State-of-the-art seismic modeling reached a point where we can model various folding/faulting structures, geobodies with unique characteristics, and realistic noises. The labeling is always on point, and deep learning has to somehow benefit from these diverse fake datasets.
Deep learning will reduce its attention to high-level representation when given a lot of generated data with different seismic features; It's similar to humans ignoring noises or other irrelevant characteristics when looking for faults, horizons, or geobodies. Whereas minuscule differences in manual labeling can derail the neural network performance; Moreover, the diversity of data is limited only to that project, so we might forget about general applicability to other projects.
The MS dissertation, "Nam Pham, 2019, Automatic channel detection using deep learning," explains how a synthetically generated dataset can solve the issue of channel detection on real seismic data. He provides results on the Browse basin (offshore Australia) and Parikhaka datasets.
Link for the LinkedIn post