Multi-modal seismic impedance inversion using a semi-supervised neural network with dual physics constraints

Seismic impedance inversion is a crucial process for subsurface characterization in seismic exploration, offering insights into rock properties by converting seismic reflection data into impedance models. Data-driven deep learning methods have been widely applied to seismic data processing since they can provide better performance compared to traditional methods. However, current deep learning techniques also face challenges in powering seismic inversion due to problems such as poor continuity, the image assimilation of seismic data, and the lack of labeled data. To address these limitations, we propose a novel semi-supervised seismic impedance inversion neural network by integrating multi-modal attention mechanisms and dual physics constraints. Our approach leverages multi-modal attention mapping to transform seismic data into multiple domains, enabling the network to capture different features and improve inversion accuracy. By incorporating both seismic reflection data and well-log information, the semi-supervised framework learns effectively from both labeled and unlabeled data. The dual physics constraints, grounded in wave propagation and guided filtering mechanism, further guide the network towards physically consistent solutions and improve the continuity of the predicted impedance models. Experimental results on synthetic and field data demonstrate that the proposed method outperforms traditional deep learning seismic inversion techniques and provides more reliable impedance models. This approach highlights the potential of combining multi-modal attention mechanisms and physics-based constraints in deep learning inversion methods to advance subsurface imaging and resource exploration.