Transofrmer Based Low-Frequency Extrapolation: A Novel Approach to Fine-Tuning

Low-frequency data are crucial for Full Waveform Inversion (FWI), as they provide the large-scale subsurface information necessary to steer inversion toward a global minimum. However, acquiring these low frequencies in field environments is particularly challenging due to seismic source limitations, sensor sensitivity, and environmental noise. To address these challenges in field data acquisition, we propose a deep learning framework that combines a U-Net architecture with a transformer backbone to reconstruct low-frequency components from high-frequency seismic data. Unlike traditional trace-bytrace methods, our approach processes all seismic traces simultaneously, thereby capturing both local features and long-range dependencies. Our framework employs a fine-tuning strategy that adapts a pretrained model specifically on field data, ensuring improved performance and robustness in practical seismic applications.