Seismic Inversion & Mineral Exploration

DU-Net Architecture & SeismoLoss Explained

Core Idea: DU-Net reframes seismic inversion as a decoupled "background-plus-anomaly" problem, enabling specialized learning for each component.

Architecture:

• Shared Encoder: Processes input seismic shot gathers to extract a rich, multi-scale feature representation.
• Dual Decoders (Parallel): The encoded features are fed into two independent decoder branches:
  • Anomaly Decoder: Outputs a single-channel image (manom) representing the probability of a pixel belonging to an anomaly (a binary segmentation mask of the ore body).
  • Background Decoder: Outputs a single-channel image (mback) representing the reconstructed background velocity model.
• Fusion Layer: The outputs of the two decoders are combined via element-wise summation to produce the final, full velocity model prediction (mfull = mback + manom).

SeismoLoss (Composite Loss Function): This novel loss function supervises each of the three outputs: background, anomaly, and final combined predictions. Crucially, the loss for the background branch is masked to ignore the anomaly region. This prevents the background decoder from being penalized for inaccurate predictions inside the anomaly, allowing it to focus purely on the background and eliminating contradictory gradient signals.

Performance & Key Advantages

DU-Net demonstrates significant improvements over standard U-Net architectures for seismic inversion in mineral exploration:

• Superior Anomaly Localization: Achieves sharper, more accurately localized anomalous bodies (ore bodies) due to dedicated anomaly segmentation, mitigating blurring artifacts observed with single-path approaches.
• Accurate Background Reconstruction: Predicts the layered velocity structure of the host rock with high fidelity, preserving geological continuity and velocity trends.
• Reduced Blurring Artifacts: The dual-path design effectively filters out environmental features that would otherwise complicate simultaneous prediction, leading to a substantial reduction in blurring caused by interference from the background response (Figure 7).
• Tunable Performance: The SeismoLoss hyperparameters (a1, a2, a3) allow practitioners to balance the contributions of background, anomaly, and final predictions, adapting the model to specific dataset characteristics and geological complexities for optimal accuracy and training stability.

Future Outlook & Applications

Several promising directions for DU-Net's evolution and application include:

• Validation on Real Field Data: Essential for addressing challenges such as data noise, source wavelet estimation, and the gap between synthetic and real-world physics.
• 3D Extension: Expanding the architecture to handle 3D seismic volumes for generating full 3D velocity models of ore bodies, critical for comprehensive resource estimation and mine planning.
• Advanced Components: Integrating attention mechanisms or multi-scale feature fusion modules to better capture fine details and long-range dependencies within the network branches.
• Multi-Class Segmentation: Modifying the anomaly branch for multi-class segmentation to differentiate between various types of mineralized zones or lithological units based on their distinct velocity signatures.
• Uncertainty Quantification: Incorporating Bayesian methods or ensemble techniques to provide uncertainty estimates alongside velocity predictions, enhancing risk-aware decision-making in exploration.