Physics-Guided Hierarchical Neural Networks for Maxwell's Equations in Plasmonic Metamaterials

Area of Science:

• Photonics
• Machine Learning
• Computational Physics

Background:

• Traditional machine learning (ML) in photonics requires extensive data, limiting practical applications.
• Resource-intensive data generation and training hinder the widespread use of ML in photonics.

Purpose of the Study:

• To develop a machine learning approach that significantly reduces data requirements in photonics.
• To improve the physics-consistency and generalizability of ML models for photonic applications.

Main Methods:

• Embedding Maxwell's equations into ML model design and training.
• Utilizing a physics-guided machine learning (PGML) approach.
• Applying the method to predict field distributions in hyperbolic meta-material photonic funnels.

Main Results:

• Demonstrated significant reduction in required training data for ML models.
• Enhanced physics-consistency and generalizability of the developed ML models.
• Successfully predicted complex field distributions in multilayered plasmonic-dielectric composites.

Conclusions:

• Physics-guided machine learning (PGML) offers a practical solution for ML in photonics.
• PGML enables the development of ML tools that do not necessitate extremely large training sets.
• Hierarchical network design facilitates knowledge transfer and the emergence of effective medium theories within neural networks.