Deep learning assisted 2D-to-3D seismic image reconstruction

Abstract

Due to the considerable expense associated with large-scale 3D seismic acquisitions, operators often rely on both legacy and new 2D seismic datasets for regional analysis and preliminary assessments during emerging frontier explorations. In this paper, we introduce an innovative deep learning-assisted workflow designed to transform 2D seismic data into pseudo-3D seismic volumes. This approach offers a cost-effective solution for risk mitigation prior to investing in full 3D surveys and rejuvenating the value of existing legacy and regionally acquired 2D seismic data. By integrating deep learning techniques, deterministic algorithms, and domain-specific geological understanding, we achieve high-quality 2D-to-3D seismic image reconstruction via the proposed workflow even with ultra-sparse 2D lines (line spacing > 5 km), without assuming any specific 2D line organization geometries. It consistently outperforms traditional methods based on geophysical signal processing by better preserving the original 2D line characteristics, maintaining realistic seismic image styles, as well as reducing directional interpolation artifacts. Furthermore, the proposed technique is substantially more economical than conventional geophysical methods. We demonstrate the effectiveness and versatility of this workflow are through applications to legacy 2D seismic datasets from diverse regions worldwide.