Deep Learning Applications in Seismic Interpretation

Deep learning has transformed seismic interpretation by automating the detection and delineation of subsurface structures with unprecedented speed and consistency. Convolutional neural networks (CNNs) now underpin workflows for fault identification, horizon picking, facies classification and structural attribute enhancement. Architectures such as encoder–decoder networks and U-Nets perform pixel-level segmentation, enabling interpreters to map complex reflection patterns at scale. Automated network design through data-driven search strategies has yielded lightweight models that fuse multi-scale features while substantially reducing computational demands. These advances support end-to-end interpretation from raw volumes to geological characterisation, facilitating faster decision-making in hydrocarbon exploration, geothermal assessment and carbon storage projects. Despite significant progress, challenges remain in securing large annotated field datasets, extending models across diverse basins and integrating deep learning with geological expertise. Emerging directions include transfer learning, semi-supervised training and interactive refinement tools that promise to bridge the gap between algorithmic predictions and interpreter knowledge.

Research from Nature Portfolio

A novel convolutional architecture discovered via automated search combines topological modules with multi-scale feature fusion units to classify seismic signals with minimal parameters. This model requires only 0.5 per cent of the storage volume of conventional CNNs and predicts reflectivity classes nearly eighteen times faster than established networks. Applied to the detection of Bottom Simulating Reflections, it demonstrated high sensitivity to gas-hydrate indicators while saliency mapping confirmed faithful capture of key seismic features. The efficiency and accuracy of this end-to-end approach herald a path towards rapid, low-cost interpretation of multiple seismic volumes.

Research from all publishers

A systematic review of deep learning for fault interpretation highlights the predominance of supervised CNNs, especially U-Net variants, trained on image-processing tasks. Drawing on over fifty studies and seventy datasets, it outlines performance gains in repeatability and speed, identifies bottlenecks such as data scarcity and domain shift, and proposes future priorities for collaboration between computer scientists and geoscientists.

A 3D CNN formulated horizon picking as a multiclass segmentation problem, using probabilistic outputs and masked loss functions to accommodate picking uncertainties. The network’s multi-scale design analyses volumes at different resolutions, allowing interpreters to fine-tune results interactively. Validation on field datasets demonstrated accurate horizon extension across complex reflectivity and facilitated rapid prestack domain application.

An open-source, expert-labelled field dataset underpinned the development of a deep CNN for automatic fault recognition. Following image augmentation and careful preprocessing, the network achieved performance comparable to human interpreters, even detecting subtle discontinuities. Case studies showed a reduction in interpretation time from months to hours, underscoring the practical impact of large-scale, high-quality training data and robust network workflows.

Deep Learning Applications in Seismic Interpretation publication trend

The graph below shows the total number of articles in deep learning applications in seismic interpretation across all publications each year (not limited to Nature Index journals).

Technical terms

Convolutional Neural Network (CNN): A deep learning model that applies convolutional filters to extract spatially localised features from seismic images.

U-Net: An encoder–decoder CNN architecture with symmetric skip connections, optimised for pixel-wise segmentation.

Segmentation: The process of partitioning seismic volumes into regions representing different geological entities or classes.

Seismic Horizon: A continuous reflection surface in seismic data that corresponds to a stratigraphic boundary.

Seismic Fault: A discontinuity in seismic reflections indicating displacement along a fracture plane.

Seismic Facies: Groups of reflection characteristics interpreted as distinct depositional or structural units.

References

1. Current state and future directions for deep learning based automatic seismic fault interpretation: A systematic review. Earth-Science Reviews (2023).
2. Extracting horizon surfaces from 3D seismic data using deep learning. Geophysics (2020).
3. Deep convolutional neural network for automatic fault recognition from 3D seismic datasets. Computers & Geosciences (2021).
4. Automated design of a convolutional neural network with multi-scale filters for cost-efficient seismic data classification. Nature Communications (2020).