Denoising-autoencoder-facilitated MEMS computational spectrometer with enhanced resolution on a silicon photonic chip

Silicon photonics enables the construction of chip-scale spectrometers, in which those using a single tunable interferometer provide a simple and cost-effective solution. Among various tuning mechanisms, electrostatic MEMS reconfiguration stands out as an ideal candidate, given its high tuning efficiency and ultra-low power consumption. Nonetheless, MEMS devices face significant noise challenges arising from their susceptible minuscule components, adversely impacting spectral resolution. Here, we propose a distinct paradigm of spectrometers through synergizing an easily-fabricated MEMS-reconfigurable low-loss waveguide coupler on a silicon photonic chip and a convolutional autoencoder denoising (CAED) mechanism. The spectrometer offers a 300 nm bandwidth and a reconstruction resolution of 0.3 nm in a noise-free condition. In a noisy environment with a signal-to-noise ratio as low as 30 dB, the reconstruction resolution of the interferograms processed by the CAED exhibits an enhancement from 1.2 to 0.4 nm, approaching the noise-free value. Our technology is envisaged to provide a powerful and cost-effective solution for applications requiring accurate, broadband, and energy-efficient spectral analysis. The Internet-of-Things era desiderates miniature spectrometers. Here, the authors present a chip-scale spectrometer through synergizing MEMS modulation and autoencoder denoising, achieving accurate, broadband, and energy-efficient spectral analysis.