Cavity Multimodes as an Array for High-Frequency Gravitational Waves Diego Blas, Yifan Chen, Yuxin Liu, Yanfei Shang, Jing Shu Abstract Microwave cavities operated in the presence of a background magnetic field provide a promising avenue for detecting high-frequency gravitational waves (HFGWs). We demonstrate for the first time that the distinct antenna patterns of multiple electromagnetic modes within a single cavity enable localization and reconstruction of key properties of an incoming HFGW signal, including its polarization ratio and frequency drift rate. Using a 9-cell cavity commonly employed in particle accelerators as a representative example, we analyze the time-domain response of 18 nearly degenerate modes, which can be sequentially excited by a frequency-drifting signal. The sensitivity is further enhanced by the number of available modes, in close analogy to the scaling achieved by a network of independent detectors, enabling sensitivity to astrophysically plausible binary sources. I Introduction Gravitational wave (GW) observations across multiple frequency bands, from kHz signals of stellar-mass binaries detected by ground-based interferometers Abbott et al. 2016 to nHz backgrounds hinted by through pulsar timing arrays (PTAs), have opened new windows into the Universe. At frequencies far above the kHz range, no strong astrophysical sources are known, making the high-frequency GW (HFGW) band a promising arena for exploring new physics such as primordial black holes (PBHs) or bosonic clumps Aggarwal et al. 2021 ; Aggarwal et al. 2025 . Detection in this regime remains largely uncharted. Several detector concepts first developed for axion searches, particularly electromagnetic (EM) resonant systems such as microwave and superconducting cavities, have recently been adapted for HFGW detection Cruise 2000 ; Bernard et al. 2002 ; Li et al. 2003 ; Ballantini et al. 2003 ; Ballantini et al. 2005 ; Goryachev and Tobar 2014 ; Herman et al. 2021 ; Berlin et al. 2022a ; Domcke et al. 2022 ; Berlin et al. 2022b ; Tobar et al. 2022 ; Ahn et al. 2024 ; Alesini et al. 2023 ; Berlin et al. 2023 ; Bringmann et al. 2023 ; Chen et al. 2025a ; Domcke et al. 2024 ; Gao et al. 2024a ; Gao et al. 2024b ; Navarro et al. 2024 ; Schmieden and Schott 2024 ; Domcke et al. 2025a ; Carney et al. 2025 ; Domcke et al. 2025b ; Schneemann et al. 2024 ; Valero et al. 2025 ; Capdevilla et al. 2025a ; Fischer et al. 2025 ; García-Cely et al. 2025 ; Campbell et al. 2025 ; Pappas et al. 2025 ; Capdevilla et al. 2025b ; Kharzeev et al. 2025 ; Kim et al. 2025 ; Liang et al. 2025 ; Takai 2025 ; Schenk et al. 2025 ; Marconato et al. . As in laser interferometers and PTA observations, extracting the properties of a detected signal is crucial for confirming its quadrupolar GW nature Chen et al. 2025b and identifying its origin. Although cosmological HFGWs are strongly constrained by Big Bang nucleosynthesis (BBN) bounds Aghanim et al. 2020 , nearby coherent or transient sources, such as PBH binaries, could still produce detectable signals Aggarwal et al. 2021 ; Aggarwal et al. 2025 . For such sources, resolving the propagation direction, polarization pattern, and waveform is essential. Similar to ultralight dark matter searches Foster et al. 2021 ; Chen et al. 2022a ; Jiang et al. 2024 ; Sulai et al. 2023 ; Zeng et al. 2025 ; Jiang et al. 2025 ; Gavilan-Martin et al. 2025 ; Reina-Valero et al. 2025 ; Arza et al. 2025 ; Wilson et al. 2025 , a network of detectors could in principle provide this information Schmieden and Schott 2024 ; Schneemann et al. 2024 , but practical challenges such as frequency overlap and phase synchronization remain, particularly for transient events. In this work, we demonstrate the use of multiple EM modes within a single cavity as an effective detector array for HFGWs, focusing on the nearly degenerate modes of a multi-cell cavity. Such multi-cell geometries, widely used in accelerator technology to improve field uniformity and energy transfer efficiency, naturally support a set of near-degenerate resonant modes Aune et al. 2000 ; Belomestnykh and Shemelin 2006 ; Jeong et al. 2020 . An inspiraling primordial black hole binary can sequentially excite these modes as its GW frequency increases Franciolini et al. 2022 ; Barrau et al. 2025a ; Gatti et al. 2024 ; Profumo et al. 2025 . By comparing the relative signal strengths and phases among the modes, which requires at least \displaystyle 5 loud modes, we show for the first time that a single cavity can localize and reconstruct key features of the GW waveform. II Cavity Multimodes and Antenna Patterns via the Inverse Gertsenshtein Effect HFGWs in the MHz-to-GHz regime, for which no strong astrophysical sources, represent an unexplored frontier in fundamental physics Aggarwal et al. 2021 ; Aggarwal et al. 2025 (see Casalderrey-Solana et al. 2022 for possible MHz signals in mergers of neutron stars). Given current sensitivities Aggarwal et al. 2021 ; Aggarwal et al. 2025 ; Tito D’Agnolo and Ellis 2025 ; Guo et al. 2025 , the detection of a cosmological background at these frequencies is strongly constrained by the BBN bound on the effective number of relativistic degrees of freedom Aghanim et al. 2020 . Nevertheless, several possible local sources within the Galaxy, such as PBH binaries or bosonic clumps, may generate detectable signals that may be detectable by upcoming precision experiments Aggarwal et al. 2021 ; Aggarwal et al. 2025 . Among the various proposed detection approaches, EM resonant systems, such as microwave and superconducting cavities, have attracted growing interest due to their conceptual similarity to axion dark matter searches. In the presence of a background EM field, a passing GW induces an effective oscillating current through the inverse Gertsenshtein effect Gertsenshtein and Pustovoit 1962 , expressed as Berlin et al. 2022a j^{\mu}_{\mathrm{eff}}=\partial_{\nu}(\frac{1}{2}\,hF_{0}^{\mu\nu}+h^{\nu}_{~\sigma}F_{0}^{\sigma\mu}-h^{\mu}_{~\sigma}F_{0}^{\sigma\nu}), (II.1) where \displaystyle F_{0}^{\mu\nu} is the background EM field-strength tensor and \displaystyle h^{\mu\nu} denotes the GW perturbation in the transverse–traceless (TT) 1 1 1 We thank Jordan Gué and Tom Krokotsch for helpful discussions on the use of TT gauge for our signals. gauge Ratzinger et al. 2024 ; Kim et al. 2025 ; Schenk et al. 2025 . The trace \displaystyle h=\eta_{\mu\nu}h^{\mu\nu} is defined with respect to the flat-spacetime metric \displaystyle\eta_{\mu\nu} . The total electric field inside the cavity can be expanded in normalized eigenmodes, \displaystyle\vec{E}(t,\vec{x})=\Sigma_{a}e_{a}(t)\,\vec{E}_{a}(\vec{x}) , where each \displaystyle\vec{E}_{a}(\vec{x}) has an eigenfrequency \displaystyle\omega_{a} . The equation of motion for each mode’s amplitude \displaystyle e_{a}(t) is then Tang et al. 2024 ; Zeng et al. 2025 \ddot{e}_{a}(t)+\frac{\omega_{a}}{Q^{L}_{a}}\dot{e}_{a}(t)+\omega_{a}^{2}e_{a}(t)=-\int\partial_{t}\vec{j}_{\mathrm{eff}}\cdot\vec{E}^{*}_{a}\,\mathrm{d}V, (II.2) where \displaystyle Q_{a}^{L} is the loaded quality factor of the \displaystyle a -th mode, and the integral is evaluated over the cavity volume \displaystyle V . For simplicity, we consider a static magnetic background field of uniform strength \displaystyle B_{0} , and note that the discussion can be readily generalized to an oscillating EM background, such as a secondary cavity-mode enabling heterodyne conversion to probe frequencies far below the cavity resonance through either EM or mechanical coupling Berlin et al. 2023 ; Goryachev et al. 2019 ; Berlin et al. 2020 ; Berlin et al. 2021 ; Thomson et al. 2021 ; Berlin et al. 2022b ; Thomson et al. 2023 ; Li et al. 2025 ; Fischer et al. 2025 ; Marconato et al. . We take the incoming GW to have momentum \displaystyle\vec{k} and expand it as \displaystyle h^{ij}(\vec{k})=(h_{+}\,\epsilon_{+}^{ij}(\hat{k})+h_{\times}\,\epsilon_{\times}^{ij}(\hat{k}))\,\exp[{-\mathrm{i}\vec{k}\cdot\vec{x}}] , in terms of the two polarization tensors \displaystyle\epsilon^{ij}_{+/\times}(\hat{k}) and the associated mode amplitudes \displaystyle h_{+/\times} . The effective current can then be written as \vec{j}_{\mathrm{eff}}(\vec{k}\,)\equiv B_{0}|\vec{k}|^{2}V^{1/3}\left(h_{+}\,\hat{j}_{+}+h_{\times}\,\hat{j}_{\times}\right), (II.3) where \displaystyle\hat{j}_{+/\times} are dimensionless spatial profiles Berlin et al. 2022a . The induced signal is proportional to the overlap between the effective current and the cavity electric-field distribution, characterized by the dimensionless overlap function \eta^{a}_{+/\times}(\vec{k}\,)\equiv\frac{1}{{V^{1/2}}}\int_{V}\hat{j}_{+/\times}\cdot\vec{E}_{a}^{*}\,\mathrm{d}V, (II.4) which depends on both the GW polarization and momentum \displaystyle\vec{k} . In practice, the measured signal represents a scalar projection of the tensorial GW field, combining the two polarization components. Recovering the full strain information therefore requires multiple independent mode responses, achievable either through a network of cavities or by exploiting multimode structures within a single cavity. A resonant cavity supports a discrete set of electromagnetic modes characterized by mode indices, with frequency spacing determined by its geometry Hill 2009 ; Navarro et al. 2024 ; Navarro-Madrid et al. 2025 . A particularly important configuration is the multi-cell cavity, which consists of a chain of resonant cells coupled through irises. Analogous to a system of coupled oscillators, each single-cell eigenmode splits into \displaystyle N collective modes in an \displaystyle N -cell structure, whose eigenfrequencies remain nearly degenerate and lie close to that of the corresponding single-cell cavity mode Aune et al. 2000 ; Belomestnykh and Shemelin 2006 . These modes, labeled by \displaystyle a=1,\ldots,N , have electric-field profiles across the cells (with the cell number labeled by \displaystyle j ) and relative amplitudes proportional to \displaystyle\sin[(j-1/2)a\pi/N] , where a negative value corresponds to the opposite field direction. As \displaystyle a increases, the field pattern evolves from all cells oscillating with the same orientation ( \displaystyle a=1 ) to neighboring cells oscillating with opposite orientations ( \displaystyle a=N ). The corresponding eigenfrequencies increase sequentially and approximately follow a cosine-shaped dispersion relation, with the frequency separation between adjacent modes set by the coupling strength through the irises Belomestnykh and Shemelin 2006 . We take the well-established 9-cell elliptical TESLA cavity as a representative example Aune et al. 2000 ; Belomestnykh and Shemelin 2006 . This structure is widely used in superconducting accelerators, operating in the fundamental \displaystyle\pi -mode ( \displaystyle a=9 ) at \displaystyle 1.3 GHz. The cavity’s internal geometry corresponds to an effective volume of about \displaystyle 25 L and can achieve intrinsic quality factors up to \displaystyle Q_{0}\sim 10^{11} when fabricated from superconducting Niobium. The modes of the axisymmetric 9-cell cavity can be described by extending the single-cell eigenmode basis into 9 coupled copies. Figure 1 : Electric-field configurations of the first three modes ( \displaystyle a=1,2,3 ) of the \displaystyle\mathrm{TE}_{111+} family in a 9-cell elliptical TESLA cavity. The cavity symmetry axis is aligned with the \displaystyle z axis. The left three panels show the electric fields on the \displaystyle y=0 plane and do not have perpendicular components. The right three panels show the fields on the \displaystyle z=0 plane in the center of the \displaystyle 5 -th cell, without perpendicular components. In each panel, arrow lengths are proportional to the normalized electric-field amplitude, normalized by maximum of each mode \displaystyle|\vec{E}_{a}^{\rm max}| . The corresponding \displaystyle\mathrm{TE}_{111-} modes are obtained by a \displaystyle\pi/2 rotation about the cavity axis. Figure 2 : Overlap functions \displaystyle\eta_{+,\times}^{a} for GWs incident from direction \displaystyle\hat{k}\equiv(\phi,\theta) , shown for the \displaystyle\mathrm{TE}_{111+} modes of the 9-cell cavity. For each mode, the GW frequency is taken to match the corresponding cavity resonance \displaystyle\omega_{a} . The plots are shown for \displaystyle\phi,\theta\in(0,\pi/2) ; overlap functions in other regions of the sky, as well as those for the \displaystyle\mathrm{TE}_{111-} modes, can be obtained via the symmetry relations discussed in Supplemental Material. We consider the fundamental transverse electric mode TE 111 , characterized by the indices \displaystyle m=1 (azimuthal), \displaystyle n=1 (radial), and \displaystyle p=1 (axial), with mode frequencies ranging from about \displaystyle 1.51 GHz at \displaystyle a=1 to about \displaystyle 1.76 GHz at \displaystyle a=9 . The \displaystyle m=1 dependence introduces a twofold degeneracy, with the electric-field component \displaystyle E_{\phi} proportional to \displaystyle\cos\phi and \displaystyle\sin\phi in the azimuthal angle around the symmetry axis ( \displaystyle z -axis), denoted \displaystyle\mathrm{TE}_{111+} and \displaystyle\mathrm{TE}_{111-} , respectively. A small frequency splitting arises from the slight breaking of axial symmetry in realistic cavity geometries, yielding a total of \displaystyle 18 modes when combined with the 9 collective modes. Figure 1 shows the electric-field distribution of the TE 111+ mode projected onto the \displaystyle y=0 plane ( \displaystyle\phi=\pi/2 ) and the \displaystyle z=0 mid-plane of the fifth cell, with color indicating field magnitude and arrows showing field direction for the lowest collective modes ( \displaystyle a=1,2,3 ). The \displaystyle\mathrm{TE}_{111-} mode corresponds to the same field pattern rotated by \displaystyle\pi/2 in \displaystyle\phi . In Supplemental Material, we demonstrate that two antenna couplers can simultaneously read out all \displaystyle 18 modes with comparable coupling strengths. In Fig. 2 , we show the overlap function \displaystyle\eta^{a}_{+,\times} for different GW propagation directions \displaystyle\hat{k}\equiv(\phi,\theta) , also referred to as the antenna pattern, assuming a uniform magnetic field applied along the \displaystyle+z direction and approximating the GW frequency by the mode’s resonant frequency. Each mode exhibits a strong directional preference in \displaystyle\theta , peaking at \displaystyle|\vec{k}|L_{z}\cos\theta\approx a\pi/9 , where \displaystyle L_{z}\sim 0.1 m is the separation between adjacent cell centers. This condition corresponds to incoming GW oscillations that are phase-matched to the modulation of the electric field across the cavity cells. We display \displaystyle\hat{k} only over one-eighth of the sky, with azimuthal angle \displaystyle\phi\in[0,\pi/2] and polar angle \displaystyle\theta\in[0,\pi/2] . The overlap functions in other regions can be obtained directly from this portion by applying the transformations \displaystyle\phi\rightarrow\pi\pm\phi and \displaystyle\theta\rightarrow\pi-\theta , under which both \displaystyle\vec{E}_{a} and \displaystyle\vec{j}_{\mathrm{eff}} undergo only a sign change of \displaystyle\pm 1 , determined by their parity under the transformation, as detailed in Supplemental Material. The eightfold degeneracy can be lifted by incorporating phase information or by applying external magnetic fields in transverse directions. III Dissecting Gravitational Wave Strain from Multi-Cell Responses For multi-cell cavities supporting several nearly degenerate modes with distinct antenna patterns, the relative amplitudes and phases of the signals among these modes can be used to reconstruct the full waveform of an incident GW. As a demonstration, we consider a waveform inspired by an inspiraling PBH binary, whose GW frequency increases linearly with time \displaystyle t , thereby sequentially exciting all \displaystyle 18 cavity modes. The waveform can be approximated as Maggiore 2018 \displaystyle h_{+} \displaystyle=\frac{h_{0}}{\sqrt{1+\kappa^{2}}}\exp\left[\mathrm{i}\left(\delta_{0}+\omega_{1}(t-t_{0})+\alpha(t-t_{0})^{2}/2\right)\right], (III.1) \displaystyle h_{\times} \displaystyle=\frac{h_{0}\kappa}{\sqrt{1+\kappa^{2}}}\exp\left[\mathrm{i}\left(\delta_{0}+\xi+\omega_{1}(t-t_{0})+\alpha(t-t_{0})^{2}/2\right)\right], where \displaystyle h_{0} denotes the overall strain amplitude, \displaystyle\kappa\equiv|h_{\times}/h_{+}| the polarization amplitude ratio, \displaystyle\xi the relative phase between the two polarizations, and \displaystyle t_{0} the reference time at which the GW frequency matches the resonant frequency \displaystyle\omega_{1} of the fundamental mode, with initial phase \displaystyle\delta_{0} . The parameter \displaystyle\alpha represents the linear frequency chirp rate that characterizes the inspiral evolution. More complex time dependences, such as merger and ringdown phases, can be incorporated within the same parameterized framework while retaining the same set of degrees of freedom. General solutions to the cavity-mode equation in Eq. ( II.2 ) can be written as e_{a}(t)=-\mathrm{i}\omega_{a}^{3}B_{0}V^{5/6}\int_{-\infty}^{t}G_{a}(t-\tau)\left(\sum_{A=+,\times}h_{A}\,\eta_{A}^{a}\right)\text{d}\tau, (III.2) where the Green function of mode \displaystyle a is defined as G_{a}(t)=\exp[{-\gamma_{a}t}]\frac{\sin\left(\sqrt{\omega_{a}^{2}-\gamma_{a}^{2}}t\right)}{\sqrt{\omega_{a}^{2}-\gamma_{a}^{2}}}H(t), (III.3) with \displaystyle\gamma_{a}\equiv\omega_{a}/(2Q_{a}^{L}) denoting the dissipation rate and \displaystyle H(t) the Heaviside step function. For the specific waveform in Eq. ( III.1 ), Eq. ( III.2 ) yields an analytic solution for \displaystyle e_{a}(t) , provided in Supplemental Material. Figure 3 : Time evolution of the loudest cavity mode, \displaystyle a=3 ( \displaystyle\mathrm{TE}_{111+} ), characterized by the mode amplitude \displaystyle e_{3}(t) , for the benchmark GW waveform. The time axis is shifted by \displaystyle t-t_{3} , where \displaystyle t_{3} denotes the moment when the GW frequency \displaystyle\omega_{g}=|\vec{k}| (indicated on the upper axis) crosses the cavity resonance \displaystyle\omega_{3} . We adopt the following benchmark parameters for the GW strain: \displaystyle(\phi,\theta)=(\pi/3,5\pi/12) , \displaystyle h_{0}=9.75\times 10^{-19} , \displaystyle\kappa=1 , \displaystyle\xi=-\pi/2 , \displaystyle\delta_{0}=0 , \displaystyle t_{0}=57.12~$\displaystyle\mathrm{\SIUnitSymbolMicro s}$ , and \displaystyle\alpha=1.20\times 10^{14}~\mathrm{rad/s^{2}} . These values correspond to a PBH binary with chirp mass \displaystyle\mathcal{M}_{c}=0.96\times 10^{-3}~M_{\oplus} ( \displaystyle M_{\oplus} denotes the Earth mass) in a circular orbit with orbital radius \displaystyle a_{\rm orb}=1.6\times 10^{-3}~\mathrm{m} , located at a distance \displaystyle r=1.9\times 10^{-1}~\mathrm{AU} , with both the orbital inclination angle and the polarization angle set to zero relative to the cavity frame. Figure 3 shows the temporal evolution of \displaystyle e_{a}(t) for the loudest mode, the \displaystyle a=3 , \displaystyle\mathrm{TE}_{111+} mode, with a loaded quality factor \displaystyle Q^{L}_{3}=1.64\times 10^{3} . Figure 4 : Posterior distributions for the GW waveform parameters \displaystyle\vec{\Theta}=(\phi,\theta,h_{0},\kappa,\xi,\delta_{0},t_{0},\alpha) , where \displaystyle(\phi,\theta) denote the GW propagation direction, \displaystyle h_{0} is the GW strain amplitude, \displaystyle\kappa and \displaystyle\xi are the two polarizations’ ratio and relative phase, \displaystyle t_{0} is the time at which the GW frequency reaches the first cavity resonance with phase \displaystyle\delta_{0} , and \displaystyle\alpha is the frequency drift rate. The red solid lines indicate the true parameter values. On the diagonal panels, black solid lines mark the posterior modes, while dashed lines indicate the \displaystyle 1\sigma ( \displaystyle 68\% ) credible intervals; the inferred values and corresponding uncertainties are listed at the top of each panel. In the off-diagonal panels, the \displaystyle 4 shaded regions represent confidence levels of \displaystyle 38\% , \displaystyle 68\% , \displaystyle 86\% , and \displaystyle 95\% , respectively. Uniform priors are assumed within the figure range for each parameter. The evolution can be approximately understood using the analytic derivation presented in Supplemental Material: \displaystyle e_{a}(t) grows on a timescale \displaystyle 1/\sqrt{\alpha} and reaches a maximum amplitude of approximately e_{a}^{\mathrm{amp}}\approx\left(1-\exp{\left[-{\frac{\gamma_{a}}{\sqrt{\alpha}}}\right]}\right)h_{0}B_{0}\omega_{a}V^{5/6}Q_{a}^{L}\eta_{\mathrm{eff}}^{a}, (III.4) where \displaystyle\eta_{\rm eff}^{a}\equiv|\eta^{a}_{+}+\exp[{\mathrm{i}\xi}]\,\eta^{a}_{\times}\kappa|/\sqrt{1+\kappa^{2}} . The amplitude then decays with a characteristic timescale of order \displaystyle 1/\gamma_{a} . In the regime \displaystyle\alpha\gtrsim 2\gamma_{a}^{2} , the time-integrated signal can be characterized by \int\left(\frac{e_{a}(t)}{e_{a}^{\mathrm{amp}}}\right)^{2}\mathrm{d}t\approx\frac{1}{4\gamma_{a}}. (III.5) We generate a mock data stream \displaystyle d_{a}^{i}=e_{a}(t_{i})+n_{a}(t_{i}) consisting of the benchmark signal plus Gaussian noise, n_{a}^{i}\sim\mathcal{N}(0,\sigma_{a}^{2}),\qquad\sigma_{a}^{2}=\frac{2\,T_{\rm eff}}{\omega_{a}\Delta t}, (III.6) where \displaystyle T_{\rm eff} is the effective noise temperature. The sampling interval is \displaystyle\Delta t=t_{\rm tot}/N_{t} , with \displaystyle t_{\rm tot} the total measurement time and \displaystyle N_{t} the number of samples; \displaystyle\Delta t must be short enough to resolve the relevant frequencies. We consider the \displaystyle 8 HFGW parameters appearing in Eq. ( III.1 ), \displaystyle\vec{\Theta}=(\phi,\theta,h_{0},\kappa,\xi,\delta_{0},t_{0},\alpha) , and construct the likelihood for a given parameter set as \displaystyle\mathcal{L}(\vec{\Theta}\,|\,d_{a}^{i})=\prod_{a}\left(2\pi\sigma_{a}^{2}\right)^{-\frac{N_{t}}{2}}\exp\!\left[-\sum_{i}\frac{\left(d_{a}^{i}-e_{a}^{\Theta}(t_{i})\right)^{2}}{2\,\sigma_{a}^{2}}\right], (III.7) where the product runs over all \displaystyle 18 cavity modes. The Fisher information matrix is then obtained by evaluating F_{jk}=\left.\Big\langle\frac{\partial^{2}[-\ln\mathcal{L}]}{\partial\Theta_{j}\,\partial\Theta_{k}}\Big\rangle\right|_{\vec{\Theta}=\vec{\Theta}_{\rm max}}, (III.8) at the maximum-likelihood point \displaystyle\vec{\Theta}_{\rm max} . The approximate signal-to-noise ratio (SNR) follows from \displaystyle{\rm SNR}^{2} \displaystyle=F_{h_{0}h_{0}}\,h_{0}^{2}=\sum_{a}\sum_{i}\left(\frac{e_{a}(t_{i})}{\sigma_{a}}\right)^{2} (III.9) \displaystyle\approx\frac{h_{0}^{2}B_{0}^{2}V^{5/3}}{2\alpha T_{\rm eff}}\sum_{a}Q_{a}^{L}\left(\omega_{a}^{2}\eta_{\rm eff}^{a}\right)^{2}. Thus the sensitivity scales as \displaystyle h_{0}\sim N^{-1/2} , effectively behaving as an \displaystyle N -detector network. For benchmark detector parameters \displaystyle B_{0}=4~\mathrm{T} , \displaystyle V=25~\mathrm{L} , \displaystyle T_{\rm eff}=10~\mathrm{mK} , \displaystyle Q_{a}^{L} ranges from \displaystyle 5.4\times 10^{2} to \displaystyle 9.1\times 10^{3} , and an average antenna factor \displaystyle\eta_{\rm eff}\sim 0.005 , the expected SNR is about \displaystyle 10 . To determine all \displaystyle 8 signal parameters, at least \displaystyle 5 cavity modes must be clearly resolved. Their relative signal strengths enable reconstruction of the GW’s overall strain amplitude, propagation direction, and the two polarizations’ ratio and relative phase. Combining their time-domain information further enables determination of the frequency drift rate, initial time, and phase, thereby matching the \displaystyle 8 unknown degrees of freedom of the signal. In our mock data analysis with the benchmark GW signal, \displaystyle 7 modes are sufficiently loud, each with an individual \displaystyle\mathrm{SNR}_{a}>1 . Fig. 4 shows the corresponding Markov chain Monte Carlo (MCMC) posterior distributions. As expected, all \displaystyle 8 parameters are successfully resolved. The parameters associated with the signal strength, \displaystyle(\phi,\theta,h_{0},\kappa,\xi) , are well reconstructed, in agreement with simple Fisher-matrix estimates, with uncertainties scaling as \displaystyle\Delta\Theta_{j}\sim F_{jj}^{-1/2}\propto\mathrm{SNR}^{-1} . In particular, the directional parameter \displaystyle\theta exhibits a much smaller relative uncertainty than \displaystyle\phi , reflecting the sharply peaked angular response in the \displaystyle\theta direction. The initial phase \displaystyle\delta_{0} is less well constrained, mainly due to the propagation of uncertainties in the frequency-drift parameter \displaystyle\alpha from louder modes to the \displaystyle a=1 mode with \displaystyle\mathrm{SNR}_{1}\simeq 0.4 . The periodicity of relative phases gives rise to secondary-peaked structures in the \displaystyle\alpha and \displaystyle t_{0} posteriors. The reconstructed signal parameters can be mapped onto a complete set of binary parameters with the same number of degrees of freedom. Notably, the usual \displaystyle\mathcal{M}_{c} – \displaystyle r degeneracy can be broken using frequency-drift information. IV Discussion HFGWs remain a largely unexplored frontier in gravitational physics, carrying potential signatures of phenomena beyond the Standard Model. EM resonant cavities, widely used in axion searches, provide a powerful platform for HFGW detection owing to their analogous microscopic coupling to EM fields. As tensorial excitations, GWs encode rich information about their sources and production mechanisms in their waveform, polarization, and propagation direction. In this work, we exploit multiple nearly degenerate modes within a single cavity to reconstruct the waveform properties of incoming HFGWs. Achieving this reconstruction requires at least \displaystyle 5 sufficiently loud modes, which motivates the use of multi-cell cavities to ensure an adequate number of detectable modes. The sensitivity scaling with the number of modes is analogous to that of a network comprising the same number of independent detectors, making a single cavity significantly more capable of probing astrophysically plausible binary sources. Our method of dissecting GW signals through comparisons among modes with distinct antenna patterns demonstrates that a single cavity can already function as an effective detector array, without the need for phase synchronization. The multi-cell configuration can be naturally generalized to a network of multiple detectors Schmieden and Schott 2024 ; Schneemann et al. 2024 . Combining internal modal diversity with spatially separated cavities, such as in a long-baseline configuration Jiang et al. 2024 , would enable even richer searches and reconstructions of HFGWs. While we have focused on HFGWs from inspiraling PBH binaries with relatively simple waveforms, more complex signals, including merger and ringdown phases Maggiore 2018 , can be naturally incorporated within our framework. In such cases, upgrading the cavity to a readout system capable of simultaneous resonant and broadband operation could significantly enhance sensitivity Chen et al. 2022b ; Wurtz et al. 2021 ; Jiang et al. 2023 ; Chen et al. 2025a . It would also be valuable to extend our analysis from time-series voltage measurements to photon-counting detection Dixit et al. 2021 ; Agrawal et al. 2024 ; Zheng et al. 2025 ; Kang et al. 2025 , which can improve the signal-to-noise ratio while sacrificing full time-domain information. Acknowledgements. We are grateful to Asher Berlin, Hanyu Cheng, Tom Krokotsch, Zhen Liu, Nick Rodd, Bo Wang, Daixu Yang, and Yanjie Zeng for useful discussions. The Center of Gravity is a Center of Excellence funded by the Danish National Research Foundation under grant No. 184. Y. C. acknowledge support by VILLUM Foundation (grant no. VIL37766) and the DNRF Chair program (grant no. DNRF162) by the Danish National Research Foundation, the European Union’s H2020 ERC Advanced Grant “Black holes: gravitational engines of discovery” grant agreement no. Gravitas–101052587, the Rosenfeld foundation in the form of an Exchange Travel Grant and by the COST Action COSMIC WISPers CA21106, supported by COST (European Cooperation in Science and Technology). Views and opinions expressed are however those of the author only and do not necessarily reflect those of the European Union or the European Research Council. Neither the European Union nor the granting authority can be held responsible for them. This project has received funding from the European Union’s Horizon 2020 research and innovation programme under the Marie Sklodowska-Curie grant agreement No 101007855 and No 101131233. J.S. is supported by the Peking University under startup Grant No. 7101302974, and the National Natural Science Foundation of China under Grants No. 12025507, No. 12450006. This publication is part of the grant PID2023-146686NB-C31 funded by MICIU/AEI/10.13039/501100011033/ and by FEDER, UE. IFAE is partially funded by the CERCA program of the Generalitat de Catalunya. This work is supported by ERC grant (GravNet, ERC-2024-SyG 101167211, DOI: 10.3030/101167211). Funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Council Executive Agency. Neither the European Union nor the granting authority can be held responsible for them. D.B. acknowledges the support from the European Research Area (ERA) via the UNDARK project of the Widening participation and spreading excellence programme (project number 101159929). Supplemental Materials: Cavity Multimodes as an Array for High-Frequency Gravitational Waves Appendix A Cavity Response to High Frequency Gravitational Waves A.1 Cavity Mode Equations in the Presence of High Frequency Gravitational Waves and a Magnetic Field Background We focus on the direct coupling of gravitational waves (GWs) to electromagnetic (EM) fields and their conversion into photon signals in the presence of a static background EM field. This process, known as the inverse Gertsenshtein effect, originates from the minimal coupling between electromagnetism and gravity, S=\int\mathrm{d}^{4}x\,\sqrt{-g}\,\left(-\frac{1}{4}\,g^{\mu\rho}g^{\nu\sigma}F_{\mu\nu}F_{\rho\sigma}\right), (S1) where \displaystyle F_{\mu\nu} is the EM field-strength tensor. Under a weak gravitational perturbation, the spacetime metric is expanded as \displaystyle g_{\mu\nu}=\eta_{\mu\nu}+h_{\mu\nu}+\mathcal{O}(h^{2}) , with \displaystyle\eta_{\mu\nu}\equiv\mathrm{diag}(-1,1,1,1) . In the presence of a background EM field \displaystyle F^{\mu\nu}_{0} , the interaction term relevant for GW–EM conversion can be written as S\supset\int\mathrm{d}^{4}x\,j^{\mu}_{\mathrm{eff}}A_{\mu},\qquad j^{\mu}_{\mathrm{eff}}=\partial_{\nu}\!\left(\frac{1}{2}\,hF_{0}^{\mu\nu}+h^{\nu}{}_{\sigma}F_{0}^{\sigma\mu}-h^{\mu}{}_{\sigma}F_{0}^{\sigma\nu}\right), (S2) where \displaystyle j^{\mu}_{\mathrm{eff}}\equiv(\rho_{\mathrm{eff}},\vec{j}_{\mathrm{eff}}) is the effective current induced by the GW in the background EM field, and \displaystyle h\equiv\eta_{\mu\nu}h^{\mu\nu} . The equation of motion for the EM field in vacuum is accordingly modified to (\nabla^{2}-\partial_{t}^{2})\vec{E}=\partial_{t}\vec{j}_{\mathrm{eff}}+\vec{\nabla}\rho_{\mathrm{eff}}. (S3) The signal EM field of interest corresponds to resonant modes inside the cavity, which receive contributions only from the effective three-current \displaystyle\vec{j}_{\mathrm{eff}} Hill 2009 . We therefore decompose the electric field in terms of a complete set of orthogonal cavity eigenmodes \displaystyle\vec{E}_{a} with eigenfrequencies \displaystyle\omega_{a} , \vec{E}(t,\vec{x})=\sum_{a}e_{a}(t)\,\vec{E}_{a}(\vec{x}),\qquad\nabla^{2}\vec{E}_{a}+\omega_{a}^{2}\vec{E}_{a}=0,\qquad\int\mathrm{d}V\,\vec{E}_{a}\cdot\vec{E}_{b}^{*}=\delta_{ab}, (S4) where \displaystyle e_{a}(t) encodes the time dependence of each mode. The volume integral is taken over the interior of the cavity. These cavity modes also approximately satisfy the conducting boundary conditions on the inner cavity surface. Projecting Eq. ( S3 ) onto a given eigenmode and including cavity dissipation, we obtain the equation of motion for \displaystyle e_{a}(t) , \ddot{e}_{a}(t)+\frac{\omega_{a}}{Q_{a}^{L}}\,\dot{e}_{a}(t)+\omega_{a}^{2}e_{a}(t)=-\int\partial_{t}\vec{j}_{\mathrm{eff}}\cdot\vec{E}_{a}^{*}\,\mathrm{d}V, (S5) where \displaystyle Q_{a}^{L} is the loaded quality factor of the cavity mode. Here, we retain only the contribution from the effective three-current \displaystyle\vec{j}_{\mathrm{eff}} . We work in the transverse-traceless (TT) gauge for the GW and expand it in terms of the two polarization states \displaystyle A=+,\times , h^{ij}(\vec{k})=\bigl(h_{+}\,\epsilon^{ij}_{+}(\hat{k})+h_{\times}\,\epsilon^{ij}_{\times}(\hat{k})\bigr)e^{-\mathrm{i}\vec{k}\cdot\vec{x}}. (S6) Here \displaystyle h_{+/\times} denote the time-dependent amplitudes of the GW strain. The polarization tensors for a plane GW propagating in the direction \displaystyle\hat{k}\equiv(\theta,\phi) are given by \epsilon^{+}_{ij}(\hat{k})=\hat{l}_{i}\hat{l}_{j}-\hat{m}_{i}\hat{m}_{j},\qquad\epsilon^{\times}_{ij}(\hat{k})=\hat{l}_{i}\hat{m}_{j}+\hat{m}_{i}\hat{l}_{j}, (S7) where \displaystyle\hat{k} , \displaystyle\hat{l} , and \displaystyle\hat{m} form an orthonormal triad. In the cavity Cartesian frame \displaystyle(x,y,z) , expressed in terms of spherical coordinates, these unit vectors take the form \displaystyle\hat{k} \displaystyle=\left(\sin\theta\cos\phi,\,\sin\theta\sin\phi,\,\cos\theta\right)^{\mathrm{T}}, (S8) \displaystyle\hat{l} \displaystyle=\left(\cos\theta\cos\phi,\,\cos\theta\sin\phi,\,-\sin\theta\right)^{\mathrm{T}}, \displaystyle\hat{m} \displaystyle=\left(-\sin\phi,\,\cos\phi,\,0\right)^{\mathrm{T}}. In our setup, we consider a static background magnetic field of uniform strength \displaystyle B_{0} . The effective current induced by the GW can then be decomposed as \vec{j}_{\mathrm{eff}}(\vec{k})\equiv B_{0}|\vec{k}|^{2}V^{1/3}\bigl(h_{+}\hat{j}_{+}+h_{\times}\hat{j}_{\times}\bigr), (S9) where \displaystyle\hat{j}_{+/\times} are dimensionless spatial profiles determined by the GW polarization and momentum. The induced signal in a given cavity mode is proportional to the overlap between the effective current and the cavity electric-field distribution. This is characterized by the dimensionless overlap function, or antenna pattern, \eta^{a}_{+/\times}(\vec{k})\equiv\frac{1}{V^{1/2}}\int_{V}\hat{j}_{+/\times}\cdot\vec{E}_{a}^{*}\,\mathrm{d}V, (S10) where the integral is taken over the cavity volume. In general, the antenna pattern depends on both the GW propagation direction \displaystyle\hat{k} and the GW frequency \displaystyle|\vec{k}| . In the main text, however, we evaluate \displaystyle\eta^{a}_{+/\times} at the cavity resonance, \displaystyle|\vec{k}|\simeq\omega_{a} , since the relevant signal is confined to a narrow bandwidth around the resonant frequency. With these reparameterizations, Eq. ( S5 ) can be rewritten as \ddot{e}_{a}(t)+\frac{\omega_{a}}{Q_{a}^{L}}\,\dot{e}_{a}(t)+\omega_{a}^{2}e_{a}(t)\simeq-\mathrm{i}\,\omega_{a}^{3}V^{5/6}B_{0}\sum_{A=+,\times}h_{A}\,\eta^{a}_{A}. (S11) A.2 Time-Domain Waveforms of GW-Excited Cavity Modes We next compute the time evolution of GW-excited cavity modes driven by the waveform given in Eq. (III.1) of the main text, which is motivated by a binary inspiral. To this end, we first reparameterize the source term on the right-hand side of Eq. ( S11 ) as \sum_{A=+,\times}h_{A}\,\eta_{A}^{a}\equiv h_{0}\,\eta_{\rm eff}^{a}\,\exp\left[{\mathrm{i}(\varphi_{a}+\omega_{1}(t-t_{0})+\alpha(t-t_{0})^{2}/2})\right], (S12) where \displaystyle h_{0} is the GW strain amplitude, \displaystyle\eta_{\rm eff}^{a} is the effective overlap function for mode \displaystyle a , and \displaystyle\varphi_{a} encodes the relative phase. The relevant quantities are defined as \displaystyle\displaystyle h_{0} \displaystyle\displaystyle\equiv(h_{+}^{2}+h_{\times}^{2})^{1/2}, (S13) \displaystyle\displaystyle\eta^{a}_{\rm eff} \displaystyle\displaystyle\equiv\frac{1}{\sqrt{1+\kappa^{2}}}\sqrt{(\eta_{+}^{a})^{2}+(\kappa\eta_{\times}^{a})^{2}+2\left|\kappa\eta_{+}^{a}\eta_{\times}^{a}\right|\cos{(\Omega^{a}_{\times}-\Omega^{a}_{+}+\xi)}},\qquad\kappa\equiv|h_{\times}/h_{+}|,\qquad\Omega_{+/\times}^{a}\equiv{\rm arg}[\eta^{a}_{+/\times}], (S14) \displaystyle\displaystyle\varphi_{a} \displaystyle\displaystyle\equiv\delta_{0}+\Omega_{+}^{a}+{\rm arctan}\left[\frac{\kappa\left|\eta^{a}_{\times}\right|\sin{(\Omega^{a}_{\times}-\Omega^{a}_{+}+\xi)}}{\left|\eta^{a}_{+}\right|+\kappa\left|\eta^{a}_{\times}\right|\cos{(\Omega^{a}_{\times}-\Omega^{a}_{+}+\xi)}}\right], (S15) Here, \displaystyle t_{0} denotes the time at which the GW frequency reaches the first cavity resonance with initial phase \displaystyle\delta_{0} , \displaystyle\kappa and \displaystyle\xi are the two polarizations’ ratio and relative phase, and \displaystyle\alpha is the linear angular frequency drift rate. Substituting Eq. ( S12 ) back into Eq. ( S11 ), the mode amplitude can be solved using the Green’s function method: e_{a}(t)=-\mathrm{i}\omega_{a}^{3}B_{0}V^{5/6}h_{0}\eta^{a}_{\rm eff}\exp[{\mathrm{i}\varphi_{a}}]\int_{-\infty}^{t^{\prime}}G_{a}(t^{\prime}-\tau)\exp[{\mathrm{i}(\omega_{1}\tau+\alpha\tau^{2}/2})]\text{d}\tau,\qquad G_{a}(t)=\exp[{-\gamma_{a}t}]\frac{\sin\left(\sqrt{\omega_{a}^{2}-\gamma_{a}^{2}}t\right)}{\sqrt{\omega_{a}^{2}-\gamma_{a}^{2}}}H(t), (S16) where \displaystyle t^{\prime}\equiv t-t_{0} , \displaystyle\gamma_{a}\equiv\omega_{a}/(2Q_{a}^{L}) is the dissipation rate and \displaystyle H(t) is the Heaviside step function. This expression admits a closed-form analytic solution, e_{a}(t)=\frac{-\omega_{a}^{2}B_{0}V^{5/6}h_{0}\eta^{a}_{\rm eff}}{2}\exp[{\mathrm{i}\varphi_{a}}]\left(\exp[{-\lambda_{-}^{a}t^{\prime}}]I_{-}^{a}(t^{\prime})-\exp[{-\lambda_{+}^{a}t^{\prime}}]I_{+}^{a}(t^{\prime})\right), (S17) where we have defined I_{\pm}^{a}(t^{\prime})\equiv\sqrt{\frac{\pi}{2\alpha}}\exp\left[{\mathrm{i}\left(\frac{\pi}{4}+\frac{(b_{\pm}^{a})^{2}}{4p}\right)}\right]\left(\text{erf}\left[\sqrt{p}(t^{\prime}-\frac{b_{\pm}^{a}}{2p})\right]+1\right), (S18) with \displaystyle\lambda^{a}_{\pm}=\gamma_{a}\pm\sqrt{\omega_{a}^{2}-\gamma_{a}^{2}} , \displaystyle b_{\pm}^{a}=\gamma_{a}+\mathrm{i}(\omega_{1}\pm\sqrt{\omega_{a}^{2}-\gamma_{a}^{2}}) , and \displaystyle p=-\mathrm{i}\alpha/2 . A.2.1 Approximate Waveforms The exact analytic expression for the cavity-mode response derived above is rather involved. To highlight the dominant physical features of the signal, we focus here on its approximate scaling behavior and time-domain structure. Each cavity mode can be modeled as a damped harmonic oscillator driven by a chirping, nearly monochromatic GW with instantaneous angular frequency \omega_{g}(t)=\omega_{1}+\alpha(t-t_{0}). (S19) As the GW frequency sweeps across the cavity resonance \displaystyle\omega_{a} , the mode response is governed by the interplay of three characteristic time scales: (i) the inverse detuning \displaystyle|\omega_{g}(t)-\omega_{a}|^{-1} , (ii) the cavity damping time \displaystyle 1/\gamma_{a} , and (iii) the frequency-sweep time scale \displaystyle\alpha^{-1/2} . The interplay of these time scales leads to three distinct stages in the response: off-resonant tracking, resonant growth, and post-resonance decay. To make this structure explicit, it is useful to define the instantaneous saturation amplitude, namely the steady-state response that would be obtained if the GW frequency were held fixed at \displaystyle\omega_{g}(t) : e_{a}^{\rm sat}(t)\equiv\left|\frac{\omega_{a}^{3}B_{0}V^{5/6}h_{0}\eta_{\rm eff}^{a}}{\omega_{a}^{2}-\omega_{g}(t)^{2}-2\mathrm{i}\omega_{g}(t)\gamma_{a}}\right|. (S20) The actual signal amplitude \displaystyle e_{a}(t) attempts to follow this moving saturation envelope but may lag behind when the frequency sweep is sufficiently rapid. In what follows, we first focus on the intermediate regime 2\gamma_{a}^{2}\;\leq\;\alpha\;<\;\gamma_{a}\omega_{a}, (S21) for which the frequency evolution is fast enough to invalidate adiabatic tracking near resonance, yet slow enough that the signal remains nearly monochromatic at any given time. • Stage I: off-resonant tracking. At early times, when \displaystyle|\omega_{g}(t)-\omega_{a}|\gg\sqrt{\alpha/2\pi} , the GW drive is far off resonance and the saturation amplitude is small. In this regime the cavity mode remains effectively saturated, and the signal amplitude closely follows the instantaneous steady-state value, |e_{a}(t)|^{\rm(I)}\simeq e_{a}^{\rm sat}(t). (S22) Because the Lorentzian envelope in Eq. ( S20 ) grows only slowly far from resonance, the signal remains negligible during this stage. As the GW frequency approaches \displaystyle\omega_{a} , the saturation amplitude increases rapidly. Stage I ends at the time \displaystyle t_{a}^{s} when the cavity can no longer adiabatically track this growth. A useful estimate for \displaystyle t_{a}^{s} is obtained by equating the rate of change of the saturation amplitude to the intrinsic growth rate of the cavity mode, \left.\frac{{\rm d}|e_{a}|}{{\rm d}t}\right|_{t=t_{a}^{s}}\sim\alpha\left.\frac{{\rm d}e_{a}^{\rm sat}}{{\rm d}\omega_{g}}\right|_{t=t_{a}^{s}}\sim\omega_{a}^{2}B_{0}V^{5/6}h_{0}\eta_{\rm eff}^{a}. (S23) • Stage II: resonant growth. For \displaystyle t>t_{a}^{s} , the instantaneous saturation amplitude exceeds the actual signal amplitude, and the cavity mode undergoes resonant energy accumulation. During this stage, the signal grows approximately as an effectively monochromatic response within the resonant bandwidth, with the growth controlled by the cavity dissipation rate \displaystyle\gamma_{a} |e_{a}(t)|^{\rm(II)}\simeq\left[1-e^{-\gamma_{a}(t-t_{a}^{s})}\right]\left(\omega_{a}B_{0}V^{5/6}h_{0}\eta_{\rm eff}^{a}Q_{a}^{L}\right). (S24) This growth persists only while the GW frequency remains sufficiently close to resonance. Once \displaystyle\omega_{g}(t) exceeds \displaystyle\omega_{a} , the saturation amplitude \displaystyle e_{a}^{\rm sat}(t) decreases rapidly. Stage II ends at time \displaystyle t_{a}^{e} , defined by the condition e_{a}^{\rm sat}(t_{a}^{e})\sim|e_{a}(t_{a}^{e})|^{\rm(II)}. (S25) The duration of the resonant growth window is primarily controlled by the frequency sweep rate and can be estimated as t_{a}^{e}-t_{a}^{s}\sim\frac{1}{\sqrt{\alpha}}. (S26) This timescale is longer than the naive estimate based on the interval during which the GW frequency remains within the cavity bandwidth, \displaystyle\sim\gamma_{a}/\alpha . This indicates that significant signal growth can occur even when the GW frequency lies outside the nominal resonance bandwidth, as the cavity mode has not yet reached its saturation limit. • Stage III: post-resonance decay. For \displaystyle t>t_{a}^{e} , the driving frequency has moved off resonance and the cavity mode undergoes free decay with damping rate \displaystyle\gamma_{a} , accompanied by weak beat oscillations. The signal amplitude in this stage is approximately |e_{a}(t)|^{\rm(III)}\simeq\left[1-e^{-\gamma_{a}(t_{a}^{e}-t_{a}^{s})}\right]\left(\omega_{a}B_{0}V^{5/6}h_{0}\eta_{\rm eff}^{a}Q_{a}^{L}\right)e^{-\gamma_{a}(t-t_{a}^{e})}. (S27) Since the signal growth during Stage I is negligible, the maximum oscillation amplitude is well approximated by the value reached at the end of Stage II, e_{a}^{\rm amp}\equiv|e_{a}(t_{a}^{e})|\simeq\left[1-e^{-\gamma_{a}/\sqrt{\alpha}}\right]\left(\omega_{a}B_{0}V^{5/6}h_{0}\eta_{\rm eff}^{a}Q_{a}^{L}\right). (S28) At any given time, the signal remains nearly monochromatic, allowing the time-integrated power to be estimated by averaging over oscillations. One finds \displaystyle\int_{-\infty}^{\infty}e_{a}(t)^{2}\,{\rm d}t \displaystyle\approx\int_{t_{a}^{s}}^{t_{a}^{e}}\frac{1}{2}\left[e_{a}^{\rm amp}\left(1-e^{-\gamma_{a}(t-t_{a}^{s})}\right)\right]^{2}{\rm d}t+\int_{t_{a}^{e}}^{\infty}\frac{1}{2}\left[e_{a}^{\rm amp}e^{-\gamma_{a}(t-t_{a}^{e})}\right]^{2}{\rm d}t (S29) \displaystyle\simeq\frac{(e_{a}^{\rm amp})^{2}}{4\gamma_{a}}\left[1+\mathcal{O}\!\left(\left(\frac{\gamma_{a}}{\sqrt{\alpha}}\right)^{3}\right)\right]. For sufficiently slow frequency evolution, \alpha\leq 2\gamma_{a}^{2}, (S30) the cavity mode remains adiabatically saturated throughout the resonance crossing. In this regime, the full time-domain response is well approximated by |e_{a}(t)|\simeq e_{a}^{\rm sat}(t), (S31) and no distinct resonant growth stage develops. Appendix B 9-Cell Cavity B.1 Mode Properties In this work, we adopt a 9-cell elliptical TESLA cavity as the benchmark configuration, as shown in Fig. S1 . Such a multi-cell cavity can be viewed as a natural generalization of a single-cell cavity: each mode with indices \displaystyle(m,n,p) in an axisymmetric single-cell cavity splits into \displaystyle N=9 nearly degenerate modes when extended to an \displaystyle N -cell structure, analogous to a system of \displaystyle N coupled harmonic oscillators. Figure S1 : Side view of the 9-cell cavity along the \displaystyle y direction. Two straight antenna couplers are employed: the left coupler predominantly reads out the \displaystyle\mathrm{TE}_{111+} modes, while the right coupler reads out the \displaystyle\mathrm{TE}_{111-} modes. Within this multiplet, the \displaystyle a -th mode ( \displaystyle a=1,\ldots,N ) exhibits relative electric-field amplitudes and phases across the cells such that the field in the \displaystyle j -th cell is proportional to \displaystyle\sin[(j-1/2)a\pi/N] , as illustrated in Fig. 1 of the main text. These modes range from the in-phase configuration, where all cells oscillate coherently ( \displaystyle a=1 ), to configurations in which adjacent cells oscillate out of phase ( \displaystyle a=N ). In addition, axial symmetry implies an extra two-fold degeneracy for modes with \displaystyle m\neq 1 , which we label by \displaystyle\pm following the \displaystyle(m,n,p) indices. \displaystyle a 1 2 3 4 5 6 7 8 9 \displaystyle\mathrm{TE}_{111+} \displaystyle f_{a} [GHz] 1.51 1.52 1.54 1.56 1.60 1.63 1.67 1.71 1.76 \displaystyle Q^{L}_{a} 8600 3192 1640 1037 753 608 542 563 785 \displaystyle\beta_{a} 1.1 4.7 10.4 17.6 26.2 34.4 40.6 40.1 33.2 \displaystyle|\eta_{\mathrm{eff}}^{a}|\times 10^{3} 1.0 12.4 20.4 13.1 0.6 5.1 1.3 2.7 2.6 \displaystyle\mathrm{SNR}^{2}_{a} 0.44 24.36 34.89 9.63 0.02 0.96 0.06 0.30 0.41 \displaystyle\mathrm{TE}_{111-} \displaystyle f_{a} [GHz] 1.52 1.53 1.55 1.58 1.61 1.64 1.68 1.72 1.76 \displaystyle Q^{L}_{a} 9112 3448 1785 1136 832 681 624 696 1557 \displaystyle\beta_{a} 1.0 4.5 9.9 16.6 24.2 31.1 35.3 31.7 17.2 \displaystyle|\eta_{\mathrm{eff}}^{a}|\times 10^{3} 0.6 0.72 11.9 7.6 1.1 2.9 1.8 1.1 4.1 \displaystyle\mathrm{SNR}^{2}_{a} 0.16 9.18 13.28 3.62 0.06 0.36 0.14 0.12 1.99 Table 1 : Properties of the 18 cavity modes, including the \displaystyle\mathrm{TE}_{111+} and \displaystyle\mathrm{TE}_{111-} families. Listed are the resonant frequencies \displaystyle f_{a}\equiv\omega_{a}/(2\pi) , loaded quality factors \displaystyle Q_{a}^{L} , and coupling coefficients \displaystyle\beta_{a} . Also provided are the effective overlap functions \displaystyle|\eta_{\mathrm{eff}}^{a}| , as well as the expected contributions to the squared signal-to-noise ratio, \displaystyle\mathrm{SNR}_{a}^{2} , for the benchmark GW signals. We focus on the lowest transverse electric modes \displaystyle\mathrm{TE}_{111\pm} with \displaystyle a=1,\ldots,9 . Their properties are simulated using COMSOL, and the resulting resonant frequencies \displaystyle\omega_{a}\equiv 2\pi f_{a} are summarized in Table 1 . To lift the degeneracy between the \displaystyle\mathrm{TE}_{111+} and \displaystyle\mathrm{TE}_{111-} modes, we introduce a small ellipticity of approximately \displaystyle 2\% in the cavity geometry. These modes are read out using two straight antenna couplers oriented in orthogonal directions, as also shown in Fig. S1 . The left antenna predominantly couples to the \displaystyle\mathrm{TE}_{111+} modes, while the right antenna couples to the \displaystyle\mathrm{TE}_{111-} modes. The corresponding loaded quality factors \displaystyle Q_{a}^{L} and coupling coefficients \displaystyle\beta_{a}\equiv Q^{0}_{a}/Q_{a}^{L}-1 span the range from critical coupling to the over-coupled regime. The frequency separation between different modes can be readily covered by broadband amplifiers, such as a Josephson traveling-wave parametric amplifier (JTWPA) Bartram et al. 2023 . The overlap functions \displaystyle\eta^{a}_{+/\times} are also computed by extracting the electric-field configurations from COMSOL simulations. In this calculation, we assume that the GW frequency \displaystyle|\vec{k}| matches the cavity resonant frequency \displaystyle\omega_{a} for each mode. For a benchmark GW incident direction, we evaluate each mode’s contribution to the squared signal-to-noise ratio, \displaystyle\mathrm{SNR}_{a}^{2} , as well as the corresponding effective overlap functions \displaystyle\eta_{\mathrm{eff}}^{a} , following the procedure described in the main text. B.2 Overlap Functions Figure 2 in the main text shows the overlap functions \displaystyle\eta_{A}^{a} for GWs incident from direction \displaystyle\hat{k}=(\phi,\theta) , for the two polarizations \displaystyle A=+,\times . In this subsection, we explain the origin of their angular dependence and the symmetry relations connecting different regions of the sky. We first note that the overlap functions for each cavity mode exhibit a pronounced preference for specific ranges of the polar angle \displaystyle\theta . This behavior can be understood as a phase-matching condition between the electric-field distributions in the 9 cavity cells and the variation of the GW plane-wave phase along the \displaystyle z direction. Denoting the distance between the centers of two adjacent cells by \displaystyle L_{z}=0.114~\mathrm{m} , we obtain the parametric scaling \left|\eta^{a}_{+/\times}\right|\propto\left|\sum_{j=1}^{N}e^{\mathrm{i}kjL_{z}\cos\theta}\sin\!\left[(j-\tfrac{1}{2})\frac{a\pi}{N}\right]\right|\propto\left|\frac{\sin\!\left(\tfrac{N}{2}\psi\right)}{\sin\!\left(\tfrac{\psi}{2}\right)}\right|,\qquad\psi\equiv kL_{z}\cos\theta-\frac{a\pi}{N}. (S32) This expression naturally leads to the multi-peak structure observed in \displaystyle\theta , with the global maximum attained when the phase-matching condition \displaystyle\psi=0 is satisfied. In addition, the azimuthal dependence can be understood by projecting the effective current onto the \displaystyle x direction, along which the electric field of the \displaystyle\mathrm{TE}_{111+} modes predominantly lies. One finds \hat{j}_{+}\cdot\hat{x}\propto\sin\theta\,\sin\phi,\qquad\hat{j}_{\times}\cdot\hat{x}\propto\sin\theta\,\cos\theta\,\sin\phi. (S33) This \displaystyle\phi dependence is in good agreement with the numerical overlap functions. The additional factor of \displaystyle\cos\theta in \displaystyle\hat{j}_{\times}\cdot\hat{x} further explains the relatively suppressed amplitude of \displaystyle\eta^{a}_{\times} compared to \displaystyle\eta^{a}_{+} . In Fig. 2 of the main text, we display the overlap functions \displaystyle\eta_{A}^{a}(\hat{k}) only over one-eighth of the sky, corresponding to \displaystyle\phi\in[0,\pi/2] and \displaystyle\theta\in[0,\pi/2] , for the \displaystyle\mathrm{TE}_{111+} modes. The overlap functions in the remaining regions of the sky, as well as those for the \displaystyle\mathrm{TE}_{111-} modes, can be obtained straightforwardly using discrete symmetry transformations, as detailed below. The combined system of the nearly axisymmetric cavity, the background magnetic field, and an incident plane-wave GW admits three independent discrete parity symmetries acting on the celestial angles. These symmetries induce a \displaystyle Z_{8} degeneracy in the overlap functions defined in Eq. ( S10 ). Explicitly, the system is invariant (up to overall sign changes) under the following transformations: (i) \displaystyle\phi\rightarrow\pi+\phi , corresponding to a parity transformation in the \displaystyle x – \displaystyle y plane; (ii) \displaystyle\phi\rightarrow\pi-\phi , corresponding to a parity transformation along the \displaystyle x direction; (iii) \displaystyle\theta\rightarrow\pi-\theta , corresponding to a parity transformation along the \displaystyle z direction. Under each of these transformations, the overlap function \displaystyle\eta_{A}^{a} changes only by an overall sign, determined by the parity of its constituent factors. The transformation properties of the GW polarizations \displaystyle h_{+/\times} follow from those of the polarization basis vectors \displaystyle\hat{l} and \displaystyle\hat{m} . The background electromagnetic field strength \displaystyle F^{ij}_{0} changes sign only under transformation (ii), reflecting the presence of a static magnetic field oriented along the \displaystyle z direction. The cavity eigenmode profiles \displaystyle\vec{E}_{a} for the \displaystyle\mathrm{TE}_{111+} modes acquire a factor of \displaystyle-1 under transformations (i) and (ii), consistent with their \displaystyle l=1 , \displaystyle m=1 angular structure. The transformation (iii) depends on the parity of the mode index \displaystyle a , reflecting whether an even or odd number of \displaystyle\pi phase shifts is accumulated under reflection about the \displaystyle z axis. Combining all contributions, the transformation properties of each component entering \displaystyle\eta_{A}^{a} for the \displaystyle\mathrm{TE}_{111+} modes are summarized in Table 2 . \displaystyle h_{\times} \displaystyle h_{+} \displaystyle F^{ij}_{0} \displaystyle\hat{j}_{\times} \displaystyle\hat{j}_{+} \displaystyle\vec{E}_{a}(\vec{r}\,) \displaystyle\eta_{\times}^{a} \displaystyle\eta_{+}^{a} \displaystyle\phi\rightarrow\pi+\phi 1 1 1 1 1 \displaystyle-1 \displaystyle-1 \displaystyle-1 \displaystyle\phi\rightarrow\pi-\phi \displaystyle-1 1 \displaystyle-1 1 \displaystyle-1 \displaystyle-1 \displaystyle-1 1 \displaystyle\theta\rightarrow\pi-\theta \displaystyle-1 1 1 \displaystyle-1 1 \displaystyle(-1)^{a-1} \displaystyle(-1)^{a} \displaystyle(-1)^{a-1} Table 2 : Parity properties of each component entering the overlap functions \displaystyle\eta_{A}^{a} for the \displaystyle\mathrm{TE}_{111+} modes. The \displaystyle\mathrm{TE}_{111-} modes are related to the \displaystyle\mathrm{TE}_{111+} modes by a rotation of \displaystyle\pi/2 about the \displaystyle z axis, corresponding to the transformation \displaystyle\phi\rightarrow\phi+\pi/2 . Due to the above symmetries, the overlap functions in eight distinct sky octants are identical up to an overall sign, resulting in an apparent eightfold degeneracy in the GW arrival direction. When multiple cavity modes with different indices \displaystyle a are measured, their distinct symmetry behaviors, particularly under transformation (iii), allow one to access the relative phase between the two GW polarizations. Since \displaystyle 4 of the \displaystyle 8 octants differ by a relative sign between the \displaystyle+ and \displaystyle\times components compared to the other \displaystyle 4 , this information reduces the degeneracy from eightfold to fourfold. The remaining fourfold degeneracy could, in principle, be lifted by introducing controlled inhomogeneities in the background magnetic field, which would break the discrete symmetries discussed above and yield quadrant-dependent overlap functions. Implementing and calibrating such magnetic-field configurations would require additional experimental complexity, and we defer a detailed investigation of this possibility to future work. Appendix C Benchmark Primordial Black Hole Binary The GW strain adopted in Eq. (III.1) of the main text is motivated by an inspiraling primordial black hole (PBH) binary. We consider a GW with instantaneous angular frequency \displaystyle\omega_{g}=|\vec{k}| that drifts slowly with a constant rate \displaystyle\alpha=1.2\times 10^{14}~\mathrm{rad}\,\mathrm{s}^{-2} , and with the two GW polarizations differing by a relative phase of \displaystyle-\pi/2 . At this drift rate, the GW frequency sweeps across all eighteen cavity modes, which have a typical frequency spacing of \displaystyle\sim 0.25~\mathrm{GHz} , within approximately \displaystyle 13~\mu\mathrm{s} . We further assume a circular orbit observed face-on. For a circular inspiral, the GW strain amplitude and frequency drift rate are related to the PBH binary parameters by Maggiore 2018 \displaystyle\displaystyle h_{0} \displaystyle\displaystyle=\frac{4}{r}\,\left(G\mathcal{M}_{c}\right)^{5/3}\left(\frac{\omega_{g}}{2}\right)^{2/3}, (S34) \displaystyle\displaystyle\alpha \displaystyle\displaystyle=\frac{12}{5}\,2^{1/3}\left(G\mathcal{M}_{c}\right)^{5/3}\omega_{g}^{11/3}, (S35) where \displaystyle\mathcal{M}_{c} is the chirp mass of the binary and \displaystyle r is the luminosity distance to the source. In this work, we retain only the linear term in the frequency evolution; a more accurate treatment of the chirp can be straightforwardly incorporated into our framework without introducing additional parameters for circular orbits. We leave the investigation of more complex waveforms, including those beyond the standard circular-orbit case Barrau et al. 2025a ; Teuscher et al. 2024 ; Barrau et al. 2025b ; Cusin et al. 2024 ; Jamet et al. 2025 , for future work. Solving for the chirp mass and distance in terms of the benchmark strain amplitude \displaystyle h_{0} , frequency drift rate \displaystyle\alpha , and GW frequency \displaystyle\omega_{g} , we obtain \displaystyle\displaystyle\mathcal{M}_{c} \displaystyle\displaystyle=0.96\times 10^{-3}\,M_{\oplus}\left(\frac{\alpha~[\mathrm{rad}/\mathrm{s}^{2}]}{1.2\times 10^{14}}\right)^{3/5}\left(\frac{1.55~\mathrm{GHz}}{\omega_{g}/(2\pi)}\right)^{11/5}, (S36) \displaystyle\displaystyle r \displaystyle\displaystyle=0.19~\mathrm{AU}\,\left(\frac{\alpha~[\mathrm{rad}/\mathrm{s}^{2}]}{1.2\times 10^{14}}\right)\left(\frac{10^{-18}}{h_{0}}\right)\left(\frac{1.55~\mathrm{GHz}}{\omega_{g}/(2\pi)}\right)^{3}. (S37) Note that in our setup there is no degeneracy between the two, owing to the additional information provided by the frequency drift rate \displaystyle\alpha .