Deep Learning-Based Single-Shot Computational Spectrometer Using Multilayer Thin Films

Abstract

Computational spectrometers hold significant potential for mobile applications, such as on-site detection and self-diagnosis, due to their compact size, fast operation time, high resolution, wide working range, and low-cost production. Although extensively studied, prior demonstrations have been confined to a few examples of straightforward spectra. This study demonstrates a deep learning (DL)-based single-shot computational spectrometer capable of recovering narrow and broad spectra using a multilayer thin-film filter array. Our device can measure spectral intensities of incident light by combining a filter array, fabricated using wafer-level stencil lithography, with a complementary metal-oxide-semiconductor (CMOS) image sensor through a simple attachment. All the intensities were extracted from a monochrome image captured with a single exposure. Our DL architecture, comprising a dense layer and a U-Net backbone with residual connections, was employed for spectrum reconstruction. The measured intensities were input into the DL architecture to reconstruct the spectra. We collected 3,223 spectra, encompassing both broad and narrow spectra, using color filters and a monochromator to train and evaluate the proposed model. We reconstructed 323 test spectra, achieving an average root mean squared error of 0.0288 over a wavelength range from 500 to 850 nm with a 1 nm spacing. Additionally, the proposed multilayer thin-film filters were validated through scanning electron microscope (SEM) analysis, which confirmed uniform layer deposition and a high fabrication yield. Our computational spectrometer boasts a compact design, a rapid measurement time, a high reconstruction accuracy, a broad spectral range, and CMOS compatibility, making it well-suited for commercialization.