Multi-objective optimization design of bidirectional metamaterial absorber based on hybrid semi-supervised learning and reinforcement learning

Electromagnetic coupling effects among multilayer structures in nanophotonic devices often lead to intricate trade-offs in optical performance, posing significant challenges for conventional simulation-driven design. In this study, we propose an intelligent hybrid optimization framework (SSL-RL) integrating semi-supervised learning (SSL) with reinforcement learning (RL) to enable multi-objective trade-off optimization design for reconfigurable bidirectional metamaterial absorbers (BMAs). A mean teacher-based SSL model is trained using limited labeled simulation data and high-quality unlabeled data generated via a denoising diffusion probabilistic model (DDPM), achieving accurate predictions of both broadband and narrowband absorption spectra while substantially reducing computational cost. The trained SSL model was integrated into the reinforcement learning model based on the advantage actor-critic (A2C) algorithm, serving as an environment surrogate to accelerate the optimization of BMA structures. The optimized BMA structure achieved an average broadband absorption of 97.20% across the 0.6-1.6μm spectral range under backward illumination (-z-direction), while demonstrating a narrowband absorption peak of 99.3% with a 22.5 nm full-width at half-maximum (FWHM) under forward illumination (+z-direction). Furthermore, the designed BMA structure exhibits excellent reconfigurability, enabling flexible modulation of absorption responses through liquid crystal (LC) molecular angle control, while demonstrating outstanding angular tolerance and polarization insensitivity. This approach provides an efficient and scalable solution for the intelligent design of complex nanophotonic devices.