Abstract

Seismic data acquisition often results in missing traces due to limitations in survey design, equipment failures, or obstructions. To address this, a deep learning-based interpolation framework was developed using a modified UNet architecture optimized with a hybrid loss function combining L1 loss, Structural Similarity Index Measure (SSIM) loss, and Frequency loss. The model was trained on synthetic shot gathers from the BP benchmark dataset, with 50% of traces randomly removed to simulate data with missing traces. The trained model was tested on both synthetic and real seismic gathers to evaluate its interpolation performance. The model performed well on both synthetic and real dataset. Spectral domain analysis, including amplitude and frequency–wavenumber (FK) spectra comparisons, confirmed excellent recovery of both spatial and spectral content. The results emphasize the model’s robustness on synthetic data and its limitations on real-field data, underlining the need for incorporating diverse field datasets during training to enhance generality of the model.