Deep-learning-accelerated meta-beam-splitter design accounting for near-field coupling effects.

Area of Science:

• Optics and Photonics
• Metamaterials
• Computational Electromagnetics

Background:

• Metasurfaces offer precise control over light wavefronts.
• Designing complex metasurfaces, like beam splitters, is computationally intensive.
• Near-field coupling effects significantly impact metasurface performance.

Purpose of the Study:

• To develop a rapid, deep-learning-based method for meta-beam-splitter design.
• To incorporate near-field coupling effects into the design process.
• To enable fast gradient-based optimization of metasurface structures.

Main Methods:

• A convolutional neural network (CNN) predicts coupling-aware near-field responses.
• Meta-atoms are encoded using a differentiable image representation.
• Gradients are backpropagated through the near-field response to structural parameters for optimization.

Main Results:

• A 1x5 meta-beam splitter was designed in 24 seconds (simulation).
• Achieved 92.33% uniformity in the designed meta-beam splitter.
• Demonstrated fast gradient-based optimization through differentiable near-field prediction.

Conclusions:

• Deep learning accelerates meta-beam-splitter design.
• The method effectively accounts for near-field coupling.
• Metasurfaces show promise for optical communications and sensing.