Plasmonic ommatidia for lensless compound-eye vision

Traditional lens-based cameras face a fundamental tradeoff between compact size and wide field-of-view due to aberrations at large incidence angles. In nature, compound eyes in insects and crustaceans achieve nearly hemispherical vision with compact, multi-ommatidia architectures that offer infinite depth of field and high motion sensitivity. These attributes motivate compound-eye-inspired cameras for applications like endoscopy, surveillance, wearables, swallowable cameras, and autonomous navigation. Prior artificial compound-eye systems typically combine microlens arrays with image sensors in either curved or planar geometries. Curved implementations closely mimic biology but are incompatible with standard planar microelectronics, requiring complex fabrication, packaging, or bulky relays. Planar microlens approaches suffer from crosstalk and f-number-limited fields of view (<±35°). Other angle-sensitive techniques (stacked gratings, optical phased arrays, nanowire photodetectors) face limited FoV, complex deconvolution, or specialized functionalities. The research question addressed here is whether a fully planar, lensless architecture using metasurface-coated pixels can realize compound-eye-like directional sensitivity with wide FoV and enable high-quality image reconstruction via computational imaging. The authors propose integrating each pixel with a plasmonic metasurface that transmits only a small angular band of incident light while reflecting others, and combining many differently oriented/pitched pixels with computational reconstruction to achieve wide-angle, lensless compound-eye vision.

The paper reviews artificial compound-eye systems: planar and curved microlens arrays combined with sensors; curved arrays offer wide FoV but complicate CMOS compatibility, necessitating optical relays or complex fabrication on curved substrates. Planar microlens implementations suffer limited FoV (<±35°) and crosstalk unless blocking layers are used. Angle-sensitive photodetectors using stacked diffraction gratings yield sinusoidal angle dependence but require global deconvolution and have reported FoV <±50°. Optical phased-array receivers can steer sensitivity but require heterodyne mixing with a local oscillator. Two-nanowire photodetectors can sense angle for ranging/triangulation rather than full imaging. These limitations motivate a new planar approach with wide FoV, dense pixel integration, and computational reconstruction.

Metasurface design and operating principle: Each pixel’s active region (Ge MSM photoconductor) is coated with a composite plasmonic metasurface: a 100-nm Au film sandwiched between SiO2 layers (insulation and coupling control), with an array of rectangular Au nanoparticles (NPs) forming (i) a periodic grating coupler, (ii) a grating reflector, and (iii) subwavelength slits in the Au film. p-polarized incident light at the designed angle θp is diffracted by the grating coupler into surface plasmon polaritons (SPPs) on the top Au-air interface. SPPs propagate toward the slits, where extraordinary optical transmission-like coupling scatters them predominantly into the Ge substrate, generating photocurrent. Light at the opposite angle −θp excites SPPs directed to the grating reflector, which, via a linear phase-gradient NP array suppressing all diffraction orders except q = −1, scatters SPPs back to free space, preventing cross-pixel coupling. Angular selectivity is controlled mainly by the grating-coupler period Λ via the first-order SPP diffraction condition. NP number and width are optimized to maximize diffraction efficiency and avoid coupling to localized resonances.

Electromagnetic simulations: Designs are optimized with 2D FDTD (x–z plane) using Lumerical, computing transmission versus polar angle θ at λ0 = 1550 nm for p and s polarizations. Materials (Ge, SiO2, Au) use built-in complex permittivity. Full 3D angular response maps (θ, φ) are computed via a reciprocity-based far-field method: a dipole source placed in Ge below the slits radiates into air; by reciprocity, the far-field pattern is proportional to the metasurface’s angular transmission. The 3D-to-2D proportionality factor is calibrated by matching to 2D FDTD line cuts.

Device fabrication: Substrate: nominally undoped (100) Ge. Steps: (1) Deposit 60 nm SiO2 (rf sputter); open contact windows via photolithography and wet etch. (2) Define slits and deposit metal: HSQ/PMMA bilayer EBL patterns HSQ lines; RIE removes exposed PMMA; evaporate 5 nm Ti/100 nm Au; lift-off in acetone forms slits; pattern Au film and device electrodes by photolithography and wet etch. (3) Deposit top 60 nm SiO2; open wire-bond windows. (4) Fabricate grating coupler and reflector via single-layer PMMA EBL, evaporate 5 nm Ti/50 nm Au NPs, lift-off. (5) Deposit/pattern a Ti window covering the device with an opening over the metasurface to block spurious electrode-region photocurrent. Devices are mounted and wire-bonded.

Device characterization: Angle-resolved photocurrent measured on a custom optical goniometer with a 1550 nm diode laser. The device is DC biased; incident intensity modulated at 1 kHz; photocurrent read via bias-tee and lock-in. Typical bias 2 V, power 0.5 mW. Polarization-maintaining fiber, beam expander (5 and 10 mm lenses), half-wave plate, and iris control beam size and polarization. Polar angle θ varied ±85° in 1° steps by rotating the optics about the focal point aligned to the metasurface; azimuth φ varied 0–90° in 5° steps by rotating the sample; remaining quadrants filled by symmetry. Two orthogonal input polarizations (parallel/perpendicular to grating lines) recorded. Linear interpolation in θ used for smooth maps.

Image formation and reconstruction: The scene is assumed in the far field; each pixel integrates intensity weighted by its angular response. Formulate y = A x, where each pixel’s angular response forms a row of the sensing matrix A mapping object intensity x (discretized over incident angles) to measurements y. Reconstruction uses truncated SVD (TSVD): x̂ = Σ_{i=1..L} σ_i^{-1} u_i v_i^T y, with L chosen by manual tuning for best visual quality. Pixel angular responses for designed devices come from simulated maps (Fig. 2d–f and Supplementary Fig. 5); additional responses are interpolated for intermediate θp and φ to achieve uniform coverage. For experiments, measured maps (Fig. 3d–g, Supplementary Fig. 8) are used, with interpolation to cover ±65°.

Array design and sampling: Choose angular spacings Δθ = 1.5° and Δφ = 3° to ensure coverage and avoid fringe artifacts, yielding 6240 pixels for ±75° FoV (simulated devices) and 5280 pixels for ±65° (measured devices). Larger Δφ produces radial fringe artifacts; larger Δθ degrades resolution at high θ.

Noise and SNR modeling: Add white Gaussian noise to y. Baseline single-pixel SNR = 56 dB (y_signal/y_noise ≈ 631) achievable with ~8 µm CMOS pixels. Higher SNR by binning identical pixel designs: N = 5 and 50 give SNR ≈ 63 and 73 dB, with total pixel counts ≈260k (measured-map case) and ≈310k (simulated-map case).

Broadband illumination modeling: For finite bandwidth δλ around λ0 = 1550 nm, the main effect is peak broadening δθp ≈ (δλ/λ0)(n_spp + sin θp)/cos θp, with n_spp ≈ 1.06. Implement by 2D convolution of monochromatic angular responses with a Gaussian of width δθp during reconstruction. Bandpass filters of ~155 nm (10%) or ~77 nm (5%) bandwidth are considered.

Devices under test: MSM Ge photoconductors with electrode separation d = 300 µm; metasurface sections repeated 5–6 times between electrodes; reflector of one section adjacent to slits of next for measurement convenience.

• Simulated metasurfaces (λ0 = 1550 nm, p-pol): tunable peak detection angles θp across ±75°; angular selectivity FWHM 3–14° increasing with |θp|; peak transmission T_p ≈ 35–45%; peak-to-average background ≈ 6; s-pol transmission <0.2% at all angles. Full 2D angular responses show C-shaped high-transmission regions consistent with first-order SPP excitation; responses decrease with increasing azimuth φ due to reduced Ex components.
• Experimental devices (Ge MSM photoconductors) with metasurfaces targeting θp ≈ 0°, 12°, 28°, 65° exhibit highly directional photocurrent with C-shaped maps. Measured FWHM along θ = 0° line cuts: ~4–21° (increasing with θp). Peak-to-average background ≈ 3. Strong polarization dependence: xz (p-like) dominates; yz contribution negligible. SEM-measured periods and widths: Λ ≈ 1440, 1180, 1030, 775 nm; w ≈ 240, 560, 526, 256 nm for the four devices, respectively.
• Responsivity calibration vs bare sample: At θp = 12° and 65°, p-pol responsivity peaks are ~42% and ~36% of bare device (no Au/NPs), matching simulations. Absolute responsivity ~10 mA W−1 V−1 for devices with similar dark resistance (~1.5 kΩ), limited by large electrode separation (d = 300 µm).
• Imaging simulations: Using simulated angular responses with Δθ = 1.5°, Δφ = 3°, 6240 pixels achieve high-quality reconstructions over ±75° FoV at SNR = 56 dB (simple object) and SNR = 73 dB (complex object). Using measured maps with 5280 pixels over ±65° FoV also yields recognizable reconstructions; resolution somewhat reduced due to broader peaks and higher background. Binning to N = 5 and 50 boosts SNR to ~63 and 73 dB with total pixel counts ~260k–310k, feasible with CMOS.
• Broadband robustness: Accounting for angular peak broadening enables successful reconstructions under δλ/λ0 = 10% (simple object, SNR 56 dB) and 5% (complex object, SNR 73 dB), with only modest resolution loss. Implementable via a bandpass filter of 155 or 77 nm bandwidth at 1550 nm.

The work demonstrates that integrating angle-selective plasmonic metasurfaces with standard planar photodetectors yields pixel-level directional sensitivity analogous to biological ommatidia, enabling lensless compound-eye vision on flat substrates. The metasurface-mediated SPP coupling/scattering architecture allows each pixel to transmit light only from a designed angular sector while reflecting others, thus suppressing crosstalk and allowing dense packing without microlenses. Computational imaging (TSVD) combines many such pixels with overlapping angular responses to reconstruct wide-FoV images. Simulations and measurements confirm narrow angular selectivity and high peak transmission for p-polarized light, and successful image reconstructions over ±65° to ±75° FoV at realistic SNRs and pixel counts. This addresses the core challenge of achieving wide FoV without curved sensors or bulky optics, while maintaining compatibility with CMOS processes and enabling extreme miniaturization. The approach is extensible to larger FoV (approaching hemispherical) with design optimization, and can incorporate polarization independence, broader bandwidth, or multiwavelength operation via metasurface engineering. Computationally, more advanced priors, iterative inverse solvers, or machine learning could further enhance reconstruction quality and robustness.

The paper introduces a lensless, planar compound-eye camera concept using plasmonic metasurface-coated pixels that act as artificial ommatidia. It details design, fabrication, and characterization of near-infrared devices with tunable angle sensitivity and demonstrates, through simulations and measured responses, high-quality image reconstruction over wide fields of view (up to ±75° in design and ±65° with current prototypes) at practical SNRs. The approach eliminates microlenses, reduces thickness, increases pixel density, and maintains CMOS compatibility. Future directions include: extending FoV toward a full hemisphere; implementing dielectric metasurfaces on visible-range photodiodes for lower loss and CMOS foundry scalability; engineering polarization-independent and broadband responses; leveraging spectral diversity for color imaging; and adopting advanced computational reconstruction (iterative methods, machine learning) to improve image quality and enable additional functionalities such as polarization sensing and wavefront reconstruction.

• Polarization dependence: current devices are highly sensitive to xz (p-like) polarization and largely insensitive to yz (s-like), reducing sensitivity under unpolarized illumination; requires metasurface redesign for polarization-independent operation.
• Experimental performance gaps: measured angular selectivity is broader and background higher than simulations, likely due to surface roughness and fabrication deviations in NP dimensions/periods; this slightly degrades reconstruction resolution.
• Field of view: fabricated devices demonstrated up to ±65° θp; designs reach ±75°, but extending toward a full hemisphere requires further optimization and managing cosθ power fall-off at high θ (reduced intercepted power), potentially needing higher gain or larger active areas for high-θ pixels.
• Device variability: significant variation in dark resistance across samples affects responsivity, complicating quantitative comparison and indicating fabrication/process variability.
• Bandwidth sensitivity: diffractive nature introduces wavelength-dependent angular responses; while accounted for computationally for δλ/λ0 up to 5–10%, broader bandwidths would need dispersion-engineered metasurfaces.
• SNR-resolution trade-offs: higher SNR achieved via pixel binning increases total pixel count or reduces effective resolution/area efficiency; practical implementations must balance pixel pitch, well capacity, and binning.
• Current demonstrations at 1550 nm on Ge MSM photoconductors with large electrode spacing (300 µm) limit absolute responsivity and photoconductive gain; translation to compact, high-speed, visible-range photodiodes is future work.