Abstract: Deep learning-based fault detection methods typically rely on pre-interpreted fault labels for network training. However, practical applications face significant challenges due to the labor-intensive nature of manual fault interpretation and insufficient labels for the effective training of a deep learning model. The utilization of synthetic labels is an alternative. However, structural and lithologic variations in different regions lead to notable discrepancies between synthetic and field seismic data, which severely limit the generalization of a synthetic data-trained model for fault identification. To address these limitations, this study proposes a transfer learning approach employing a domain-adaptive neural network for cross-domain fault detection. A domain discriminator is designed in the U-Net architecture to simultaneously process both synthetic and real seismic data. A feature extractor is designed to learn domain-invariant representations from both data types, while the domain discriminator engages in adversarial training to distinguish the data source of extracted features. Through iterative optimization, the network eventually achieves domain confusion when the discriminator cannot reliably identify the origin of input features. At this equilibrium state, the model effectively captures the shared characteristics between synthetic and real seismic data, enabling the robust generalization to field data for accurate fault detection. Comparative experiments using seismic data from Netherlands North Sea F3 block and western China demonstrate the enhanced accuracy of our method compared to conventional deep learning approaches and its good performance of addressing domain shifts in seismic interpretation.