.png){kind=icon}
.png){kind=icon}
Return
Enhanced bidirectional neural network for tailoring high-Q Fano resonances in a Y-shaped metasurface array
.png){kind=icon}
PRE
AI
B
Y
Q
Q
Q
G
S
H
S
李
DOI：10.1088/1361-6463/ae611f.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Optical resonances with high quality (Q) factor offer advantages for various research fields such as sensing, lasing and filtering. However, designing and optimizing high-Q resonances in all-dielectric metasurfaces is physically complexing, computationally demanding and time-consuming. In this work, we propose a metasurface array with Y shaped a-Si blocks unit cell which could support high-Q Fano resonance. With detailed analysis, it is confirmed that the destructive interference between electric dipole and electric quadrupole reduces the radiative loss of the resonant mode, leading to the formation of a resonant dip in the transmission spectrum. The Q factor of the resonance could be improved when the resonant wavelength approaches the wavelength of the Rayleigh anomaly. A two-stage deep learning framework is further employed to optimize the Q factor of the resonance and inverse design a structure supporting resonance with desired Q factor. At the first stage, a forward neural network (FNN) based on a residual multilayer perceptron (ResMLP) is integrated with the non-dominated sorting genetic algorithm II multi-objective optimizer to explore the trade-off between the Q factor and modulation depth, resulting in an optimized design with a Q factor of 5521—representing a 41% improvement over the highest configuration in the data set. In the second stage, a bidirectional neural network (BNN) with an architecture similar to that of the FNN is proposed, enabling on-demand generation of Fano resonances defined by mathematical parameters in the Fano formula. This model exhibits strong data efficiency and achieves accurate inverse design using a compact dataset of approximately 4000 samples, reaching a mean squared error of . This study offers a general and computationally efficient approach for designing high-Q dielectric metasurfaces, accelerating the discovery of high-performance devices across a wide range of nanophotonic applications.
Keywords:
High-Q Fano resonance
all-dielectric metasurfaces
deep learning optimization
Y-shaped a-Si blocks
inverse design
Journal
J
IF:
3.2
Papers:
726
Citations:
0
Organization
Cited Papers