Design of macroscale optical systems with metaoptics using transformer-based neural networks

Area of Science:

• Optics and Photonics
• Computational Electromagnetics
• Materials Science

Background:

• Metaoptics offer advanced wavefront control in a compact form factor.
• Simulating metaoptics across multiple length scales presents computational challenges, balancing speed and accuracy.
• Existing methods either are computationally intractable or suffer from predictive inaccuracies.

Purpose of the Study:

• To develop a computationally efficient and accurate method for simulating metaoptic devices.
• To overcome the speed-accuracy trade-off in metaoptic optical response prediction.
• To enable faster optimization and application of metaoptics.

Main Methods:

• Implemented a transformer-based neural network solver for metaoptic optical response.
• Integrated the neural network with commercial ray optics software using Fourier propagation.
• Compared simulation results with traditional finite-difference time-domain (FDTD) methods and approximation techniques.

Main Results:

• The neural network method achieved over three orders of magnitude speedup compared to FDTD.
• The method demonstrated a low deviation of 0.47% in total irradiance versus full wave simulation.
• Accuracy was nearly two orders of magnitude better than standard approximation methods for metaoptics.

Conclusions:

• The developed neural network approach provides a highly accurate and efficient solution for metaoptic simulation.
• This method facilitates the optimization of metaoptics, accelerating their integration into optical systems.
• The findings pave the way for broader and faster adoption of metaoptic technology.