Acoustic impedance prediction using an attention-based dual-branch double-inversion network

Seismic impedance inversion is a critical technique in reservoir characterization, which relies on high-precision and high-resolution results for accurate reservoir interpretation. In recent years, researchers have observed significant advancements in deep learning methods within this field. However, traditional single-branch semi-supervised networks, using raw seismic and well-log data, often fail to capture high-frequency weak signals, leading to low-resolution and inaccurate inversion results. Besides, the uncertainty of seismic wavelets further limits the application of these methods to field seismic data. To address these challenges, we proposed an Attention-based Dual-branch Double-Inversion Network for Impedance prediction (ADDIN-I). This dual-branch architecture overcomes the limitations of single-branch semi-supervised networks and enhances high-frequency weak signal detection in sequence modeling. By integrating attention mechanisms into the network, ADDIN-I significantly enhances the stability of learning high-frequency features, resolving the issue of unstable loss curves during training. Additionally, this research constructs a deep learning forward operator, enabling the adaptive fitting of seismic wavelets and facilitating the method’s effective application to field seismic data. ADDIN-I demonstrates exceptional performance in both model and field data applications. In tests using the Marmousi2 model, the method achieves a correlation coefficient of 0.9851, while in field data applications, it attains a PCC coefficient of 0.8827. These results provide compelling evidence of the method’s significant advantages in improving the accuracy and resolution of seismic impedance inversion, offering robust support for high-precision reservoir characterization.