Reliable, efficient, and scalable photonic inverse design empowered by physics-inspired deep learning

Area of Science:

• Metamaterials and Photonics
• Computational Electromagnetics
• Artificial Intelligence in Engineering

Background:

• On-chip computing metasystems offer high-speed, low-power processing but face design challenges.
• Current numerical and analytical design methods lack efficiency and accuracy.
• A novel inverse design paradigm is needed for complex metasystem development.

Purpose of the Study:

• To propose a physics-inspired deep learning architecture, the electromagnetic neural network (EMNN), for efficient inverse design.
• To enable rapid, accurate, and flexible design of on-chip computing metasystems.
• To overcome limitations of existing design methodologies in metamaterial-based computing.

Main Methods:

• Developed EMNN, comprising EMNN Netlet for local field solving and Huygens-Fresnel Stitch for prediction concatenation.
• Utilized EMNN for direct, rapid, and accurate full-wave field predictions based on arbitrary input fields and structures.
• Applied EMNN to design metasystems capable of handwritten digit and speech command recognition.

Main Results:

• EMNN achieved a 17,000-fold increase in design speed compared to analytical models.
• EMNN reduced modeling error by two orders of magnitude relative to numerical models.
• Demonstrated interpretability, generalization ability, and high fidelity in EMNN designs.

Conclusions:

• EMNN provides an efficient, reliable, and flexible paradigm for inverse design of computing metasystems.
• The integration of deep learning with physics principles enhances design capabilities.
• EMNN is applicable to complex, large-scale devices like on-chip optical diffractive networks, advancing computing metasystems.