Broadband Tunable Deep-UV Emission from AI-Optimized Nonlinear Metasurface Architectures

Abstract

Metasurfaces represent a pivotal advancement in nonlinear optics, leveraging high-Q resonant cavities to enhance harmonic generation. Multi-layer metasurfaces (MLMs) further amplify this potential by intensifying light-matter interactions within individual meta-atoms at the nanoscale. However, maximizing nonlinear efficiency demands extreme field confinement through optimized designs of large geometric and material parameters, which exceed traditional simulation's computational ability. To overcome this, we introduce NanoPhotoNet-NL, an AI-driven design tool employing a hybrid deep neural network (DNN) that synergizes convolutional neural networks (CNNs) and Long Short-Term Memory (LSTM) models. This framework accelerates nonlinear MLM design speed by four orders of magnitude while maintaining over 98.3% prediction accuracy relative to physical simulators. The optimized MLMs achieve quality factors exceeding 50, enabling broadband third-harmonic generation (THG) in the deep ultraviolet (DUV) from wavelengths 200 nm to 260 nm via parametric sweeps. Furthermore, NanoPhotoNet-NL facilitates dynamically tunable DUV nanolight sources with 20 nm spectral coverage in the UVC band using low-loss nonlinear phase change materials. This work marks a transformative leap in nonlinear metasurface engineering, unlocking high-performance, reconfigurable platforms for nonlinear and quantum optical nanodevices.