Accelerate Machine Learning processing time from weeks to minutes The Machine Learning seismic facies prediction workflow we published in the March 2021 issue of First Break, ha...
Accelerate Machine Learning processing time from weeks to minutes
The Machine Learning seismic facies prediction workflow we published in the March 2021 issue of First Break, had one practical problem: it took weeks to apply the trained Lenet model.
This is because an “image-to-point” model only predicts the central sample position from a block of seismic data surrounding the evaluation point. To reduce processing time, we had to limit the application area. Here, we present an extended workflow that allowed us to process the entire volume in a matter of minutes. The modified workflow consists of three-steps with steps 1 and 2 as described in the FB paper.
Step 1: We use OpendTect’s Thalweg tracker to create labeled point sets for 8 seismic facies classes. The shape and relative positions of the tracked bodies are the main interpretation criteria for labeling the facies classes.
Step 2: We train a Lenet CNN to classify the seismic input cubelets extracted at each labeled position. We apply the trained network in a small area.
Step 3: We train a 3D Unet on the Lenet output. The Unet transforms a large block of seismic data (128x128x64 samples) into a similar-size block of seismic facies predictions. Training the model on 11K+ examples took 10 hours. Application of the trained Unet was a matter of a few minutes.
The movie shows the three steps we took to segment the seismic response into 8 seismic facies classes. The dataset is from the Maui in New Zealand.
Author: Paul de Groot
#MachineLearning #OpendTect
