Abstract: Metasurfaces have emerged as a cutting-edge research focus in optics for their ability to manipulate electromagnetic waves with high efficiency at the sub-wavelength scale. Conventional design strategies, typically demand prolonged optimization cycles and rely heavily on the designer's domain expertise. To address these limitations, we propose an enhanced bidirectional neural-network framework (EBI-DNN) that integrates a forward prediction neural network (FPNN) with an inverse prediction neural network (IPNN). This hybrid architecture rapidly derives the requisite geometric parameters of a metasurface from a specified phase profile while resolving the "one-to-many" ambiguity inherent in conventional inverse design. We demonstrate the EBI-DNN framework on circular-polarization and linear-polarization multiplexed metasurfaces, validating its superior holographic performance through full-wave simulations, with peak signal-to-noise ratios (PSNR) of 18.75 dB and 16.82 dB, respectively. Training the EBI-DNN requires only 41 s, representing a 21-fold speed-up over the traditional bidirectional design network (BI-DNN), even when trained on the same dataset, it still achieves a 1.73-fold increase in training speed. Once trained, the network generates a complete metasurface design in mere seconds. Its streamlined architecture not only shortens development time but also lowers the level of prior expertise needed, in line with current demands for integrated and automated design workflows. Simulation results confirm that EBI-DNN delivers robust performance and offers broad application prospects, particularly for multi-functional and multi-channel metasurface architectures. This work thus provides a new paradigm for the rapid design of metasurfaces and lays the groundwork for large-scale fabrication of complex devices in the future.