Nanometer-scale photon confinement in topology-optimized dielectric cavities

The study addresses whether dielectric nanocavities can confine light deep below the diffraction limit while maintaining practical quality factors and compatibility with semiconductor devices. Prior dielectric cavity approaches have achieved high Q but not deep subwavelength mode volumes, whereas plasmonic systems offer subdiffraction confinement at the cost of severe absorption losses and low Q. The authors propose a fundamentally different route: fabrication-constrained topology optimization (inverse design) of dielectric bowtie cavities to maximize local density of optical states inside the material, not at surfaces, thereby enhancing light–matter interaction in compact silicon membranes. The purpose is to realize and experimentally validate a realistic, fabricable design that confines photons within nanometer-scale silicon bridges, overcoming limitations of intuition-driven designs and numerical artifacts associated with lightning-rod effects. This has broad importance for nanophotonic devices requiring strong interaction over usable bandwidths, including nonlinear optics and quantum photonics.

Multiple cavity mechanisms exist—distributed Bragg reflection, total internal reflection, Fano resonances, bound states in the continuum, and topological confinement—which improved temporal confinement but generally do not push mode volumes into the deep subwavelength regime. Plasmonic nanostructures achieve subdiffraction confinement but are limited by metal absorption, keeping Q well below 100. Earlier inverse-designed dielectric bowtie cavities suggested reduced mode volume and ring-grating-induced Q increases, yet unconstrained inverse design often yields unrealizable geometries. Previous experimental claims of subdiffraction dielectric confinement were undermined by numerical inconsistencies and reliance on lightning-rod effects at sharp tips that do not yield confined modes inside the dielectric. The authors build on theoretical advances in topology optimization for photonics and recent recognition that realistic fabrication constraints must be embedded in inverse design to avoid non-fabricable outcomes. They also adopt a robust mode-volume evaluation at the cavity center to avoid divergences at tips and corners that make max-normalized definitions unreliable.

Design and simulation: The team performs size- and tolerance-constrained topology optimization to maximize the projected local density of optical states (LDOS) at the geometric center of a silicon membrane cavity, forcing the center to be solid to avoid surface lightning-rod artifacts. Physics is modeled via time-harmonic Maxwell’s equations in a finite domain, exploiting three-fold spatial symmetry and first-order absorbing boundary conditions. The finite-element method with first-order Nedelec elements is used. Topology optimization is set up as a continuous constrained optimization with one design variable per finite element in the selected 2D (x–y) design domain (membrane thickness fixed at 240 nm and variables linked along z). Filtering and thresholding regularize the design; a material interpolation scheme links design variables to permittivity. The globally convergent method of moving asymptotes solves the optimization. The design targets a resonance near λ ≈ 1550 nm and maximizes the LDOS at the specified center point. Fabrication constraints: Constraints are experimentally measured and embedded in the optimizer as minimum attainable radii of curvature: solid features r ≥ 10 nm, void features r ≥ 22 nm, and at the cavity center r ≥ 4 nm to realize an average bridge width of ~8 nm. Devices are realized in 240 nm-thick crystalline (100) silicon membranes (n = 3.48), patterned by 100 keV electron-beam lithography in chemically semi-amplified resist (CSAR), followed by a low-power cyclic reactive-ion etch (modified CORE sequence using SF6 etch and O2 passivation) and selective vapor-phase HF undercut. The etch process is optimized to minimize critical dimensions while tolerating periodic sidewall scallops. Proximity effects are mitigated through layout spacing and local mask corrections. Geometry tuning and SEM metrology: To precisely assess bridge widths near resolution limits of SEM, three device sets are fabricated under global geometry-tuning δ, shrinking exposed (air) areas in steps of 2 nm. Six nominally identical copies per δ are produced. Representative 40°-tilted SEM images verify high-fidelity replication and systematic variations. Mean bowtie bridge widths measured: 8 nm, 10 nm, and 16 nm for the respective δ sets. Electromagnetic analysis: Quasi-normal modes are computed for the full structure (including suspension tethers). An effective mode volume V and Q are extracted, with V evaluated robustly at the cavity center r0 to avoid tip-induced divergences. Simulations yield V ≈ 0.08 × (λ/(2n))^3 and Q ≈ 1100 around λ ≈ 1551 nm. Far-field spectroscopy: Confocal cross-polarized reflection microscopy records broadband spectra. The DBC resonance interferes with a low-Q background (air-gap vertical mode) producing a Fano lineshape. Local fits to the Fano model provide resonance wavelength λ0 and Q for all devices. Resonance shifts across δ quantify fabrication reproducibility. Near-field spectroscopy and imaging: Scattering-type scanning near-field optical microscopy (s-SNOM) with pseudo-heterodyne detection probes the cavity immediately above the surface, strongly suppressing far-field background. A CW laser is scanned, and the AFM silicon tip (nominal 10 nm radius) oscillates while mapping amplitude near ~5 nm above the surface. Frequency-domain Lorentzian fits give λ0 and loaded Q (tip-induced loss). Spatial scans on resonance map the mode hotspot. To compare with simulations, the tip is modeled as a polarizable scatterer, and an instrument function f(σ), taken as a Gaussian with standard deviation σ, is convolved with the calculated field amplitude above the device. Overlap is quantified via the Bhattacharyya coefficient τ to determine σ and assess agreement. Regions where the tip falls into voids are excluded from analysis to focus on surface-proximal fields.

• Realized a compact silicon dielectric bowtie cavity (DBC) with simulated mode volume V ≈ 3 × 10^−4 λ^3 (equivalently ≈ 0.08 × (λ/(2n))^3) and quality factor Q ≈ 1100 at telecom wavelengths (λ ≈ 1550 nm), within a footprint ~4 λ^2.
• Fabricated structures confine photons inside silicon bridges as narrow as 8 nm with aspect ratios ~30, achieved by embedding measured fabrication constraints directly into the inverse design.
• Far-field confocal reflection: Fano-resonant cavity features observed near 1520–1550 nm. Across three δ steps (−2 nm increments), the mean spectral shift is Δλ = 40.4 ± 0.6 nm per step; device-to-device standard deviation within each set ≤ 4 ± 0.6 nm. Six nominally identical copies per δ are consistent within |δ| ≤ 0.2 nm. Q values from far-field fits are consistent with designed moderate Q (~10^3).
• SEM-derived mean bridge widths for the three geometry-tuned sets: 8 nm, 10 nm, and 16 nm.
• Near-field s-SNOM: At ~5 nm height above the surface, measured resonance λ0 = 1489.4 ± 0.1 nm with loaded Q = 370 ± 40 (reduced by tip-induced loading). Spatial maps show a single localized hotspot with excellent background suppression.
• Instrument-function analysis: Best agreement between measured and simulated fields for a Gaussian instrument function with σ = 37 ± 5 nm (FWHM ≈ 87 nm), yielding τ = 0.984 overlap for δ = −4 nm device and τ = 0.991 for δ = −6 nm device; cross-measurement overlap τ = 0.996. The instrument width exceeds the mode size, so measurements set an upper bound indicating subdiffraction confinement (mode volume below (λ/(2n))^3) even after convolution.
• Purcell factor: The directly optimized LDOS corresponds to an estimated Purcell enhancement ~6 × 10^3 over up to ~2 nm bandwidth, beneficial for nonlinear optics and broadband-emitter applications.

The work addresses the central challenge of achieving deep subdiffraction confinement of light inside a dielectric while retaining practical Q and fabricability. By incorporating measured fabrication limits (minimum radii/critical dimensions) into topology optimization, the authors obtain designs that are both realistic and high-performing. Simulations predict V ≪ (λ/(2n))^3 with Q ~ 10^3, and experiments corroborate subdiffraction localization through near-field mapping, despite the instrument’s finite spatial response, establishing an upper bound on the mode size below the diffraction limit. The results resolve shortcomings of prior demonstrations that relied on lightning-rod surface effects and ambiguous mode-volume definitions by evaluating V at the cavity center and showing global confinement within the dielectric material. The significance is substantial: strong light–matter interaction without extreme Q enables broader bandwidth operation (useful for short pulses, nonlinear effects, and broadband quantum emitters), while extending the design domain could increase Q further without sacrificing the ultra-small V. The reported Purcell bandwidth and enhancement indicate potential advances in quantum photonics (e.g., with erbium dopants in silicon) and integrated nanophotonic devices.

This study introduces a fabrication-constrained topology-optimization framework that yields silicon dielectric bowtie cavities with experimentally verified subdiffraction photon confinement inside the material. The cavities exhibit ultra-small mode volumes (≈ 0.08 × (λ/(2n))^3) with moderate Q (~1100) and are fabricated with high fidelity, including 8 nm bridges at high aspect ratios. Far-field and near-field measurements, combined with rigorous modeling of the instrument response, confirm strong mode localization and reproducible device performance. The approach sets a new paradigm for photonic inverse design, demonstrating that measured process constraints can be intertwined with optimization to produce realizable, high-performance nanophotonic structures. Future directions include enlarging the design domain to boost Q while maintaining ultra-small V, integrating suitable narrow-linewidth emitters (e.g., erbium in silicon) for exploring extreme Purcell regimes, reducing sidewall roughness via improved masks, and generalizing the methodology to other nanotechnologies (optomechanics, nanoelectromechanics, quantum photonics).

• Precise experimental determination of the absolute mode size is limited by the near-field probe’s finite instrument function; only an upper bound to V is obtained.
• s-SNOM measurements introduce loading, reducing the observed Q compared to the intrinsic cavity Q.
• The design is highly sensitive to nanoscale dimensions (e.g., bridge width and void geometry), placing stringent demands on fabrication control and metrology; SEM resolution challenges persist near a few-nanometer scale.
• Numerical artifacts (e.g., field divergences at sharp tips) necessitate careful definitions of mode volume and robust evaluation at the cavity center; small geometric discretization errors can affect simulations if not controlled.
• The optimized structure likely represents a local optimum within a vast design space; global optimality cannot be guaranteed.
• The compact design domain used here trades off extremely small V against moderate Q; achieving much higher Q would require larger footprints and potentially more complex fabrication steps.