Deep learning-driven multi-frequency seismic inversion for enhanced thin-layer stratigraphic characterization

Acoustic impedance derived from seismic data through inversion algorithms serves as a critical parameter for stratigraphic characterization. However, the limited resolution of seismic data poses significant challenges for accurately predicting acoustic impedance. In this study, we demonstrate that multi-frequency band seismic signals, obtained via continuous wavelet transform (CWT) based on a Ricker wavelet, provide a more precise representation of thin-layer acoustic impedance variations than the original seismic signal. Additionally, the Multi-Head Self-Attention Mechanism enables flexible, nonlinear mapping between parameters by capturing contextual relationships. Therefore, integrating a Multi-Head Self-Attention Mechanism with multi-frequency band seismic signals is essential for improving impedance prediction. To leverage these advantages, we introduce the CWT-CNNTrans algorithm. This method combines stratigraphic encoding data with multi-frequency seismic signals, which are processed by multi-scale convolutional neural network (CNN) modules to extract essential stratigraphic features as multi-scale geological priors. These priors are then integrated into a Transformer architecture, enhancing the accuracy of acoustic impedance predictions and improving the characterization of thin-layer structures. Comparative experiments confirm that incorporating multi-scale geological priors and optimizing the network architecture significantly improves prediction accuracy, enhances stratigraphic continuity, and mitigates overfitting in scenarios with limited geophysical data.