Dual-Input Autoencoder–Based Method for Enhancing Repeatability in Time-Lapse Seismic Data

Time-lapse seismic exploration is a crucial tool for monitoring subsurface changes, with applications in hydrocarbon production assessment and CO₂ migration tracking in carbon capture and storage (CCS) projects. However, accurate detection of subsurface changes is challenging due to nonrepeatable effects. Traditional cross-equalization techniques and advancements in machine learning have been aimed at addressing these challenges, but often face limitations, particularly in handling real-world variables. We present a dual-input autoencoder (DIAE) designed to improve the repeatability of time-lapse poststack seismic data by simultaneously processing baseline and monitoring datasets. The DIAE model introduces a custom loss function that combines reconstruction loss and repeatability loss, enabling flexible control between signal preservation and noise reduction. We evaluate the proposed method using two real-world datasets: the Enfield oil field dataset from Western Australia and the Sleipner CCS dataset from Norway. The results demonstrate that DIAE effectively suppresses nonrepeatable noise while minimizing signal distortion. By achieving a practical balance between repeatability enhancement and signal preservation across different geological settings, DIAE presents a robust solution for improving time-lapse seismic monitoring in both production and CCS environments.