Metasurfaces have demonstrated extensive application potential in multiple fields. During the design of metasurfaces, it is necessary to optimize the components based on factors such as polarization, amplitude distribution, and phase distribution. This process typically requires the involvement of experts and is time-consuming. A method is proposed for inverse design of components that integrates high-precision ultra-wideband spectral forward prediction using neural networks and genetic algorithms. This method can simultaneously predict the amplitude and phase of a 16×16 high-degree-of-freedom discrete grid structure within the frequency range of 0.5~2 THz. The amplitude prediction accuracy can reach 0.019, and the phase prediction accuracy can reach 4.332°. The average optimization time for a single metasurface unit is 1.5 min. Two sets of 3 bit frequency-multiplexed and polarization-multiplexed metasurfaces are designed and simulated, and the simulation results validate the effectiveness of the proposed method. The method provides a new paradigm for the rapid design of components facing complex application scenarios and is of significant reference value to metasurface designers.