Abstract: Multilayer composite metamaterial absorber, has multi-level impedance matching characteristics. Compared with single-layer MMA, it has a wider operating bandwidth and a wider application range. However, MC-MMA has numerous structural parameters. In the process of structural design, time-consuming electromagnetic simulation calculations are often relied upon, resulting in low design efficiency. To improve both design efficiency and prediction accuracy, we propose an inverse de-sign method (FPETNet-DNRGA)which combines a Frequency Positional Embedded Transformer (FPETNet) with a Dynamic Neighborhood-Regulated Genetic Algorithm (DNRGA). FPETNet serves as a forward predictor for MC-MMA, rapidly and accurately replacing the electromagnetic simulation. DNRGA adaptively adjusts the search neighborhood during population evolution, markedly enhancing global convergence and avoiding local optima. To validate the effectiveness of the proposed approach, we conduct algorithmic tests in three representative design scenarios: broadband response design, ultra-thin structure design, and thick-ness-constrained design. Experimental results show that FPETNet-DNRGA reduces the design time to just 0.05% of that required by electromagnetic simulation, while the optimized structures exhibit reflectance below -10 dB across the 2~18 GHz (S-to Ku-band) frequency range. Compared with electromagnetic simulation, the mean absolute percentage errors (MAPE) are 4.42%, 7.12%, and 4.75% for the three scenarios, respectively, further demonstrating the significant advantages of our method in both efficiency and accuracy.