Induced eccentricity splitting in disordered optical microspheres for machine learning enabled wavemeter
Accurate measurement of light wavelength is critical for applications in spectroscopy, optical communication, and semiconductor manufacturing, ensuring precision and consistency of sensing, high-speed data transmission and device production. Emerging reconstructive wavemeters synergize physical systems capable for pseudo-random wavelength dependent pattern formation with computational techniques to offer a promising alternative against established methods such as frequency beating and inteferometry for high-resolution and broadband measurements in compact and cost-effective devices. In this paper, we propose a novel type of compact and affordable reconstructive wavemeter based on the disordered chip with thousands of high quality-factor whispering gallery mode microcavities as physical model and a hybrid machine learning approach utilizing boosting methods and variational autoencoders implemented as wavelength interpreter. We leverage eccentricity mode splitting obtained via controllable deformation of the spherical microresonators in order to ensure the uniqueness of the wavelength patterns up to ultra-wide (~100 nm) spectral window while guaranteeing high (~100 fm) intrinsic sensitivity. The latter allocates the proposed model right next to the ultimate reconstructive wavemeters based on integrating spheres, but with superior miniaturization options and chip-scale integrability.