Abstract: To address the challenge of balancing performance and efficiency in the design of electromagnetic metamaterials, a joint optimization framework integrating a Residual Fully-Connected Network (RFCN) with an Improved Genetic Algorithm (IGA) is proposed. First, an RFCN model was constructed to efficiently predict the reflectance curves of a stepped-cone lattice structure in the 2–40 GHz frequency band (test set RMSE = 0.38). Subsequently, an IGA incorporating catastrophic mutation, large-scale mutation, and precise screening mechanisms was introduced for global parameter optimization. Simulation and experimental results showed that the number of convergence generations was reduced from 36 to 14, indicating a significant improvement in optimization efficiency. The optimized structure achieved a reflectance below -10 dB in the 3.4–36.3 GHz and 39–40 GHz bands, with effective absorption bandwidths of 91.2% (simulation) and 89.3% (measurement), respectively. Significant absorption peaks were observed at 3.79, 15.5, and 35.3 GHz, with a minimum reflectance of -41.77 dB. The proposed method overcomes the limitations of traditional physics-based design and provides an efficient and accurate approach for electromagnetic metamaterial design, demonstrating strong potential for applications in spacecraft stealth and electromagnetic compatibility design.