Active Learning Optimisation of Binary Coded Metasurface Consisting of Wideband Meta-Atoms.

Area of Science:

• Electromagnetics and Metamaterials
• Computational Science and Engineering
• Machine Learning Applications

Background:

• Metasurface array design for radar cross-section minimization is a key research area.
• Conventional optimization algorithms like genetic algorithm (GA) and particle swarm optimization (PSO) are computationally intensive, limiting their use for large arrays.
• High computational complexity hinders efficient optimization of complex metasurface designs.

Purpose of the Study:

• To introduce and evaluate active learning as a machine learning optimization technique for metasurface arrays.
• To demonstrate the significant reduction in computational time offered by active learning compared to traditional methods.
• To achieve comparable or superior optimization results for radar cross-section minimization using active learning.

Main Methods:

• Application of active learning, a machine learning optimization strategy.
• Comparison of active learning performance against the genetic algorithm (GA).
• Utilizing accurately trained surrogate models within the active learning framework.

Main Results:

• Active learning reduced optimization time for a 10x10 metasurface array from 13,260 minutes (GA) to 65 minutes.
• For a 60x60 metasurface array, active learning was 24x faster than GA for similar results.
• Machine learning optimization drastically cut computational expenses, especially for larger metasurface designs.

Conclusions:

• Active learning offers a computationally efficient alternative to genetic algorithms for metasurface array optimization.
• The proposed method significantly accelerates the design process, making large-scale optimizations feasible.
• Leveraging surrogate models further enhances the speed and efficiency of active learning in this domain.