Global Optimization of Dielectric Metasurfaces Using a Physics-Driven Neural Network.

Area of Science:

• Metasurface design
• Computational electromagnetics
• Machine learning for physics

Background:

• Topology optimization is crucial for designing efficient metasurfaces.
• Current methods can be computationally intensive and may not fully explore the design space.

Purpose of the Study:

• To introduce a novel global optimizer based on a conditional generative neural network.
• To enable the generation of highly efficient topology-optimized metasurfaces across various parameters.

Main Methods:

• A conditional generative neural network was trained using electromagnetic simulations and gradient-based backpropagation.
• The network generates an initial distribution of devices, then refines it towards optimal regions.
• Metagratings across wavelengths and angles served as a model system.

Main Results:

• The generative network produced metasurface devices with efficiencies matching or exceeding traditional adjoint-based optimization.
• The method demonstrated a lower computational cost compared to existing techniques.
• The approach successfully optimized metagratings for performance across diverse operating conditions.

Conclusions:

• Reframing adjoint-based optimization for generative neural network training is a generalizable method.
• This technique offers a computationally efficient alternative for optimizing physical systems using gradient information.
• The developed global optimizer effectively generates high-performance metasurface designs.