Owing to the limitations of surface conditions, the distribution of earthquake station arrays, even dense arrays, is uneven and spatially irregular. The station intervals are too large with respect to migration algorithm requirements. Therefore, the regularization of irregular station data is an important preprocessing step before imaging. Several methods, such as the curvelet transform method based on the sparse transform, have been developed for regularizing teleseismic data. In this study, we present a novel deep learning (DL) approach for teleseismic waveform regularization in 2D surveys. We designed an earthquake waveform regularization network (EWR‐Net) based on a deep generative model and residual network, consisting of a transposed convolution block, convolution block, and full connection block. The convolution block was variable and adjusted according to different data complexities to improve adaptability. The network was able to capture complex mapping between station locations and waveforms, and could be used to regularize both randomly and regularly sampled data without spatial smoothing. Unlike other DL methods, the EWR‐Net was trained and used for each event. It was trained using recorded teleseismic waveforms at irregular stations, and was then used to predict waveforms at regular stations. To avoid overfitting, L2 regularization, dropout in the full connection block, and early stopping were employed. The test results on both synthetic and field data showed that EWR‐Net could generate more accurate earthquake waveforms at virtual stations than the curvelet method. Reverse‐time migration imaging tests using regularized data demonstrated the feasibility of the proposed method.