Terahertz (THz) Electromagnetically Induced Transparency (EIT) metasurface enable precise control and modulation of THz wave propagation, offering unique opportunities for a wide range of technological applications, including sensing, imaging, and communications. In recent years, the rapid advancement of deep learning has significantly accelerated the design process of these metasurfaces. Researchers have increasingly applied deep learning techniques to explore the underlying relationships between metasurface structures and their electromagnetic responses, addressing the complexities that traditional methods struggle to resolve efficiently. Consequently, the integration of deep learning into THz EIT metasurface design has gained great importance in both research and practical applications. This paper primarily focuses on the application of convolutional neural networks (CNNs) in the design methodology of THz metasurfaces. The main contributions include generating datasets for training purposes, mapping the actual metasurface structures into two-dimensional structure matrices, and developing both forward prediction and inverse design networks based on CNNs. These networks facilitate intelligent metasurface design by accurately predicting electromagnetic responses and enabling the reverse engineering of metasurface structures based on specified electromagnetic properties, significantly enhancing the efficiency and precision of the design process.