Controlled-source audio-frequency magnetotellurics (CSAMT) uses artificial sources, providing strong anti-interference capabilities. It is widely used in oil exploration, mineral surveys and other areas. Traditional 2D inversion technology is mature, and deep learning has recently made some research advancements in geophysical exploration. There is still a research gap in applying deep learning to CSAMT inversion. Therefore, developing a 2D inversion algorithm for CSAMT based on deep learning is highly significant for advancing the use of deep learning in electromagnetic exploration. The characteristics of deep learning components such as convolutional layers, pooling layers, fully connected layers, and the UNet network were introduced. An explanation was provided on how to construct the training dataset, the UNet network used in this study, and how to set various training parameters. The network was saved after training. When the inversion was needed, the net was loaded and the algorithm could predict the result. Several theoretical models were designed for inversion, and the experiment results verified the reliability and effectiveness of the algorithm. The time of the deep learning inversion and the tranditional inversion was recorded. Building training set needed much time, but the time of deep learning inverison was much less than the tranditional inversion. The deep learning inversion is more efficient than the traditional inversion.