2019 AAPG Annual Convention and Exhibition:

Datapages, Inc.Print this page

Automated Lithology Prediction From Core Images and Well Log Data Using Machine Learning Models: A Case Study From the Greater Schiehallion Area, West of Shetland, United Kingdom

Abstract

Machine learning models have been used previously to predict facies from well log data, most notably from the Society of Exploration Geophysicists 2016 machine-learning contest. These techniques have recently been improved, but still primarily only utilize digital well log information, which like any remotely sensed measurements, have resolution constraints. Core, on the other hand, is the only borehole data that is true to geologic scale and heterogeneity. However, core description and analysis are time-intensive and therefore most core data are not utilized to their full potential. We apply machine-learning models to core data to distinguish the core from its box/tray, align the core true to wellbore depth, and analyze the rock color, texture, and brightness. This image data allows for higher-resolution interpretation compared to well log data alone, and enables geoscientists to interpret hundreds of meters of core in hours rather than weeks.

Quadrant 204 of the United Kingdom continental shelf (UKCS) provides a large publicly available dataset to test this workflow. We selected 18 wells with logs and core from the Schiehallion, Foinaven, Loyal, and Alligin fields, a submarine fan system that contains turbidites and hybrid-event beds. We interpreted training data from over 150m of core at the sub-centimeter scale using four lithology-based facies appropriate for this submarine fan and channel setting that contains deposits of varying flow rheology: sandstone, clay-prone sandstone, sandy mudstone, and mudstone. Using these lithologic interpretations, we train a variety of supervised machine-learning models to predict lithology in core photos at a ~0.5cm scale. Using core image data alone, our model predicts the correct lithology with more than 70% accuracy (i.e., the predicted lithology output from the model is the same as the interpreted lithology). Our current focus is on increasing the volume of labeled data, testing the viability of predictions on un-cored intervals, and experimenting with multimodal neural network models that can capture subtle interactions between well logs and image features.

Machine learning can unlock warehouses full of high-resolution data in a multitude of geological settings. The workflow developed for this study can be applied to other areas where large amounts of legacy data are available.