ResearchPod Summary
Acoustic impedance imaging (AII) is a critical task in subsurface exploration, yet it is severely ill-posed due to the scarcity of labeled well-log data (often <1% of total traces), the unknown nature of the seismic wavelet, and the band-limited nature of seismic observations. The authors seek to develop a deep learning framework that can effectively leverage abundant unlabeled seismic data while maintaining physical consistency without relying on inaccurate prior wavelet assumptions or complex, unstable auxiliary networks.
The authors introduce RD-SCL (Regularized Deconvolution Semi-Supervised Cross-Learning). The framework centers on a differentiable, closed-form first-order Tikhonov deconvolution operator that dynamically estimates the latent seismic wavelet in the frequency domain during training. This operator provides a physics-guided feedback loop that avoids the need for fixed wavelet priors. Building on this, the authors implement a symmetric cross-learning strategy that enforces consistency between predictions on labeled and unlabeled traces. By treating the wavelet as a latent bridge variable, the model transfers information between labeled and unlabeled data, effectively regularizing the training process and mitigating the overfitting risks associated with few-shot learning.
RD-SCL consistently outperforms state-of-the-art supervised and semi-supervised methods on the SEAM and Marmousi 2 benchmarks. It achieves higher signal-to-noise ratios (SNR) and structural similarity (SSIM) while maintaining lower mean absolute error (MAE) and mean squared error (MSE). Notably, the framework demonstrates superior training stability, evidenced by smaller error bounds across multiple random seeds. The model is also computationally efficient, utilizing only 56.5k learnable parameters, making it a practical solution for complex subsurface imaging tasks.
This research provides a robust, physically consistent, and efficient solution to the few-shot problem in seismic inversion. By eliminating the need for auxiliary networks or rigid wavelet priors, the proposed method simplifies the training pipeline and improves the reliability of subsurface characterization in scenarios where labeled data are extremely limited and expensive to acquire.
AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.