where ALPs could constitute all of the dark matter via the misalignment mechanism, making CMB spectral‑distortion experiments a decisive test of that scenario.

5. Anisotropy Constraints

5.1 Damping of small‑scale temperature anisotropies

If photons convert into ALPs after recombination, the effective photon mean free path increases, leading to extra damping of the acoustic peaks at high multipoles (\(\ell\gtrsim1500\)). The Planck 2018 temperature power spectrum is measured to a few per‑cent precision at these scales. By adding a phenomenological opacity term \(\tau_{a\gamma}\) to the Boltzmann hierarchy, one can fit the data and extract an upper limit:

\[ \tau_{a\gamma} \equiv \int_{z_{\rm rec}}^{0} \frac{dz}{(1+z)H(z)}\,\Gamma_{\gamma\to a}(z) \;<\; 3\times10^{-3}\quad (95\%\,\text{C.L.}). \]

Translating this into a coupling bound (again assuming \(B_T=1\) nG) yields

\[ g_{a\gamma\gamma}\ \lesssim\ 8\times10^{-12}\ {\rm GeV}^{-1} \quad\text{for}\quad m_a\lesssim10^{-15}\,\text{eV}. \]

The shape of the damping—whether it is scale‑independent (as in a uniform opacity) or scale‑dependent (as in resonant conversion)—helps break degeneracies with other cosmological parameters such as the effective number of relativistic species \(N_{\rm eff}\).

5.2 Polarisation‑based anisotropies

ALP‑induced photon loss also reduces the E‑mode polarisation amplitude because fewer photons are available to scatter at the last‑scattering surface. The Planck EE spectrum, measured to a relative precision of \(\sim1\%\) for \(\ell\sim200\)–\(1000\), leads to a comparable bound on \(\tau_{a\gamma}\). Moreover, the TE cross‑correlation is particularly sensitive to any differential opacity between temperature and polarisation; the absence of an anomalous phase shift constrains models where ALPs preferentially affect one polarisation state.

5.3 Constraints from the CMB lensing potential

ALP conversion also modifies the CMB lensing reconstruction because the lensing estimator relies on higher‑order correlations of the temperature map. The Planck lensing power spectrum is consistent with ΛCDM within 2 % at multipoles \(\ell\lesssim400\). A modest opacity \(\tau_{a\gamma}\) would suppress the lensing amplitude \(A_{\rm lens}\); the measured value \(A_{\rm lens}=1.01\pm0.06\) thus indirectly limits \(g_{a\gamma\gamma}\) to the same level as the primary anisotropies.

6. Complementary Probes: Polarisation Rotation and Cosmic Birefringence

Beyond intensity distortions, a background ALP field that slowly evolves can rotate the linear polarisation of CMB photons by an angle

\[ \alpha = \frac{1}{2}g_{a\gamma\gamma}\,\Delta a, \]

where \(\Delta a\) is the change in the ALP field along the line of sight. This cosmic birefringence mixes E‑ and B‑modes, generating otherwise forbidden TB and EB correlations.

The latest analysis of the Planck 2018 data (including the High‑Frequency Instrument) finds a global rotation

\[ \alpha = 0.35^\circ \pm 0.14^\circ, \]

consistent with zero at the 2.5σ level (see Cosmic Birefringence). Interpreting this as an ALP signal yields

\[ g_{a\gamma\gamma}\,\Delta a \lesssim 3\times10^{-3}\ {\rm rad}. \]

If the ALP constitutes dark matter with a coherent field amplitude \(a_0\approx f_a\theta_i\) (where \(\theta_i\) is the initial misalignment angle), then for \(f_a\sim10^{12}\) GeV one obtains a coupling bound comparable to the spectral‑distortion limit: \(g_{a\gamma\gamma}\lesssim10^{-12}\) GeV\(^{-1}\). Future missions like LiteBIRD and CMB‑S4, which will measure polarisation with sub‑µK‑arcmin sensitivity, aim to push the birefringence limit down to \(|\alpha| \sim 0.01^\circ\), tightening the ALP coupling constraint by another factor of \(\sim5\).

7. From the Cosmos to the Hive: Why Precision Matters

You might wonder how a discussion about exotic particles in the early universe connects to bee conservation. The answer lies in the shared reliance on ultra‑precise measurements.

• Bee navigation depends on the detection of subtle magnetic and polarisation cues. Honeybees can resolve changes in the Earth’s magnetic field of order 10 nT, a precision comparable to the faint CMB anisotropies that cosmologists chase.
• Quantum sensors under development for CMB spectral‑distortion experiments (e.g., transition‑edge sensor bolometers) are also being adapted to monitor honey‑bee hive temperature and humidity with millikelvin accuracy, enabling early detection of disease or stress.
• AI agents trained on massive CMB maps (using convolutional neural networks and transformer architectures) are now being repurposed to analyse bee‑flight video data, distinguishing between foraging patterns and predator avoidance. The same statistical rigor that prevents a spurious ALP detection also guards against false alarms in bee‑population monitoring.

Thus, the technology and methodology that let us set a limit \(g_{a\gamma\gamma}<10^{-12}\) GeV\(^{-1}\) also empower Apiary’s mission to protect pollinators. When we invest in high‑fidelity detectors for the sky, we simultaneously gain tools to listen to the hum of a hive.

8. Future Directions: Next‑Generation CMB Experiments and AI‑Driven Analysis

8.1 Upcoming CMB facilities

These missions will close the gap between laboratory searches (e.g., the ALPS II light‑shining‑through‑wall experiment) and astrophysical probes. In particular, the combination of spectral‑distortion limits (µ, y) and polarisation rotation will test the entire ultralight ALP window from \(10^{-33}\) eV up to \(10^{-12}\) eV.

8.2 AI for component separation and anomaly detection

Separating the tiny ALP‑induced signal from foregrounds (Galactic dust, synchrotron, free‑free emission) is a classic inverse problem. Recent advances include:

• Variational autoencoders (VAEs) trained on simulated CMB + ALP spectra, which can reconstruct residuals with < 10 % bias.
• Normalising flows that learn the full joint probability distribution of multi‑frequency maps, enabling likelihood‑free inference on \(g_{a\gamma\gamma}\).
• Self‑governing AI agents (as described in Self‑Governing AI Agents) that negotiate the trade‑off between false‑positive ALP detection and over‑fitting to noise, using a game‑theoretic framework that parallels how bee colonies allocate foragers.

These techniques not only sharpen our cosmological constraints but also provide templates for real‑time monitoring of bee colonies, where AI agents must quickly distinguish between normal behavioural variability and early signs of disease.

9. Why It Matters

The CMB is a cosmic ledger that records any process capable of injecting, removing, or rotating photons after the Big Bang. Axion‑like particles, while hypothetical, are a well‑motivated extension of the Standard Model and a plausible dark‑matter constituent. By exploiting the spectral purity of the CMB and the fine‑scale structure of its anisotropies, we have already ruled out large swaths of ALP parameter space—down to couplings of order \(10^{-12}\) GeV\(^{-1}\) for ultralight masses.

Future experiments promise to tighten these bounds by two more orders of magnitude, potentially confirming or excluding the ALP dark‑matter hypothesis. The same instruments and AI tools that make these measurements possible are also being turned toward bee conservation, demonstrating that the quest to understand the Universe can directly benefit life on Earth. In the end, the precision that lets us read the faint whispers of the early cosmos also lets us hear the subtle songs of pollinators, ensuring a healthier planet for both particles and pollinators alike.