Physics-informed deep learning for 3D modeling of light diffraction from optical metasurfaces.

Area of Science:

• Optics and Photonics
• Computational Electromagnetics
• Metasurface Engineering

Background:

• Traditional numerical solvers for optical metasurfaces are computationally intensive.
• Data-driven deep learning methods require large datasets, limiting their applicability.
• Accurate simulation of light-matter interactions is crucial for metasurface design.

Purpose of the Study:

• To develop a data-free deep learning approach for efficient simulation of light diffraction from 3D optical metasurfaces.
• To model polarization effects and wavefront manipulation using physics-informed neural networks (PINNs).
• To accelerate the computation of electromagnetic field (EMF) responses.

Main Methods:

• Utilizing a physics-informed neural network (PINN) trained solely on governing physics principles.
• Incorporating vector Maxwell's equations, Floquet-Bloch boundary conditions, and perfectly matched layers (PML) into the PINN.
• Developing a PINN-based EMF solver for rapid simulation of light scattering.

Main Results:

• The PINN accurately simulates near-field and far-field responses, including polarization impacts.
• The model efficiently handles variations in meta-atom geometry and illumination settings.
• Trained PINN solver achieves millisecond-level inference for multiple inputs, demonstrating significant speed-up.

Conclusions:

• PINNs offer a data-free, efficient, and accurate alternative for simulating light diffraction from optical metasurfaces.
• This method accelerates computational electromagnetics, enabling faster design and optimization of metasurfaces.
• The approach overcomes limitations of traditional solvers and data-driven networks by leveraging physics principles.