Design of a high-efficiency achromatic metalens for mid-infrared broadband applications enabled by deep-learning algorithms

This work presents a deep-learning-assisted inverse design framework for high-efficiency broadband achromatic reflective metalenses in the mid-infrared regime, aiming to overcome common limitations of conventional reflective metalenses, including relatively low focusing efficiency and difficulty in further improving overall optical performance. The method combines an MLP forward predictor with genetic algorithm optimization to design reflective binary meta-atoms that satisfy multi-wavelength phase matching and high-reflectance requirements. Using this framework, the designed metalens achieves near-diffraction-limited focusing over 8−12µm, with an average focusing efficiency of 74.47% and a maximum focal-length deviation of 3.1µm. After training, the MLP predicts the optical responses of approximately 500 meta-atoms in less than 1 s, enabling rapid candidate screening. Compared with existing inverse design methods, this framework directly handles non-differentiable 0/1 pixelated structures without differentiable solvers, continuous relaxation, or post-binarization, offering a practical route toward fast binary meta-atom and metalens design.