<![CDATA[The Asri Basin, located in the Java Sea, Indonesia, is a significant hydrocarbon province with regions that remain underexplored. The available legacy seismic data, however, are limited in quality, particularly due to their narrow frequency bandwidth and the absence of low-frequency components. This limitation poses a significant challenge for advanced seismic imaging techniques such as Full Waveform Inversion (FWI), which rely low-frequency data to generate accurate and reliable subsurface models. This study aims to reconstruct the missing low-frequency (<10 Hz) components from the band-limited seismic data to enhance the applicability of FWI. A real-data-driven, self-supervised learning approach for low-frequency extrapolation is implemented to address this challenge. Using a modified U-Net architecture, the framework is trained directly on the available band-limited seismic data, eliminating the need for synthetic or labeled datasets. The self-supervised workflow employs a frequency-specific masking strategy that enables the model to learn and predict the missing low-frequency content from higher-frequency inputs. The results demonstrate that the proposed method effectively recovers low-frequency signals, achieving accurate reconstruction down to <5 Hz, reducing residual amplitudes compared to conventional methods, and preserving the mid-to-high frequency spectrum. This approach provides a promising solution for overcoming data limitations and mitigating cycle-skipping issues in FWI applications within the Asri Basin and comparable geological settings.]]>
