Client Background
Our client is an American energy corporation and a leading oil company involved in petroleum exploration and production worldwide.
Business Challenge
Seismic interpretation typically requires geophysicists to manually label seismic data volumes based on visual inspection. This labeled data informs critical business decisions and feeds directly into downstream processes.
Our client wanted to address two persistent challenges in this workflow:
- Time consumption: Manual seismic data labeling can take weeks or months to complete.
- Interpretation errors: Manual interpretation is prone to human error and depends heavily on individual expert knowledge.
Project Description
The goal was to test the feasibility of a deep learning (DL) approach for segmenting geological volumes and assigning rock formation labels to each pixel in a given volume. SoftServe evaluated the approach across several deep neural network architectures, using training data manually labeled by the client's in-house experts.
Our team applied state-of-the-art segmentation architectures, including U-Net and FPN, for automatic geological data segmentation. The resulting proof of concept (PoC) tool takes unlabeled SEG-Y file volumes as input and generates labeled SEG-Y cubes as output. Each 3D cube was split into 2D slices, allowing the team to train networks effectively even with limited 3D data, using transfer learning.
The solution also incorporated generated statistics and predicted classes into an ontology tree built from volume predictions. This tree maps the relationships between geo-bodies within a cube: the region itself serves as the root node, while child nodes represent the corresponding geo-bodies identified in the volume. The same structure can be built for any geo-body to examine further relationships or distinguishing features.
Technology Stack
- GPU-powered GCP instance for model training and data processing
- PyTorch, an open-source deep learning library
- GitLab repository for source code management
- Jupyter Notebook for data visualization
Value Delivered
The PoC proved that seismic data interpretation using computer vision is feasible, along with an end-to-end pipeline for model training and 3D seismic volume interpretation. The approach can be extended to interpret other types of seismic geo-bodies.
- Deep learning enables analysis of large volumes of geological data and helps uncover relationships across multiple data types simultaneously.
- The solution interprets large volumes of seismic data significantly faster than manual methods, while reducing the risk of human bias.
- Faster, more accurate seismic interpretation supports quicker decisions on future exploration activity and enables more precise basin models for predicting production and field profitability.
This project shows what's possible when deep learning is applied to a problem experts have solved manually for decades. The client's geophysicists get a faster, more consistent path from raw seismic data to actionable insight, and a framework that extends to other geo-body types as the approach scales beyond proof of concept.
