Abstract: Metasurfaces, as two-dimensional planar materials composed of subwavelength structural units, exhibit significant advantages over conventional optical components, including being more compact, lightweight, and flexible, as well as offering greater functional diversity. However, traditional forward design approaches for metasurfaces suffer from inefficiency and a limited capacity to address complex design requirements. In recent years, inverse design has gained increasing prominence in metasurface design. It does not rely on manual experience or intuition, circumventing the time-consuming processes such as parameter scanning. Consequently, it shortens the design cycle, improves efficiency, meets multi-functional requirements, and even creates counter-intuitive structures. This review systematically categorizes the current mainstream inverse design methods for metasurfaces into two categories: those based on optimization algorithms and those based on deep learning. It further introduces their fundamental principles, advantages and drawbacks, and their specific applications within the domain of metasurface research.