Note, this repo is part of research that has been accepted in Frontiers Earth Science, please cite that paper with DOI (doi: 10.3389/feart.2021.659611). Full paper is here!
This repo contains code for modeling lithology and facies in core photo + well log datasets, using both deep learning / computer vision and traditional machine learning approaches.
For pre-processing of UK core image data, check out CoreBreakout.
Base requirements:
Python 3.xNumpy+Scipy+Matplotlibscikit-imagescikit-learntensorflow.keras
Individual scripts and notebooks may require some other libraries.
Use pip to do a local develop mode install:
$ pip install -e path/to/coremdlrThe coremdlr module contains a number of submodules:
coremdlr.config : settings and default paths / labels / dataset args / viz properties
coremdlr.datasets : loading / preprocessing / generating data
coremdlr.models : hierarchical set of generic model classes
coremdlr.networks : tf.keras network construction functions
coremdlr.layers/.ops : tf.keras custom layers and tensor operations
coremdlr.viz : plotting data and analysis (e.g., confusion matrices)
The final experiments for the paper mostly took place in experiments, and more specifically the notebooks_* subdirectories.
The notebooks folder contains assorted notebooks and a figures subdirectory in which paper figures were generated.
Current data consists of 12 UK Contiential Shelf wells from Q204 and Q205. Please check out the data folder for more information and licensing. The full dataset is here: https://figshare.com/articles/dataset/UKCS_Q204_Q205_Subsurface_Data_products_for_Machine_Learning_Study/14265689 including the images.