Towards flexible demultiple with deep learning

Deep learning (DL) methods have demonstrated promising advancements in seismic demultiple, addressing issues of traditional workflows. However, a key challenge is the limited flexibility of DL solutions in the demultiple process. Once a DL model has been trained, it produces one demultiple solution for a given input data. However, interpreting seismic events as multiples or primaries is often subjective. Moreover, multiple discrimination in Common Depth Point (CDP) domain relies on the accurate Normal Moveout (NMO) velocity estimation. To address this, we propose a supervised DL training method for demultiple based on moveout discrimination in the CDP domain. Our novel approach generates several multiple models based on moveout discriminations for a given input CDP gather, enhancing flexibility without additional computational costs. We validate the generalization ability of the Convolutional Neural Network (CNN) trained with the proposed methodology on synthetically generated data and on field datasets.