MxDiffusion: A Physics-Aware Maxwell's Law-Guided Diffusion Model Strategy for Inverse Photonic Metasurface Design.

Area of Science:

• Photonics
• Computational Electromagnetics
• Materials Science

Background:

• Inverse design of photonic structures is crucial for advanced optical devices.
• Existing data-driven methods often struggle with physical constraints and accuracy.

Purpose of the Study:

• To introduce MxDiffusion, a hybrid physics- and data-driven framework for efficient and accurate photonic structure generation.
• To improve the fidelity and physical consistency of inverse design processes.

Main Methods:

• Developed a two-stage diffusion model incorporating Maxwell's equation-based loss.
• The first stage embeds physical insights, and the second maps representations to geometries with high fidelity.
• Validated on gold pattern optimization and tunable bandpass filter design.

Main Results:

• MxDiffusion demonstrates superior performance compared to conventional data-driven diffusion models.
• Achieved higher accuracy, especially for out-of-distribution targets and constrained resonance conditions.
• Successfully generated complex photonic structures for specific optical properties.

Conclusions:

• MxDiffusion offers a powerful and generalizable physics-guided inverse design paradigm.
• The hybrid approach significantly enhances the efficiency and accuracy of photonic structure design.
• This framework paves the way for novel photonic device development.