In recent years, the integration of terahertz metasurfaces with the rapidly developing machine learning technologies has injected new vitality into the development. Thanks to the efficient iterative rates and diverse network architectures of machine learning, it has been widely applied in both the surrogate modeling and intelligent design of full-wave numerical simulations for terahertz metasurfaces. On the other hand, terahertz metasurfaces provide a platform for optical computing, enabling computation at the speed of light. This paper focuses on the application of machine learning in the terahertz metasurfaces. It first elaborates on the design frameworks and principles based on deep neural networks and generative adversarial networks, revealing the advantages of machine learning in multi-dimensional parameter optimization. Secondly, using the all-optical diffraction neural network architecture as an example, it introduces the physical mechanisms by which terahertz metasurfaces can be used to execute machine learning algorithms, along with potential derived applications. Finally, the paper looks ahead to the development trends and application prospects of the integration of machine learning with terahertz metasurfaces.