Miniature computational spectrometer with a plasmonic nanoparticles-in-cavity microfilter array

Optical spectrometers are crucial for analyzing light-matter interactions, but conventional designs are often bulky and complex. Miniaturization efforts face a trade-off between size reduction and performance. This research addresses this challenge by developing a computational spectrometer that uses a microfilter array to sample light, with the spectrum reconstructed computationally. Traditional spectrometers use dispersive elements like diffraction gratings, leading to large size and slow response times. Alternatives like Fourier transform spectrometers (FTSs) using Michelson interferometers also suffer from bulkiness and slow response. Computational spectrometers offer a solution by using wavelength-selective components and reconstruction algorithms. CMOS/CCD image sensors are attractive for miniaturized spectrometers due to their compact size, robustness, and ability to integrate numerous photosensors. High resolution is achievable by increasing the number of spectrum-disparate filters in the array. Various microfilter technologies have been explored, including quantum dots, photonic crystals, plasmonic encoders, metamaterials, and multilayer film filters, each with limitations in scalability and cost. This paper proposes a novel approach leveraging a plasmonic nanoparticles-in-cavity microfilter array and artificial intelligence (AI) for high-performance miniature spectroscopy.

Existing miniaturized spectrometers employ diverse technologies to achieve compact size and spectral resolution. These include thin-film optical filters, perovskite films, single nanowires, superconducting nanowires, tunable van der Waals junctions, folded digital metalenses, integrated photonic chips, and wavelength-selective photodetectors. CMOS or CCD image sensor-based spectrometers are particularly promising due to their integration capabilities and compactness. However, creating microfilter arrays with many spectrum-disparate elements remains a challenge. Previous approaches, such as using quantum dots, photonic crystals, or metamaterials, often involve complex and expensive nanofabrication techniques, limiting scalability. Multilayer film filters offer flexibility but suffer from complex, multi-step fabrication. Plasmonic 2D chirped gratings have been used, but their low transmittance restricts operation to reflection mode. While metamaterials and metasurfaces show promise in achieving sub-nanometer resolution, their fabrication processes remain time-consuming and costly.

This research introduces an AI-powered miniature spectrometer based on a scalable plasmonic nanoparticles-in-cavity microfilter array. The device samples incident light using the array, and a digital image sensor captures the resulting light pattern. The spectrum is then reconstructed using a machine learning algorithm. The plasmonic nanoparticles-in-cavity structure harnesses the strong coupling between Fabry-Pérot (FP) resonance and Mie resonance of silver nanoparticles (AgNPs). The size of AgNPs and the length of the FP cavity are precisely controlled to achieve spectral diversity. A digital ultraviolet (UV) exposure-based fabrication method is employed to directly print size-controlled AgNPs and length-varying FP cavities. This method avoids expensive nanofabrication techniques, ensuring scalability. The fabrication involves five steps: sputtering deposition of silver and SiO2 layers; sol-gel spin coating of TiO2; direct printing of AgNPs; grayscale patterning and nanoscale thickness tuning of polymer FP cavities; and sputtering deposition of a top silver and SiO2 layer. Precision photoreduction technology, modified to address challenges associated with UV light exposure through an opaque substrate, is utilized for AgNP printing. Grayscale photopolymerization technology, based on oxygen inhibition of free-radical polymerization, is used to achieve nanoscale thickness tuning of the polymer FP cavities. Both fabrication techniques utilize a digital micromirror array (DMD) for fast, parallel UV exposure control. The machine learning algorithm reconstructs the spectrum from the CMOS image data. The algorithm is trained using single-narrow-peak light spectra from a tunable monochromator and solves an optimization problem that minimizes a loss function incorporating least squares, L1 and L2 regularization, total variation (TV), and quadratic variation (QV). The algorithm generalizes several existing techniques, enabling adaptation through hyperparameter tuning. The experimental setup includes a tunable monochromator, broadband light sources, a monochrome camera with a CMOS image sensor, and a commercial spectrometer for training and testing. The microfilter array is not directly attached to the CMOS sensor, but instead uses a lens to project the light pattern onto the sensor. Image data are pre-processed with pixel binning and data filtering to improve signal-to-noise ratio.

Numerical simulations demonstrated the strong coupling effects between FP resonance and Mie resonance, resulting in Rabi splitting and wavelength shifts. The fabrication process successfully produced a large-scale (1152 elements) plasmonic microfilter array with high-density spectral peaks (2436 peaks). Measured transmission spectra confirmed the predicted spectral diversity, showing sharp transmission peaks varying in wavelength. Rabi splitting was observed in the measured reflection and transmission spectra, particularly in the short-wavelength region. Polarization dependence was minimal, indicating insensitivity to input light polarization. The machine learning algorithm effectively reconstructed spectra, accurately measuring peak wavelengths with a root-mean-square error (RMSE) of approximately 0.03 nm. The spectrometer achieved an average spectral resolution of 0.65 nm, successfully resolving adjacent spectral peaks with a wavelength separation of 0.8 nm. It also accurately reconstructed broadband spectra with different profiles, demonstrating its ability to handle diverse input light spectra. Analysis of the number of microfilters showed a significant enhancement in resolution up to ~500 filters, with diminishing returns thereafter. A comparison of various reconstruction algorithms showed that a hybrid approach, combining aspects of LASSO, ridge regression, and TV regularization, yielded the best performance in terms of both peak wavelength accuracy and spectral resolution. Testing with intentionally noisy data demonstrated the expected sensitivity of the system's resolution to noise levels.

The developed miniature spectrometer successfully addresses the challenges of miniaturizing spectrometers while maintaining high resolution. The use of a plasmonic nanoparticles-in-cavity microfilter array, fabricated using a scalable digital UV lithography method, provides a significant advantage over previous approaches that rely on complex and costly nanofabrication. The machine learning algorithm, with its flexibility in hyperparameter tuning, proves effective in reconstructing spectra from the CMOS image data. The achieved spectral resolution of 0.65 nm demonstrates significant improvement compared to previous miniature spectrometers. The high accuracy in peak wavelength measurement and the ability to handle both narrowband and broadband spectra further validate the effectiveness of this approach. The sensitivity to noise highlights a key area for future improvement, which could involve further refinement of the fabrication process to enhance uniformity and the development of more robust reconstruction algorithms.

This work presents a highly scalable, AI-powered miniature spectrometer with sub-nanometer resolution. The novel plasmonic nanoparticles-in-cavity microfilter array and the robust machine learning algorithm enable accurate and high-resolution spectral measurements across the visible range. Future work should focus on improving fabrication techniques to enhance the uniformity of the microfilter array and developing more noise-resistant algorithms. Exploring alternative materials and optimizing the CMOS image sensor could further improve performance. This technology has significant potential for applications such as portable medical diagnostics and spectral imaging.

The current spectrometer exhibits sensitivity to noise, as evidenced by the observed impact of noise levels on the achievable spectral resolution. The performance of the spectrometer in measuring narrow spectral peaks in a smartphone-based prototype needs further improvement. The optimization of hyperparameters in the machine learning model requires careful consideration, and finding optimal values might be computationally intensive. Further improvements to the fabrication process, such as increasing the uniformity of the AgNPs, and the development of more robust and noise-resistant spectral reconstruction algorithms are required to fully realize the potential of this technology.