A bidirectional deep learning framework for inverse design and optical response prediction of mid-infrared metamaterial absorbers

A unified bidirectional deep learning framework is developed for mid-infrared metamaterial absorbers. It comprises a Transformer-based multilayer perceptron (TRMLP) for forward spectral prediction and a FiLM-residual-attention multilayer perceptron (FRAM) for inverse structural design, where FiLM denotes feature-wise linear modulation. The models were trained on 30,000 finite-difference time-domain (FDTD)-generated samples containing nine geometric parameters and 500 spectral points over 6–18 μ ⁢ m . Using identical sample indices and evaluation protocols, TRMLP achieved a mean squared error (MSE) of 0.048, a mean absolute error (MAE) of 0.122, and an 𝑅 2 of 0.977. Relative to the global- and local-spectrum-aware Transformer (GLSaT) baseline, it reduced MSE and MAE by 4.0% and 19.2%, respectively, and increased 𝑅 2 by 0.087. It predicted a 500-point spectrum in 35 ms, providing a 9743-fold speedup over a 341-s FDTD simulation without particle swarm optimization (PSO). FRAM achieved an MSE of 0.024 and an MAE of 0.085, reducing the corresponding errors of the conditional diffusion model DiffMeta by 14.0% and 6.4%. Permutation importance identified ℎ 2 and ℎ 4 as the most and least influential parameters, respectively. FDTD reconstruction, electromagnetic-field analysis, angular-sensitivity simulations, fabrication, and optical measurements demonstrated broadband absorption over 7–14 μ ⁢ m , supporting the physical feasibility of the framework.