Using synthetic seismic data
Author: Suman Gautam
This is an ongoing project on building a deep learning learning models to accurately detect fault from seismic images. You may also find my blogpost posted in Towards Data Science for an early result of this project at : Medium blogpost
Seismic images provides a structural snapshot of Earth's subsurface at a given time. Because of the many geological and tectonic processes, the Earth's layers are folded and faulted. A seismic fault is basically a crack in the rock layers along with the block of rocks/sediments moves. If the displacemnt is large enought, they are visible in the seismic data. These faults are of primary interest in hydrocarbon exploration because they can trap oil in places or may act as conduit for oil to escape. They also pose significant hazard in drilling process and therefore a correct identification and mapping of faults is critical.
The synthetic data were taken from FORCE competition provided by XEEK.ai. To get more information, follow the link at XEEK.ai
An example of a seismic images with fault overlay:
The introductory data exploration will be provided in a separate notebook while the machine learning models will be provided under each type of the machine learning framework name. The current model uses a simple U-Net framework using PyTorch implementation:
The example below are taken randomly from 3 different epochs. Within 20 epochs, the model starts to pick faults effectively.
See the Data Exploration in the Data Exploration See the UNet model in the U-Net_Model
├── Overview
├── images
├── notebook_versions
├── Data_exploration
├── Pytorch_based_UNet_Model_versions.ipynb
└── README.md