Comparison of algorithms using deep reinforcement learning for optimization of hyperbolic metamaterials

Area of Science:

• Materials Science
• Optics
• Energy Conversion

Background:

• Hyperbolic metamaterial absorbers offer broadband absorption for photo-thermoelectric devices.
• Optimizing metamaterial geometry is complex due to numerous parameters and spectral alignment needs.

Purpose of the Study:

• To compare three deep reinforcement learning algorithms for optimizing hyperbolic metamaterial absorbers.
• To assess the prediction accuracy of these algorithms with limited datasets.
• To determine the impact of algorithm choice on optimization reliability.

Main Methods:

• Implementation and comparison of three distinct deep reinforcement learning algorithms.
• Analysis of absorption spectra generated by each algorithm.
• Evaluation of prediction accuracy based on limited absorption spectra datasets.

Main Results:

• Single algorithm optimization with limited data can lead to structural misestimations.
• Multiple algorithms are crucial for accurate and reliable optimization of metamaterial absorbers.
• The optimal algorithm identified increased metamaterial thermoelectric conversion power by five times.

Conclusions:

• The selection of appropriate optimization algorithms is critical for accurate metamaterial design.
• Employing multiple deep reinforcement learning algorithms enhances the reliability of structural optimization.
• Advanced metamaterial absorbers significantly improve photo-thermoelectric device performance.