Search for a command to run...
Seismic velocity inversion remains a critical challenge in subsurface reservoir characterization, particularly within structurally complex salt and fault-dominated environments. This study presents a hybrid deep-learning framework that couples regression-based velocity prediction with geological mask-aware segmentation to achieve physically consistent, high-resolution P-wave velocity models. The network extends a conventional U-Net with dual output heads and a composite loss that integrates Mean Absolute Percentage Error (MAPE), Structural Similarity Index (SSIM), and a salt-weighted regularization term derived from domain priors. A curated dataset of 5000 synthetic shot gathers and paired ground-truth velocity models, each shaped (300 × 1259), was used for training. The data-engineering pipeline incorporated well-log extraction, geometry-aware normalization, and geology-based augmentation (elastic deformation, noise injection, depth shift). Model evaluation followed a multi-layered test plan comprising continuous-integration checks, data-integrity verification, well-log consistency, and robustness testing under trace corruption and masking scenarios. The proposed framework achieved a MAPE of 0.0337 on the hold-out dataset, reducing column-wise MSE at borehole positions by ≈ 20 %, and improving structural boundary fidelity relative to baseline regression U-Nets. Performance degraded by less than 5 % MAPE under 20 % synthetic trace corruption, demonstrating resilience to realistic acquisition noise. The integration of geology-driven constraints and automated validation workflows produced velocity models that are both numerically accurate and geologically interpretable. The results highlight a path toward interpretable, domain-aware seismic inversion and future foundation-model integration in reservoir digital twins.