ScoreInver: 3D seismic impedance inversion based on scoring mechanism

Abstract

In recent years, the introduction of deep learning has significantly advanced the field of seismic impedance inversion (SII). However, existing methods generally rely heavily on large volumes of expensive well logs, limiting their broader applicability, particularly in scenarios beyond mature or synthetic data. To reduce the dependency on well logs in deep learning-based SII research, this paper proposes a 3D data-driven SII approach based on the pseudo-labeling strategy in semi-supervised learning, termed the ScoreInver framework. The core of the ScoreInver framework lies in the design and training of a Scorer, which can precisely select high-quality pseudo-labels from seismic data, thereby enhancing data utilization and extracting geological information while minimizing the need for extensive well logs. This framework is highly versatile, capable of seamless integration into various semi-supervised learning architectures. Experimental results demonstrate that, when using only 9 well logs as training samples on synthetic data, the semi-supervised learning architectures based on the ScoreInver framework significantly outperforms traditional supervised learning methods, with improvements of 3.3% in Structural Similarity Index (SSIM) and a reduction of 29.1% in Mean Squared Error (MSE). Moreover, tests on field data reveal that the application of the ScoreInver framework yields more robust and reliable results, further validating its effectiveness and practicality in real-world exploration environments.