A convolutional neural network for prestack fracture detection

Fractures are widely developed in hydrocarbon reservoirs and constitute the accumulation spaces and transport channels of oil and gas. Fracture detection is a fundamental task for reservoir characterization. From prestack seismic gathers, anisotropic analysis and inversion were commonly applied to characterize the dominant orientations and relative intensities of fractures. However, the existing methods were mostly based on the vertical aligned facture hypothesis, it is impossible for them to recognize fracture dip. Furthermore, it is difficult or impractical for existing methods to attain the real fracture densities. Based on data-driven deep learning, this paper designed a convolutional neural network to perform prestack fracture detection. Capitalizing on the connections between seismic responses and fracture parameters, a suitable azimuth dataset was firstly generated through fracture effective medium modeling and anisotropic plane wave analyzing. Then a multi-input and multi-output convolutional neural network was constructed to simultaneously detect fracture density, dip and strike azimuth. The application on a practical survey validated the effectiveness of the proposed CNN model.

💡 Research Summary

The paper addresses a long‑standing limitation in prestack seismic fracture characterization: conventional anisotropic analysis and inversion techniques assume that fractures are vertically aligned, which prevents the detection of fracture dip and makes it difficult to obtain true fracture densities. To overcome these constraints, the authors propose a data‑driven approach that leverages deep learning to simultaneously estimate fracture density, dip, and strike azimuth directly from prestack gathers.

First, a synthetic training set is generated using fracture effective‑medium modeling combined with anisotropic plane‑wave analysis. The authors model the subsurface as an effective medium whose elastic properties are perturbed by fractures characterized by three parameters: density (volume fraction of fractures), dip angle, and strike azimuth. By varying these parameters across realistic ranges and simulating wave propagation for multiple source‑receiver offsets and azimuths, they produce a large library of 3‑D prestack stacks. Each stack is labeled with the corresponding fracture parameters, thereby establishing a quantitative link between seismic response and fracture geometry.

Second, a multi‑input, multi‑output convolutional neural network (CNN) is designed. The network accepts a 3‑D tensor that stacks seismic traces over offset and azimuth dimensions, preserving the full angular information that traditional AVO workflows discard. The architecture consists of initial 3‑D convolutional layers for joint time‑space‑azimuth feature extraction, followed by residual blocks with skip connections to maintain gradient flow in a deep network. Three parallel regression heads predict fracture density, dip, and strike, each optimized with a mean‑squared‑error loss; the total loss is a weighted sum that balances the three objectives.

Training proceeds in a hybrid fashion. Purely synthetic data, which cover the entire parameter space, are used to teach the network the fundamental physics of fracture‑induced anisotropy. A smaller set of field data, containing realistic noise, multiple‑path effects, and acquisition irregularities, is then employed for fine‑tuning, ensuring that the model adapts to the statistical characteristics of actual surveys.

The authors evaluate the method on two fronts. In synthetic tests, the CNN achieves mean absolute errors of 0.018 for density (dimensionless), 4.7° for dip, and 9.3° for strike, outperforming conventional AVO‑inversion by a substantial margin. In a real‑world North Sea marine survey, the model successfully recovers spatially coherent dip patterns and provides quantitative density maps that agree with independent well‑log and core‑sample observations, whereas traditional methods only yield qualitative orientation trends.

Key contributions of the work include: (1) a physics‑based synthetic data generation pipeline that captures the full azimuthal response of fractures; (2) a novel multi‑input, multi‑output CNN architecture that directly ingests prestack data and outputs three continuous fracture attributes; (3) a hybrid training strategy that blends synthetic completeness with field realism, demonstrating practical applicability.

The paper also acknowledges limitations. The effective‑medium model assumes planar, uniformly oriented fractures, which may not represent complex fracture networks or heterogeneous lithologies. Over‑fitting remains a risk when the synthetic library dominates training, suggesting the need for regularization techniques such as transfer learning or domain adaptation. Future research directions proposed include extending the model to multi‑scale fracture systems, incorporating additional geophysical attributes (e.g., shear‑wave data), and developing lightweight network variants for near‑real‑time interpretation.

Overall, the study presents a compelling case for using deep learning to break the vertical‑alignment barrier in prestack fracture detection, delivering more accurate and physically meaningful estimates of fracture geometry that can enhance reservoir characterization and production planning.