Drill Cuttings Properties Prediction Using Deep Learning

Abstract:

During drilling a well, drill cuttings and fluids are constantly retrieved at the surface. As it passes over a shale shaker, cuttings are separated from the drilling mud and further sampled for an analysis. Information on drill cuttings analysis can be used to determine mineralogy, lithology and grain sizes of the formation that is currently being drilled. The process of characterizing of drill cuttings at a wellsite can be automated by utilizing recent developments in image recognition, machine learning and deep learning.

In this work we built the training dataset by collecting digital photographs of drill cuttings and corresponding laboratory measurements. We evaluated various deep neural network architectures for data training and measured the prediction performance as a percentage of characteristics correctly predicted by the deep neural network in comparison with real laboratory analysis. We further tested the accuracy and performance of various algorithms to predict drill cuttings properties from their digital images from a new well. Our results show that the neural network approach for predicting properties from images can perform as accurate as a manual characterization by an expert and be a complimentary tool at the wellsite. Model performance could be further improved by collecting and utilizing more data of digital images of drill cuttings.