Abdelraouf and Colleagues Designs Nanophotonet-Pinl Framework for Nonlinear Metasurface Optimisation

Omar A. M. Abdelraouf, Institute of Materials Research and Engineering, and colleagues have created a new artificial intelligence framework, NanoPhotoNet-PINL, that designs multi-layer metasurfaces to sharply enhance the efficiency of second-harmonic generation, a key process for advancements in nanophotonics and quantum technologies. Their approach overcomes limitations of conventional 3R-phase molybdenum disulfide (3R-MoS2) by optimising light-matter interaction and phase matching within nanoscale cavities, achieving over three orders of magnitude enhancement in second-harmonic generation intensity compared to a bare 3R-MoS2 flake. This intelligent inverse design paradigm promises a generalisable route towards high-efficiency nonlinear optical devices.

AI-driven design unlocks substantial second harmonic generation enhancement in multi-layer

Second harmonic generation (SHG) intensity increased by over three orders of magnitude with a new multi-layer metasurface design, surpassing the performance of a bare 3R-MoS2 flake on a planar substrate. Previously, efficient frequency conversion proved difficult due to weak light-matter interaction and poor phase matching within materials, making substantial enhancement unattainable. The fundamental principle behind SHG involves the nonlinear response of a material to intense light, where photons with frequency ω are effectively doubled to produce photons with frequency 2ω. However, the efficiency of this process is heavily dependent on constructive interference between the generated second-order nonlinear polarisation and the driving field. In atomically thin 2D materials like 3R-MoS2, achieving this phase matching is particularly challenging due to their weak light-matter interaction and the difficulty in controlling the propagation of both the fundamental and second harmonic waves. The Institute of Materials Researches and Engineering team developed NanoPhotoNet-PINL, an AI framework that intelligently maps desired optical properties to precise nanoscale structures, optimising geometry and material composition for maximum efficiency. Validating the framework, consistently high prediction accuracy exceeded 99.2% when designing multi-layer metasurfaces (MLMs). It successfully reconstructed target resonances across a broad spectrum, from 390 nanometres to 800 nanometres, achieving Q-factors of 123 at 390nm, 106 at 500nm, 32 at 700nm, and 45 at 800nm. These Q-factors represent the resonance sharpness, with higher values indicating a more confined and enhanced electromagnetic field. Finite-Difference Time-Domain simulations confirmed these findings, demonstrating excellent agreement between predicted and simulated optical responses. The optimised MLMs yielded up to a 123-fold improvement in second harmonic generation in the ultraviolet region compared to a simple thin-film structure. This substantial enhancement is crucial for applications requiring efficient frequency doubling, such as high-resolution microscopy and optical sensing. Although these enhancements are significant, current work does not address the scalability of fabrication processes needed to consistently and cost-effectively produce these nanoscale structures for widespread application. Advanced lithographic techniques, such as electron beam lithography or nanoimprint lithography, may be required, presenting significant manufacturing challenges.

Optimised Metasurface Design via Artificial Intelligence Constrains Complex Spectral Targeting

The authors acknowledge a key constraint within their reported 99.2% inverse-design prediction efficiency, specifying it applies only “along the linear spectral manifold”. This suggests performance may diminish when targeting more complex spectral characteristics, limiting the framework’s adaptability to a wider range of optical designs. The ‘linear spectral manifold’ refers to the range of wavelengths where the AI model has been thoroughly trained and validated. Deviating significantly from this range, or attempting to engineer more intricate spectral responses, could lead to reduced prediction accuracy. Furthermore, the current demonstration is specifically tailored to 3R-MoS2, and further investigation is needed to determine the extent to which NanoPhotoNet-PINL can be readily applied to other nonlinear materials or device configurations. The choice of 3R-MoS2 is based on its relatively strong second-order nonlinearity and its compatibility with 2D material-based nanophotonics. However, other materials, such as gallium nitride or lithium niobate, also exhibit significant nonlinearities and could potentially benefit from this AI-driven design approach. Integrating Maxwell-based nonlinear electrodynamics into the AI training process enabled physics-guided optimisation of the MLM structures, achieving a substantial increase in nonlinear conversion and radiation efficiency. This physics-guided approach is critical for ensuring that the AI-generated designs adhere to fundamental physical principles, preventing the creation of unrealistic or non-functional structures. The resulting 3R-MoS2/MLM platform offers a scalable route towards compact, CMOS-compatible nonlinear metasurface devices, potentially enabling applications including on-chip frequency converters, quantum light sources, and all-optical signal processing. CMOS compatibility is particularly important for integration with existing microelectronic circuits, paving the way for highly integrated photonic systems. Quantum light sources, which generate non-classical states of light, are essential for quantum communication and computation, while all-optical signal processing offers the potential for faster and more energy-efficient data processing.

AI-driven design of molybdenum disulfide metasurfaces for nonlinear optical frequency doubling

An artificial intelligence framework, NanoPhotoNet-PINL, has been developed by The researchers of Materials Research and Engineering (IMRE), Agency for Science, Technology and Research (A*STAR) in Singapore, to optimise multi-layer metasurfaces (MLMs) for enhanced second harmonic generation (SHG) in 3R-phase molybdenum disulfide (3R-MoS2). This new approach directly maps desired spectral responses to the required geometry and material composition of the metasurface, maximising light-matter interaction and phase matching, both vital for efficient frequency doubling. Traditional trial-and-error methods for designing complex, multi-layered structures are impractical due to the vast number of possible configurations and the simultaneous need to optimise both linear and nonlinear optical responses. The design space for MLMs is immense, with numerous parameters governing the layer thickness, material composition, and arrangement of individual nanostructures. Exhaustively searching this space using conventional optimisation techniques is computationally prohibitive. NanoPhotoNet-PINL overcomes these challenges by integrating physics-based modelling with a hybrid deep learning architecture, achieving an inverse-design prediction efficiency exceeding 99.2% within a specific range of light wavelengths. The hybrid architecture likely combines convolutional neural networks (CNNs) for feature extraction with other deep learning techniques to capture the complex relationships between structure, material properties, and optical response. This success prompts investigation into broadening its application beyond the current material and exploring its potential with more complex light interactions. The team suggest NanoPhotoNet-PINL establishes a generalizable method for designing nonlinear multi-layer metasurfaces and phase-matched cavities, potentially leading to compact, CMOS-compatible nonlinear metasurface devices. The ability to design phase-matched cavities is crucial for enhancing nonlinear effects by ensuring that the fundamental and second harmonic waves propagate in phase, maximising the efficiency of frequency conversion. Further research will likely focus on expanding the AI framework’s capabilities to handle more complex designs and materials, as well as developing scalable fabrication techniques for realising these advanced nonlinear optical devices.

Researchers developed NanoPhotoNet-PINL, an artificial intelligence framework that designs multi-layer metasurfaces to significantly enhance the efficiency of second harmonic generation in 3R-phase molybdenum disulfide. This approach overcomes limitations of previous designs by optimising both material composition and geometry to maximise light-matter interaction and phase matching. The framework achieved over 99.2% prediction efficiency and resulted in a more than three orders of magnitude increase in second harmonic generation intensity compared to the bare material. The authors intend to expand the framework’s capabilities to accommodate more complex designs and materials, establishing a generalisable method for nonlinear metasurface design.