Deep Reinforcement Learning Inverse Design of Dielectric Metasurfaces Supporting Dual Quasi‐BIC Resonances

ABSTRACT Optimizing resonance characteristics in dielectric metasurfaces, including the resonance wavelength, quality factor, and near‐field enhancement, is fundamental to nanophotonic applications such as sensing, nonlinear optics, and enhanced light–matter interactions. This task is especially challenging in dual quasi‐bound states in the continuum (quasi‐BIC) metasurfaces, where narrow‐linewidth resonances, strong modal coupling, and sensitivity to geometric perturbations create a highly nonconvex and constrained design landscape. Existing optimization methods, including heuristic algorithms and deep learning, are useful in simpler settings, but they often become inefficient, rely on large training datasets, or struggle to coordinate tightly coupled objectives in quasi‐BIC systems. To address these limitations, we introduce a deep reinforcement‐learning framework for the inverse design of a dielectric metasurface supporting dual quasi‐BIC resonances. A dueling double deep Q‐network with rollout is used to mitigate myopic search under coupled objectives and unstable reward responses. The resulting framework coordinates two resonance wavelengths, achieves Q factors of approximately 1219 and 1104 for the two quasi‐BIC modes, reaches an electric‐field amplitude enhancement of approximately 81, reduces the design time from 1 month to 82 h, and achieves 5 × higher sample efficiency than other mainstream high‐performance algorithms. These results demonstrate the potential of reinforcement learning for strongly coupled multi‐objective photonic optimization.