Deep Learning-Based Estimation of Petrophysical Properties for Unconventional Shales from SEM Images

The use of machine learning algorithms to process images has significantly enhanced our ability to extract meaningful information in diverse areas such as weather forecasting, medical imaging, satellite image analyses, and driverless vehicle navigation. Borrowing on these technologies, we demonstrate the use of deep-learning features acquired from an image classifier to predict petrophysical properties in unconventional shales. By directly analyzing images, that can easily be acquired from drill cuttings or even rock chips in highly friable formations which make routine core analyses challenging, we are able to provide very rapid turnover at a fraction of the cost of lab-based methods that are more rigorous, but also more resource-intensive. Our dataset consists of over 8000 images from 22 shale plays across North and South America and Europe. We first construct a convolutional neural network-based (CNN) classifier trained to recognize the source play associated with each image. In this work, we use the well-known ResNet50 architecture. We then extract the deep learning features from the model after training, and relate these to various petrophysical properties such as porosity, mineralogy and elastic properties such as Young's Modulus. We obtain a very high degree of precision, recall and accuracy over all 22 classes with the CNN-based classifier. This demonstrates that the CNN is able to capitalize and exploit microstructural differences between each of the 22 plays to identify the source of an image. However, the misclassifications that did occur point to similarities in the microstructure of two or more plays. In situations such as these, the different plays can be considered to be digital geologic analogs that can provide useful information for completion and well design as well as the planning of drilling operations. By extracting the deep-learning features from this trained CNN-based network, we were able to show remarkable predictive capability for petrophysical properties such as Young's Modulus, porosity and mineralogy. Prior approaches to predict petrophysical properties from images have relied on image segmentation (identifying the pixels associated with grains, pores and other features in an image) and then relating them to petrophysical properties. Image segmentation, for the complex microstructure associated with the SEM images can take upwards of two-three weeks per image, making the traditional approach prohibitively expensive. In this work, by using higher-level deep learning features from a trained classifier, we circumvent this problem and present a more robust approach to image-based estimation of petrophysical properties.