This code demonstrates fault probability prediction on 3D seismic images using 3D Shallow U-Net model. A large number of synthetic 3D seismic images are generated by randomly choosing parameters to represent geological deformations, fault strike/dip, number of faults, central frequency of wavelet, etc. This techinique has advantage against human labeling as fault labeling can be done automatically during data generation process.

GPU: NVIDIA GeForce GTX 1080 Ti
Model architecture: Shallow 3D U-net
Training Data: 200
Validataion Data: 20
Batch size: 3 (Data Augmentation)
Data Augmentation: z-axis rotation (Randomly choose from 0, 90, 180, 270 deg.)
Feature size: 128 x 128 x 128

Netherlands Offshore F3 Block (dGB Open Seismic Repository)
https://terranubis.com/datainfo/Netherlands-Offshore-F3-Block-Complete

FaultSeg3D: Using synthetic data sets to train an end-to-end convolutional neural network for 3D seismic fault segmentation, Xinming Wu et al., Geophysics, 2019