Abstract: The X-discontinuity in the upper mantle of the Earth’s interior, typically located at depths of approximately 250–350 km, is critical for unraveling mantle composition and dynamics. However, its global distribution and underlying formation mechanisms remain poorly constrained. Conventional methods for identifying this discontinuity using teleseismic P-wave receiver functions rely heavily on post-stacking empirical interpretation, which is inefficient and unsuitable for processing large volumes of seismic data. To address the challenges posed by massive datasets and the scarcity of manually labeled samples, this study proposes a transfer learning strategy termed “pretraining on synthetic data followed by fine-tuning with real data”. We construct a convolutional neural network-based binary classification model to determine the presence or absence of the X-discontinuity. The procedure is as follows: First, synthetic receiver function datasets—with and without the X-discontinuity—are generated through forward modeling based on three classical velocity models (AK135, IASP91, and PREM). Noise augmentation and temporal trimming are applied to improve model generalization. Second, high-quality observed data are selected from global seismic networks and manually annotated to create a real dataset. Finally, a two-phase training approach is adopted, where the model is sequentially trained on the synthetic and real datasets to achieve automated detection of the X-discontinuity. Experimental results show that the trained model achieves an accuracy of approximately 90% in classifying receiver function images. In determining the presence of the X-discontinuity beneath seismic stations, the model exhibits about 80% agreement with results derived from conventional methods, indicating its effectiveness in automated detection. Based on this approach, subsequent research will systematically investigate the global distribution of the X-discontinuity.