The development of sophisticated photonic technologies, demands precise control over light propagation across multiple dimensions. Metasurfaces achieve this by manipulating the phase at each meta-unit, tuning amplitude, phase, and polarization states. However, the conventional design approach for these meta-units, heavily dependent on iterative simulations, imposes significant computational burdens. Addressing these constraints, we introduce a deep neural network combining forward modeling and inverse design, achieving both computational efficiency and prediction fidelity. As proof of principle, we successfully designed, simulated, and experimentally validated polarization-multiplexed metalens. This versatile strategy is readily applicable to the design of diverse wavefront-shaping components, such as beam deflectors and holography, paving the way for a new generation of customizable, multifunctional meta-devices.