DNN-assisted waveguide width extraction via optical measurement of a single low-order Mach-Zehnder interferometer

Abstract

Abstract An accurate waveguide width extraction methodology is demonstrated by developing a deep-learning neural network (DNN) framework synergized with a single low-order Mach-Zehnder interferometer (MZI). Utilizing its intrinsic capacity for complex nonlinear capture, the proposed DNN achieves a simulation-to-prediction accuracy of 0.15 nm on a test dataset, establishing a robust geometry mapping. By leveraging fabrication tolerance analysis and group-index estimation from C-band MZI interference spectra, our method directly extracts the as-fabricated waveguide width via a pre-trained DNN. Experimental validation on 30 devices demonstrates an experimental estimation accuracy of 3.28 nm when compared to SEM measurements. Additionally, the system operates with a single low-order MZI featuring a reduced arm-length difference of 26.672 μm, while the conventional ones exceeding 100 μm. The superior precision and miniaturization capability make our approach an effective strategy for fabrication monitoring and high-density photonic circuit characterization.