Transfer learning for metamaterial design and simulation.

Area of Science:

• Electromagnetism
• Materials Science
• Computer Science

Background:

• Deep learning models, particularly residual neural networks (ResNets), require substantial data for effective training.
• Metasurface arrays, crucial in electromagnetic applications, present complex multi-scale problems that are data-intensive to simulate.
• The data bottleneck is a significant challenge in applying deep learning to novel metasurface designs.

Purpose of the Study:

• To demonstrate transfer learning as a method to improve the efficiency of training deep learning models for metasurface arrays.
• To assess the effectiveness of transfer learning across different problem domains with varying degrees of similarity to the original training task.
• To mitigate the data bottleneck in deep learning for electromagnetic metamaterials research.

Main Methods:

• Utilized residual neural networks (ResNets) for deep learning model training.
• Employed a quasi-analytical discrete dipole approximation (DDA) method for simulating electrically large metasurface arrays.
• Applied transfer learning to adapt pre-trained ResNet models to new metasurface design problems.

Main Results:

• Transfer learning significantly reduces the amount of data required for training deep learning models.
• Achieved a test mean absolute relative error of 3% with a pre-trained neural network in optimal transfer scenarios.
• Demonstrated data reduction by a factor of 1000 in best-case transfer learning applications.
• Showcased the efficiency of transfer learning across a spectrum of related electromagnetic metamaterial problems.

Conclusions:

• Transfer learning is a powerful tool for accelerating the design and analysis of metasurface arrays.
• This approach effectively overcomes the data bottleneck in deep learning for electromagnetic metamaterials.
• Leveraging pre-trained models via transfer learning enables rapid training and high accuracy even with limited data.