Broadband on-chip spectral sensing via directly integrated narrowband plasmonic filters for computational multispectral imaging

2.1 Narrowband near – infrared plasmonic bandpass transmission filter

For on-chip spectroscopy, a high-efficiency narrowband bandpass filter operating across a broad wavelength range with low background has long been pursued as a pivotal component. We developed a novel plasmonic bandpass filter with simple architecture and direct detector integrability, comprising a dielectric nanohole array perforated in a thin metal film (Figure 2a). The resonant response arises from surface plasmon polariton excitation by the periodic nanohole array at the metal–dielectric interface [30], [40]. Shallow nanoholes (H < 80 nm) suppress out-of-plane radiative leakage, yielding a narrow transmission linewidth (FWHM < 10 nm). The optical response is primarily governed by three structural parameters: (i) Etching depth (H), which influences mode confinement and propagation loss; (ii) Au film thickness (t_Au), which affects coupling efficiency and Ohmic dissipation; and (iii) Filling factor (FF = d/P, where d is the nanohole diameter and P is the period), which regulates inter-hole coupling and lattice diffraction. As quantitatively shown in Figures S1–3, these parameters collectively modulate the balance between resonant enhancement and intrinsic losses in the Au/SiN _x heterostructure, enabling wavelength-specific narrowband filtering across 900–1,700 nm. As shown in Figure 2b, a transmittance exceeding 55 % can be maintained while achieving an FWHM as narrow as 5 nm. Furthermore, the central wavelength of the bandpass filter can be precisely tuned by varying the nanohole periodicity, enabling wafer-scale fabrication of filter arrays through a single lithographic step.

Figure 2:

Period tunable narrowband near – infrared plasmonic bandpass transmission filters. (a) Schematic cross-section of the Au/SiN _x nanohole array structure. (b) Simulated transmission spectra for the device structure in (a). The nanohole period P = 1,500 nm, the depth H = 70 nm, the Au film thickness t_Au = 25 nm, and the fill factor FF = 0.7. Electric field distributions of device at characteristic wavelengths, I. xy-plane field intensity at 1,305 nm resonance wavelength (aperture center plane); II. xz-plane cross-section at 1,305 nm, illustrating the SPP resonance at the Au/SiN _x boundary within the nanohole array; III. xy-plane distribution at 1,500 nm off-resonance wavelength (>80 % intensity reduction); IV. xz-plane profile at 1,500 nm, showing the absence of boundary electric field enhancement under off-resonance conditions. (c) Calculated transmission spectra of filter arrays at H = 70 nm, t_Au = 30 nm and FF = 0.7. Chip is a 16-unit array with P = 1,050, 1,106, …, 1,950 nm. (d) The transmission spectra of the 16-period nanohole of the coefficient correlation matrix.

Building upon the optimized geometric parameters (H = 70 nm, t_Au = 25 nm, FF = 0.7, P = 1,500 nm), we numerically validated the plasmonic resonance filter via full-field electromagnetic simulations. Figure 2b demonstrates exceptional narrowband filtering across a 900–1,700 nm operational window: >55 % peak transmittance at resonance Ultra-narrow bandwidth (FWHM ≈ 5 nm, Q ≈ 130) Strong out-of-band rejection (<10 % transmittance at off-resonance wavelengths) Field confinement analysis reveals the physical origin: at resonance (1,310 nm), surface plasmon polaritons (SPPs) couple to Au-dielectric interfaces within the nanoholes, generating intense near-field enhancement (|E|/|E₀| > 5.8). Conversely, off-resonance fields (1,500 nm) exhibit suppressed transmission (<10 %), confirming effective bandpass functionality. To extend spectral coverage, we engineered nanohole arrays with linearly scaled periods (P = 1.05–1.95 μm), achieving 16 discrete channels that quasi-continuously cover 900–1,700 nm (Figure 2c). This design enables computational spectrometry by constructing a high-dimensional spectral transfer matrix. Crucially, we evaluate the matrix quality through cross-correlation [2] analysis:

where f _i (λ) denotes the transmission spectrum of the i-th (or j-th) filter, and C _ij quantifies the spectral orthogonality between channels, which forms the basis for high-accuracy spectral reconstruction. As quantified by the correlation matrix in Figure 3d, the filters exhibit ultra-low mutual correlation with a mean cross-correlation coefficient of 0.25.

Figure 3:

Spectral reconstruction comparisons to the reference spectrum. (a) Reconstructed target narrowband spectra (dotted lines, FWHM = 5 nm) using calculated filter responses as shown in Figure 3c. The black lines are the reference spectra. (b) Reconstructed single Gaussian peak target spectra (dotted lines, FWHM = 260 nm); (c) reconstructed two different half-width combinations target spectra (dotted lines, FWHM = 107/132 nm); (d) reconstructed target broadband spectra. (e) Transmission spectra generated by sixteen filter units with lattice periods ranging from 1.76 to 1.81 µm, covering the spectral range of 1,520–1,560 nm. (f) Reconstruction of a dual-peak signal centered at 1,550 nm with a 1 nm wavelength separation. The dashed box highlights the magnified view of the dual-peak region.

Infrared spectroscopy provides critical molecular fingerprint identification for portable precision matter analysis. While rich in chemical information, conventional infrared spectrometers suffer from bulkiness, complexity, and high cost. Emerging Computational Spectroscopy chip-scale alternatives offer compelling advantages: ultracompact footprint, simplified optics, and high accuracy. Crucially, spectrometer performance hinges on two parameters: operational bandwidth and spectral resolution, both fundamentally governed by the filtering subsystem. Here, we introduce a plasmonic nanohole array serving as an integrated tunable filter that simultaneously delivers: Narrow linewidth (FWHM ≈ 5 nm, Q_max ≈ 284), broadband coverage (900–1,700 nm), high out-of-band rejection (>90 % suppression), enabling miniaturized computational spectrometers (Figure 2c).

Conventional grating spectrometers rely on spatially separated dispersive elements and long optical paths, with their resolution fundamentally constrained by detector pixel density. In contrast, our design integrates a periodically tunable transmission filter directly with the photodetector, thereby overcoming this limitation. This integration significantly reduces system complexity and footprint by eliminating bulky optical components. When paired with computational reconstruction, the architecture overcomes conventional resolution limits imposed by detector geometry.

Spectral reconstruction methodology:

For N = 16 filter channels spanning 900–1,700 nm (Figure 2c), the detector response is modeled as:

Here, L(λ) represents the target spectrum within the 900–1,700 nm range, while f _i (λ) denotes the transmission response of the i-th plasmonic filter. The output signal from each detector corresponds to the convolution of L(λ) with its respective filter response. To construct the full sensing matrix, we precisely characterized f _i (λ) using a wavelength-swept supercontinuum laser, scanning from 900 to 1,700 nm in 1-nm increments and recording each filter’s photoresponse. Recovering L(λ) from this under sampled dataset constitutes an ill-posed inverse problem, requiring advanced reconstruction techniques. To mitigate the influence of noise measurement, we apply L₂-norm Tikhonov regularization for robust spectral reconstruction from undersampled data [22], [41]. The modified Tikhonov formulation employed in our work is given by:

where α is the regularization weight value, and Γ is the auxiliary matrix. As demonstrated in Figure 3a–d, the recovered spectra closely match the reference in both peak location and general profile, with only minor deviations. To quantitatively assess reconstruction accuracy, we use spectral fidelity, defined as:

where x _i and x ̂ i are the original and reconstructed spectral values at the i-th wavelength point, respectively, and N is the total number of data points. Fidelity values range from 0 to 1, with values approaching 1 indicating high reconstruction accuracy. Across multiple test spectra, our platform consistently achieves fidelity values exceeding 0.99, underscoring the high precision of the proposed spectroscopic architecture.

To evaluate the spectral resolution of our approach, we reconstructed a synthetic spectrum containing two closely spaced narrowband peaks (Δλ = 1 nm) using the single-resonance response of our plasmonic filter structure, as shown in Figure 3f. Sixteen filter units, with periods ranging from 1.76 to 1.81 µm, were employed to generate transmission spectra spanning 1,520–1,560 nm (Figure 3e). Applying our reconstruction framework, we successfully resolved the characteristic double-peak profile using only 16 measurements, demonstrating the system’s high resolution. These results highlight the strong spectral discrimination capability of our plasmonic filter–InGaAs detector integrated chip, underscoring its potential for compact, on-site spectroscopic applications such as on-site material rapid inspection and point-of-care diagnostics.

Building on the high-fidelity spectral reconstruction framework used in Figure 3, we further extend the plasmonic filter–detector integrated chip toward on-chip multispectral imaging, enabling spatially resolved spectral acquisition within a compact platform. This architecture uniquely combines: Monolithic integration of tunable nanohole filters with InGaAs photodiodes, computational multiplexing enabled by the 16-channel filters, snapshot spatial-spectral capture without mechanical scanning. Hyperspectral dataset validation: As shown in Figure 4a, a hyperspectral image consisting of 463 × 241 pixels across 76 spectral bands (900–1,700 nm) was reconstructed based on the HYPERION dataset acquired over Jiangsu Province, China, on May 3, 2013 [42]. This dataset, primarily intended for monitoring fluvial material variations, demands highly accurate and faithful spectral image reconstruction. Using the 16 filter channels described in Figure 2c, we performed pixel-wise spectral reconstruction across the full dataset. Figure 4b–d present representative spectra at three distinct locations, where the reconstructed results closely match the reference spectra, achieving a spectral fidelity greater than 0.99. In addition, spectral images at eight representative wavelengths were visualized, demonstrating excellent boundary fidelity of the river morphology and clearly delineated shoreline features (Figure 4e). Furthermore, silicon-based hot-carrier photodetectors offer a promising alternative to InGaAs detectors in the near-infrared range, substantially reducing cost and improving compatibility with standard semiconductor fabrication processes. In addition, the filter design exhibits strong spectral scalability: by tuning the structural period P, the operational band can be readily reconfigured to target specific wavelength ranges. Together, these results establish our chip-based platform as a compact and scalable solution for next-generation airborne or field-deployable imaging spectrometers.

Figure 4:

Hyperspectral imaging using the narrowband near-infrared plasmonic bandpass transmission filter spectrometer. (a) Hyperspectral data set of river image on May 3rd, 2013 [42]. (b–d) Recovered reflectance spectra (dotted lines) of three selected patches in the test multispectral image Figure 4a, the black lines are the reference spectra; (e) a series of reconstructed images at selected wavelengths. The intensity range of these images is normalized.