Abstract
Accurate wavelength measurement is critical for spectroscopy, optical communications, semiconductor manufacturing, and quantum research. Emerging reconstructive wavemeters are compact, cost-effective devices that utilize pseudo-random wavelength patterns and computational techniques to provide high-resolution, broadband alternatives to solutions based on frequency beating and interferometry. We propose a novel reconstructive wavemeter that synergizes the advantages of both approaches. Its physical model is based on the integration of thousands of high-quality-factor optical microcavities, which are deformed to stimulate whispering gallery mode splitting. For realizing a wavelength interpreter, we developed a hybrid machine learning approach utilizing boosting methods and variational autoencoders. This enabled the implementation of wavelength interpretation as a rigorous regression task for the first time. The introduced novel concept ensures the uniqueness of the wavelength patterns up to ultra-wide (~100 nm) spectral window while guarantees high (~100 fm) intrinsic sensitivity. The latter allocates the proposed solution right next to the ultimate reconstructive wavemeters based on integrating spheres, but with less calibration efforts, featuring superior miniaturization options and chip-scale integrability.