Functional design of metamaterial absorbers using a combined convolutional neural network-genetic algorithm approach

Electromagnetic absorbers based on metamaterials are essential for mitigating electromagnetic interference and crosstalk in highly integrated photonic devices, yet their design remains computationally expensive and often constrained by predefined structural patterns. Here, we propose a functional design framework that integrates a convolutional neural network (CNN) with a genetic algorithm (GA) to efficiently explore the vast topological space of metamaterial absorbers. A shape grammar-based encoding scheme is introduced to represent the MXene (Ti₃C₂T_x) resonator pattern as a 10 × 10 binary matrix, enabling high-semantic, reversible definitions of complex patterns while maintaining fabrication feasibility. A CNN surrogate model is trained on 5000 samples generated by finite-element simulations, accurately predicting the absorption spectrum from 0.8 to 3.3 μm with a mean absolute error below 0.02. The GA then iteratively evolves the population of structural patterns and dielectric layer thickness, guided by the CNN predictions, to maximize absorption above 85%. The optimized absorbers achieve average absorption rates exceeding 94% and exhibit rich pattern diversity, including large irregular structures combined with small isolated features. Furthermore, by simply modifying the fitness function, the framework rapidly designs a polarization-insensitive absorber without relying on mirror or rotational symmetry. Electromagnetic field analysis reveals that short-wavelength absorption originates from intrinsic material loss and surface localization, while long-wavelength peaks arise from synergistic surface plasmon resonance and Fabry-Pérot cavity resonance. The surrogate-assisted GA reduces the number of required full-wave simulations by a factor of 30 compared with conventional GA, demonstrating strong reusability and significantly accelerating the inverse design of high-performance metamaterial absorbers.