Broadband visible-light anti-reflection enabled by a bilayer dielectric metamaterial with deep-learning-driven inverse design

Achieving ultra-low reflection across the visible spectrum and under wide incident angles is of considerable importance for broadband antireflection coatings and optical interface engineering. Here, we introduce a bilayer subwavelength dielectric metamaterial together with a rapid and generalizable inverse-design framework based on a tandem residual deep neural network architecture. The metamaterial is fabricated entirely from the same dielectric material as the substrate, ensuring excellent compatibility and straightforward integration. Owing to its all-dielectric composition, the design principle can also be readily extended to other spectral bands as required. Furthermore, the framework can directly incorporate different substrate material parameters and output the corresponding optimized geometry, demonstrating strong adaptability and scalability. The inverse-design strategy integrates a forward network (FNet) and an inverse network (INet), both constructed using lightweight residual blocks to ensure solution uniqueness and high prediction fidelity. Simulations show that the optimized bilayer metamaterial achieves a maximum reflectivity of less than 3% throughout the visible band (400–700 nm) and maintains stable ultra-low-reflection performance for incident angles up to 80°. Beyond providing a compact and fabrication-friendly route to broadband wide-angle antireflection, this work establishes a fast, accurate, and broadly applicable inverse-design paradigm for next-generation nanophotonic structures.