Inverse-Designed Optical Cavity

• The paper details an inverse-design paradigm that transforms cavity performance targets into optimized dielectric profiles using iterative, multi-objective optimization.
• It demonstrates how design variables—such as binary index fields, hole displacements, and grating widths—achieve specific figures of merit like quality factor, mode volume, and coupling efficiency.
• It highlights experimental validations where optimized cavities deliver enhanced near-field chirality, high Q factors, and tailored radiation patterns across diverse photonic platforms.

Searching arXiv for the cited paper and closely related inverse-designed optical cavity work. arXiv search query: "(Romashkina et al., 22 Dec 2025) inverse-designed optical cavity cavity inverse design photonic crystal optomechanical" An inverse-designed optical cavity is a resonant photonic structure in which cavity performance is specified rather than the geometry, and the final dielectric profile, hole pattern, reflector topology, or resonator cross section is obtained through iterative optimization. Across recent arXiv literature, the term encompasses all-dielectric mirror resonators, Fabry–Pérot cavities, photonic-crystal cavities, meta-cavities, release-free optomechanical crystals, whispering-gallery cavities, and hollow nanowire resonators. The common premise is that target observables such as quality factor, mode volume, dispersion, far-field numerical aperture, optical chirality, photon-pair generation efficiency, or photon-phonon coupling are encoded directly in a figure of merit, while fabrication rules are imposed during optimization rather than added afterward (Bi et al., 19 May 2025, Romashkina et al., 22 Dec 2025).

1. Definition and design scope

Inverse design in cavity photonics is motivated by the fact that optical cavities confine light to subwavelength volumes, enabling strong light-matter interactions, while practical devices often require simultaneous control of high quality factors, low mode volumes, and high coupling efficiencies. Because optimizing across these metrics requires exploring a large design space, recent work formulates cavity synthesis as a constrained optimization problem over a material distribution or a reduced geometry parameter set rather than as a manual sweep over intuitive shape parameters (Vij et al., 20 Sep 2025).

The design variables differ by platform. In the dielectric meta-cavity for superchiral hot spots, the variable is a binary “index-density” field $\rho(x,y)$ within a $2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$ unit cell, with $\rho\in\{0,1\}$ mapping to air or Si, and a fixed film thickness $t=200\,\mathrm{nm}$ (Romashkina et al., 22 Dec 2025). In visible SiN L3 photonic-crystal cavities, the variables are the hole displacements in the first quadrant, $p\equiv\{\delta x_i,\delta y_i\}$ for $i=1\ldots N$, with $N=16\times 16$ holes and bounds $|\delta x_i|,|\delta y_i|\le 0.1a$ (Vij et al., 20 Sep 2025). In thick-silicon-nitride Fabry–Pérot cavities, the unknown is a design-variable field $\rho(x)\in[0,1]$ mapped to local permittivity, while in an interpretable nonlinear silicon cavity the variables are the silicon grating widths $\{w_n\}$ and gaps $2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$0 of $2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$1 periods (Bi et al., 19 May 2025, Jia et al., 2023).

This breadth of parameterizations reflects a broader point: inverse-designed optical cavities are not a single geometry class but a design paradigm. Published implementations range from a compact all-dielectric parallel-plane mirror resonator with a footprint of $2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$2 (Tutgun et al., 2019) to a gallium nitride hexagonal hollow nanowire whispering-gallery cavity optimized for orbital angular momentum generation (Lim et al., 12 Dec 2025) and a release-free silicon optomechanical crystal cavity optimized for vacuum optomechanical coupling (Hambraeus et al., 5 May 2026).

2. Figures of merit and electromagnetic observables

The canonical cavity figures of merit remain the quality factor and mode volume. In the meta-cavity work, the quality factor is written as

$2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$3

where $2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$4 is the full-width at half maximum of the resonance, and the electric mode volume is

$2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$5

For small-$2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$6, high-$2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$7 dielectric cavities, a large $2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$8 ratio leads to strong field confinement and enhanced light-matter interactions (Romashkina et al., 22 Dec 2025).

Inverse design extends this conventional $2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$9- and $\rho\in\{0,1\}$0-centric description to application-specific observables. In the superchiral dielectric meta-cavity, the target is near-field optical chirality at a selected point under linearly polarized excitation,

$\rho\in\{0,1\}$1

with enhancement normalized to the chirality of a circularly polarized plane wave of the same power in free space (Romashkina et al., 22 Dec 2025). In the SiN L3 cavity framework, the composite loss simultaneously drives the quality factor toward a target $\rho\in\{0,1\}$2 and the far-field coupling efficiency toward a target numerical aperture $\rho\in\{0,1\}$3,

$\rho\in\{0,1\}$4

so that the cavity is designed not only for storage but also for a prescribed radiation pattern (Vij et al., 20 Sep 2025).

In optomechanical cavities, the figure of merit may be the single-photon optomechanical scattering rate,

$\rho\in\{0,1\}$5

which depends on the optical eigenvalue, the mechanical eigenvalue, and the moving-boundary and photoelastic contributions to $\rho\in\{0,1\}$6 (Hambraeus et al., 5 May 2026). In nonlinear and quantum photonic cavities, the objective can instead be a nonlinear overlap integral for spontaneous four-wave mixing,

$\rho\in\{0,1\}$7

with $\rho\in\{0,1\}$8, or a reflection-based merit function that targets strong reflectivity at multiple wavelengths (Jia et al., 2023, Bi et al., 19 May 2025).

A recurrent feature is that cavity optimization is no longer limited to internal eigenmode confinement. It includes externally measurable quantities such as resonant circular dichroism, Gaussian collection efficiency, output beam overlap, or even a cavity transfer function used for analog multiplication (Romashkina et al., 22 Dec 2025, Diez et al., 2022, Mathur, 18 Jul 2025).

3. Optimization methodologies

A central methodological distinction is between objective-first, adjoint-based, differentiable, and heuristic optimization schemes. An early all-dielectric parallel-plane mirror resonator used the objective-first inverse-design algorithm, which alternates between a field subproblem and a structure subproblem in the discretized Maxwell operator $\rho\in\{0,1\}$9, while constraining the permittivity between air and silicon values and adding a binarization-cost regularizer (Tutgun et al., 2019). This formulation exemplifies the original cavity inverse-design idea: the algorithm creates a bandgap in a random structure to obtain a resonance peak at the desired wavelength without intuitive scanning of the parameter space (Tutgun et al., 2019).

Recent cavity work is dominated by adjoint sensitivities and automatic differentiation. In thick-SiN Fabry–Pérot cavities, the SPINS framework computes forward fields at a set of target frequencies, defines adjoint sources proportional to $t=200\,\mathrm{nm}$0, and evaluates the gradient

$t=200\,\mathrm{nm}$1

with spatial filtering and projection enforcing minimum feature size and curvature constraints (Bi et al., 19 May 2025). In visible SiN L3 cavities, the guided-mode expansion simulator is implemented in Legume so that $t=200\,\mathrm{nm}$2 is obtained by back-propagation rather than a hand-derived adjoint, and convergence occurs in approximately $t=200\,\mathrm{nm}$3 epochs to $t=200\,\mathrm{nm}$4 (Vij et al., 20 Sep 2025).

The dielectric superchiral meta-cavity introduces a two-step inverse-design pipeline. The first stage is a rapid global search with a neural-network-parameterized topology and a differentiable RCWA solver; a small random seed vector is fed through fully connected and convolutional layers to produce a pixelated $t=200\,\mathrm{nm}$5, and Adam trains the network parameters to maximize the chirality figure of merit. The second stage is a local geometry refinement with FDTD plus adjoint sensitivities, where direct and adjoint FDTD runs compute $t=200\,\mathrm{nm}$6, followed by gradient-ascent updates and periodic spatial filtering and Heaviside projection. After approximately $t=200\,\mathrm{nm}$7 iterations, a fully binary, fabrication-aware design emerges (Romashkina et al., 22 Dec 2025).

Other cavity classes require different optimizers. The hollow GaN nanowire OAM cavity uses COMSOL eigenfrequency studies coupled to Fuzzy Self-Tuning PSO with $t=200\,\mathrm{nm}$8 particles and a stop criterion of $t=200\,\mathrm{nm}$9 consecutive epochs with no global-best improvement (Lim et al., 12 Dec 2025). The release-free optomechanical crystal combines a physics-guided starting point, XHOPE, with multiphysics inverse design; each iteration requires two eigenmode solves and two adjoint solves, for $p\equiv\{\delta x_i,\delta y_i\}$0 total FEM solves per iteration, with typical convergence in approximately $p\equiv\{\delta x_i,\delta y_i\}$1–$p\equiv\{\delta x_i,\delta y_i\}$2 iterations (Hambraeus et al., 5 May 2026).

The literature also identifies an interpretability problem. One paper states that the “black box” nature of inverse design techniques has hindered the understanding of optimized inverse-designed structures and addresses this with an effective-potential picture for a nonlinear silicon cavity (Jia et al., 2023). This criticism is not a rejection of inverse-designed cavities; rather, it marks a methodological debate over whether optimized geometries should be explained in terms of envelope functions, local band-edge frequencies, or mode-overlap sensitivity maps after convergence.

4. Representative cavity architectures

The current literature supports a broad taxonomy of inverse-designed optical cavities.

Platform	Inverse-design target	Reported result
Dielectric meta-cavity	Near-field optical chirality at the center under linearly polarized excitation	$p\equiv\{\delta x_i,\delta y_i\}$3 in simulation; $p\equiv\{\delta x_i,\delta y_i\}$4 in experiment (Romashkina et al., 22 Dec 2025)
Release-free optomechanical crystal	Maximize $p\equiv\{\delta x_i,\delta y_i\}$5	$p\equiv\{\delta x_i,\delta y_i\}$6–$p\equiv\{\delta x_i,\delta y_i\}$7; $p\equiv\{\delta x_i,\delta y_i\}$8 (Hambraeus et al., 5 May 2026)
SiN L3 photonic-crystal cavity	Simultaneous targets for $p\equiv\{\delta x_i,\delta y_i\}$9 and $i=1\ldots N$0	$i=1\ldots N$1-fold coupling-efficiency improvement and $i=1\ldots N$2-fold quality-factor improvement vs standard L3 (Vij et al., 20 Sep 2025)
Thick-SiN Fabry–Pérot cavity	Loaded finesse and multi-wavelength mirror reflectivity	$i=1\ldots N$3; mirror reflectivity peak $i=1\ldots N$4 (Bi et al., 19 May 2025)
GaN hollow nanowire cavity	Maximize normalized OAM order $i=1\ldots N$5	$i=1\ldots N$6; mode purity $i=1\ldots N$7; $i=1\ldots N$8 (Lim et al., 12 Dec 2025)

The dielectric meta-cavity is a particularly explicit example of a cavity emerging from inverse-designed lateral patterning. It consists of an inverse-designed silicon film on sapphire with interlocking lobes around a central gap of approximately $i=1\ldots N$9, where co-localized electric and magnetic hotspots yield a “meta-cavity.” Under $N=16\times 16$0-polarized incidence, the free-space wave couples into a bound resonance of the patterned film; low material loss in Si enables high-$N=16\times 16$1 resonances, and the sub-wavelength gap yields small mode volume (Romashkina et al., 22 Dec 2025).

Fabry–Pérot implementations show a different design logic. In thick SiN, each mirror occupies an $N=16\times 16$2 design cell adjoining a $N=16\times 16$3 intracavity waveguide and a $N=16\times 16$4 bus waveguide, and the free-form mirror simultaneously tapers the fundamental $N=16\times 16$5 mode, realizes a near-unity broadband reflection band, and minimizes scattering into radiation modes (Bi et al., 19 May 2025). In silicon carbide, two inverse-designed reflectors of footprint $N=16\times 16$6 define a Fabry–Pérot cavity with target reflectivity, phase, and anomalous dispersion across roughly $N=16\times 16$7–$N=16\times 16$8 (Yang et al., 2023).

Cavity inverse design also appears in photonic-crystal and ring-like resonators. Visible-wavelength SiN L3 cavities modify only the $N=16\times 16$9 holes nearest the defect, preserving $|\delta x_i|,|\delta y_i|\le 0.1a$0 symmetry while engineering both quality factor and far-field numerical aperture (Vij et al., 20 Sep 2025). A circular ring cavity optimized by guided-mode expansion and automatic differentiation breaks periodicity relative to a standard bullseye, yielding a non-periodic concentric-ring design that preserves a Gaussian-like far field while increasing $|\delta x_i|,|\delta y_i|\le 0.1a$1 to approximately $|\delta x_i|,|\delta y_i|\le 0.1a$2 (Das et al., 15 May 2026). Whispering-gallery systems extend the same logic to beam shaping and polarization control in sub-micron nanolasers (Diez et al., 2022).

5. Experimental realizations and measured performance

Experimental validation is a defining feature of the field. The dielectric meta-cavity for enantioselective detection was fabricated by EBL on a $|\delta x_i|,|\delta y_i|\le 0.1a$3 Si-on-sapphire wafer, followed by RIE pattern transfer and resist removal, with final thickness $|\delta x_i|,|\delta y_i|\le 0.1a$4 and pattern agreement within $|\delta x_i|,|\delta y_i|\le 0.1a$5. Transmission spectroscopy under H/V/R/L polarization gave an observed resonant CD up to $|\delta x_i|,|\delta y_i|\le 0.1a$6 at $|\delta x_i|,|\delta y_i|\le 0.1a$7, a measured chirality enhancement of approximately $|\delta x_i|,|\delta y_i|\le 0.1a$8 relative to circular plane-wave $|\delta x_i|,|\delta y_i|\le 0.1a$9, and an experimentally extracted $\rho(x)\in[0,1]$0 with FWHM $\rho(x)\in[0,1]$1 (Romashkina et al., 22 Dec 2025).

The release-free silicon optomechanical crystal likewise moved from inverse-designed simulation to experiment. At room temperature, the measured quantities were $\rho(x)\in[0,1]$2, intrinsic mechanical linewidth $\rho(x)\in[0,1]$3, vacuum coupling $\rho(x)\in[0,1]$4 on the blue side to $\rho(x)\in[0,1]$5 on the red side, and optomechanical scattering rate $\rho(x)\in[0,1]$6. The device retained the thermal robustness of a fully clamped geometry and enabled intracavity photon numbers up to approximately $\rho(x)\in[0,1]$7 without thermo-optic drift (Hambraeus et al., 5 May 2026).

Visible SiN L3 cavities demonstrate multi-objective optimization in a distinctly cavity-radiation context. Three inverse-designed cavities, G1, G2, and G3, were fabricated in silicon nitride; photoluminescence measurements confirmed experimental control of the far-field numerical aperture. The reported outcomes include a $\rho(x)\in[0,1]$8-fold improvement in coupling efficiency and a $\rho(x)\in[0,1]$9-fold improvement in quality factor relative to the standard L3 cavity, and the statistical sample of $\{w_n\}$0–$\{w_n\}$1 copies per design retained the enhancement trends despite nanofabrication imperfections (Vij et al., 20 Sep 2025).

Fabry–Pérot cavity realizations show that inverse-designed reflectors can support high loaded finesse in foundry-compatible platforms. In thick silicon nitride, after $\{w_n\}$2 finite-difference frequency-domain iterations, single-mirror reflectivity exceeded $\{w_n\}$3 at four design wavelengths; final 3D FDTD refinement yielded average mirror reflectivity greater than $\{w_n\}$4 over a $\{w_n\}$5 bandwidth around $\{w_n\}$6, a peak of $\{w_n\}$7, loaded finesse up to $\{w_n\}$8, and loaded linewidth $\{w_n\}$9 at $2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$00, giving $2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$01 (Bi et al., 19 May 2025). In SiC Fabry–Pérot cavities, inverse-designed reflectors delivered average reflectivity $2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$02 over $2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$03–$2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$04, measured integrated dispersion with $2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$05, and loaded quality factors on the order of $2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$06, enabling both $2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$07 Kerr comb generation and $2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$08 visible-telecom frequency conversion (Yang et al., 2023).

Nonlinear, quantum, and structured-light cavities also have direct measurements. The interpretable silicon-on-insulator cavity for spontaneous four-wave mixing generated photon pairs at $2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$09 with a coincidence-to-accidental ratio of $2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$10 and loaded cavity $2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$11 values of approximately $2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$12, $2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$13, and $2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$14 (Jia et al., 2023). The inverse-designed GaN hollow nanowire cavity realized a mode with $2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$15, mode purity of about $2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$16, and $2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$17 of approximately $2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$18 (Lim et al., 12 Dec 2025). Inverse-designed whispering-gallery nanolasers experimentally produced azimuthal, radial, and linearly polarized beams with measured overlaps of $2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$19, $2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$20, and $2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$21, respectively, and $2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$22-factors near threshold of approximately $2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$23, $2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$24, and $2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$25 (Diez et al., 2022).

6. Applications, misconceptions, and current limitations

The application space is unusually broad. Inverse-designed cavities have been used for chiral spectroscopy with ultra-compact devices (Romashkina et al., 22 Dec 2025), fast and low-noise classical and quantum optomechanics (Hambraeus et al., 5 May 2026), low-power photonics and quantum information (Vij et al., 20 Sep 2025), on-chip nonlinear optics and quantum light generation (Jia et al., 2023, Yang et al., 2023), tailored-beam nanolasers (Diez et al., 2022), polarization-encoded spin-photon interfaces (Das et al., 15 May 2026), analog optical computing with closed-loop metastructures (Cordaro et al., 2022), and photonic hardware for optical multiplication and dot-product engines (Mathur, 18 Jul 2025).

A common misconception is that inverse-designed optical cavities are intrinsically uninterpretable black boxes. The literature does document the “black box” criticism, but it also shows several routes to physical interpretation: an effective-potential picture for nonlinear cavity phase matching (Jia et al., 2023), explicit decomposition of optomechanical coupling into moving-boundary and photoelastic terms (Hambraeus et al., 5 May 2026), and non-Hermitian perturbation analysis for OAM generation in a hollow nanowire cavity near an exceptional-point-like regime (Lim et al., 12 Dec 2025). This suggests that inverse design and physical insight are not mutually exclusive, although the interpretability burden often shifts from the initial parameterization to post-optimization modal analysis.

The main limitations are likewise well documented. Gradient-based workflows can become trapped in local minima and may require a good initial guess, as emphasized in the release-free optomechanical crystal where XHOPE provides a powerful physics-inspired starting point (Hambraeus et al., 5 May 2026). Realized performance can be limited by unmodeled loss channels, sidewall scattering, oxide-cladding non-uniformity, Akhiezer damping, sidewall roughness, and radius disorder rather than by the radiation-limited design target (Hambraeus et al., 5 May 2026, Jia et al., 2023, Vij et al., 20 Sep 2025). Even when a design is robust, performance trade-offs persist: in circular cavities, the innermost ring width and central disk radius strongly affect both $2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$26 and collection efficiency under $2.6\,\mu\mathrm{m}\times 2.6\,\mu\mathrm{m}$27 fabrication variation (Das et al., 15 May 2026); in whispering-gallery nanolasers, enhanced vertical outcoupling produces a threshold penalty relative to a conventional disk (Diez et al., 2022).

A plausible implication is that the mature form of the inverse-designed optical cavity is not a single algorithm or cavity family but a fabrication-aware, multi-objective framework that couples differentiable electromagnetics with application-specific observables. The most explicit statement of this generality appears in the dielectric meta-cavity work, where the two-stage workflow is said to apply to any target figure of merit, including Purcell factor, field intensity at a point, and multi-objective combinations, and can be adapted to design single-mode lasers, modulators, sensors, or nonlinear meta-cavities by exchanging the objective function and boundary conditions (Romashkina et al., 22 Dec 2025).