Abstract: Metasurfaces, as artificially engineered planar subwavelength arrays, can generate unique electromagnetic responses that enable functionalities beyond those of conventional optical devices. This is achieved by precisely manipulating parameters of incident light waves, including amplitude, phase, polarization state, and frequency. Traditional metasurface design workflows primarily rely on numerical simulation algorithms combined with parameter optimization. However, such methods typically exhibit high computational complexity and require substantial physical intuition and accumulated empirical expertise from the designer. In recent years, with the deep integration of various neural network architectures into metasurface design, design efficiency has been significantly enhanced while substantially expanding the capabilities of metasurfaces in light field manipulation. This review aims to summarize research progress on the fusion of neural networks with metasurface design and light field manipulation: First, existing methods are categorized into four typical frameworks based on differences in design paradigms: forward prediction networks, inverse design networks, physics-informed neural networks, and end-to-end design networks. Subsequently, we analyze innovative applications of neural networks in metasurface-based light field manipulation, comparing and contrasting their core principles and technical advantages. Finally, we summarize the key challenges currently faced by metasurfaces in integrating neural networks in both design theory and practical applications, and outline future research directions.