Deep reinforcement learning hybrid models enabled broadband metamaterial solar absorbers optimization and design

Machine learning has revolutionized the design and optimization of metamaterial absorbers, yet existing approaches face notable challenges. Data-driven deep learning (DL) methods often require extensive datasets, limiting efficiency and hindering generalization beyond the training domain. Meanwhile, exploratory reinforcement learning (RL) methods eliminate the need for a preconstructed dataset but require extensive training time for exploration. In this study, we introduce, for the first time, an advanced hybrid model architecture, D-3DQN, combining DL and RL for the optimization of high performance metamaterial solar absorbers (MSAs). The proposed method integrates the fast training capabilities of DL with the exploratory advantages of RL, partially mitigating the time-consuming process of dataset acquisition, significantly reducing RL exploration time, and effectively addressing the generalization limitations of DL models beyond the dataset. Using this approach, we designed polarization-insensitive MSAs with average absorptivities of 98.51 % and 98.32 %, respectively, in the 0.4–2.8 μm wavelength range. Compared to using DL or RL models alone under the same conditions, the D-3DQN model reduced the design time by more than fivefold while delivering superior absorption performance. This method offers a novel design paradigm for high-performance metamaterial absorbers and can be extended to the design of other nanophotonic devices.