Codes for running UNet on RTM images seen in Jreij, S., Trainor-Guitton, W., Morphew, M., and Lim Chen Ning, I., 2021, The Value of Information from Horizontal Distributed Acoustic Sensing Compared to Multicomponent Geophones via Machine Learning: Journal of Energy Resources Technology, v. 143, doi: https://doi.org/10.1115/1.4048051.
U-net for fault detection below the Brady National Laboratory Quickstart guide by Michael Morphew (May 2019), modified by W. Trainor-Guitton (Nov 2020). The generated .npy are available upon request (wtrainor@gmail.com, too big to put on github).
DATA PREPROCESSING Step 1. Run your simulation in Mio, generating many hh files Step 2. Copy your data over to a new folder for safety, then run copy_files.bash to create the @hh binary files Step 3. Place the NotebookReadsInHeaderBinarydata.ipynb notebook and the read_in_rsf.py files into the directory with all the hh and @hh files Step 4. Edit the jupyter notebook with your preferred output names for the .npy files that the notebook spits Step 5. Run the notebook and generate the .npy files that will be used by the notebook. There should be four: geophone, DAS, combined, and fault labels Step 6. (optional) The Data_Checker.ipynb notebook can be used to compare data files and plot data to ensure that your data is what you think it is
RUNNING UNET Step 1. Start the Unet_Faults.ipynb notebook Step 2. Run the imports and definition cells Step 3. Edit the data = and labels = cell to import your data and labels Step 4. Run the rest of the cells up to the callbacks cell, editing model parameters as you desire (adding data augmentation, changing dropout, etc.) Step 5. Change the ModelCheckpoint argument to name your model Step 6. Change your batch size and epochs in the results = ... cell (epochs should be 50-100 to ensure training stabilizes) Step 7. Run the results = ... cell to perform the training Step 8. Run the rest of the notebook to plot your training and testing results
RESTARTING AN OLD MODEL Step 1. Start the Unet_Faults.ipynb notebook Step 2. Run the notebook up to and including the model initiation cell (model = create_unet()) Step 3. Skip the next couple cells as those cells perform the training Step 4. Uncomment and edit the model.load() cell to load in the weights you want. Step 5. Run the rest of the notebook as normal.
RUNNING UNET-CV Step 1. Run the notebook as you would the Unet_Faults notebook up until the K fold split cell (the big cell that performs CV) Step 2. Change the ModelCheckpoint argument to name your models generated from each fold Step 3. Run the rest of the notebook up until the last cell. Step 4. Change the name of the CSV file that saves the CV statistics to your preferred file name