Enhancing FWI Convergence through Self-Supervised Low-Frequency Extrapolation of Legacy Marine Data: A Case Study from the Asri Basin, Indonesia

Marine seismic data often lacks low-frequency content, limiting its usefulness for advanced modelling and inversion techniques such as Full Waveform Inversion (FWI). In this study, we applied a self-supervised deep learning (SSL) framework for low-frequency extrapolation, tested on field data from the Asri Basin, Indonesia. Our approach uses a modified U-Net architecture and a two-stage learning scheme, namely: synthetic warm-up followed by iterative data refinement (IDR), to train directly on unlabelled band-limited seismic data. When used as input for a source-independent FWI (SI-FWI) method, the low-frequency extrapolated data significantly improves inversion results, demonstrating better recovery of deep velocity structures and reduced cycle skipping compared to inversions using band-limited data as inputs. The integrated SSL and SI-FWI workflow provides practical value for reprocessing legacy datasets lacking both low-frequency energy and source wavelet information. This method offers a data-driven, label-free approach for expanding the spectral bandwidth of field seismic data and further improving inversion convergence.