Lateral constrained multi-trace seismic inversion based on deep learning

Seismic inversion is a widely used method for exploring and characterizing subsurface geological structures, especially in the context of oil and gas exploration. The data-driven approach exemplified by deep learning (DL) circumvents the need for pre-defined physical system and is used to solve seismic inversion problems. However, the DL-based inversion method is sensitive to noise and is typically performed trace-by-trace. When the inversion results are aggregated into a 2D image, the lateral continuity of the final output is often inadequate, impacting subsequent interpretation and evaluation. In contrast, conventional DL-based multi-trace seismic inversion typically treats multi-trace seismic data as input without fully accounting for the coupling relationships between neighboring traces. Therefore, we propose a lateral constrained multi-trace seismic inversion method based on DL to enhance the continuity and geological reliability of the inversion results. Given the similarities among neighboring traces, the method employs adjacent multi-trace seismic data as input to the network and designs multi-trace coupling constraints to ensure the lateral consistency of the prediction outcomes. Moreover, physical laws and low-frequency prior information are incorporated into the network training process to mitigate the dependence of data-driven methods on large amounts of training data. The effectiveness of the proposed method in enhancing both the lateral continuity and accuracy of inversion results is demonstrated by applying it to synthetic and real datasets, and comparing the results with those of conventional DL-based single-trace and multi-trace inversion methods.