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

A hyperbolic metamaterial absorber has great potential for improving the performance of photo-thermoelectric devices targeting heat sources owing to its broadband absorption. However, optimizing its geometry requires considering numerous parameters to achieve absorption that aligns with the radiation spectrum. Here, we compare three algorithms using deep reinforcement learning for the optimization of a hyperbolic metamaterial absorber. By analyzing the absorption spectra obtained from the three algorithms with limited number of datasets, we assessed the prediction accuracy of each algorithm. Our findings indicate that relying on a single algorithm for optimization, particularly with a small number of datasets, can lead to misestimations in structural optimization. This underscores the importance of using multiple algorithms to ensure accurate and reliable optimization results. Finally, by utilizing the optimal algorithm, we achieved to increase the power generation of the metamaterial thermoelectric conversion by five times.