Lithology Prediction from Drill Cutting Images Using Convolutional Neural Networks and Automated Dataset Cleaning

Abstract

Abstract The task of automating the detection of lithology in drill cuttings is an essential aspect of reservoir engineering. As a wellbore is drilled, the rotating bit breaks down the rocks at the bottom of the hole, and these fragments are then transported to the surface through the drilling mud. These fragments are separated from the mud by a shaker upon reaching the surface, enabling the mud to be recycled. The leftover rock fragments, known as drill cuttings, can provide a wealth of information about the geology of the wellbore, drilling speed, and the oil and gas content in the rocks. Once the cuttings are cleaned of the drilling mud, their various properties can be examined using a range of techniques. The importance of automating this analysis process lies in its ability to rapidly evaluate the drilling procedure and predict possible emergencies. Moreover, an accurate analysis of the geological properties of the drill cuttings can provide real-time data similar to well logging. In the domain of oil and gas exploration and drilling, lithology identification can be achieved using a variety of data samples. These include well log data (such as acoustic logs, resistivity logs, gamma-ray logs, and spectral gamma-ray logs) [1,2], laboratory data (such as core samples, X-ray diffraction (XRD), X-ray fluorescence (XRF), and Near-infrared (NIR) spectroscopy, scanning electron microscopy (SEM) [3, 4]), and data collected during drilling (like white light or UV images, images captured under fluorescent light, or real-time drilling parameters such as weight on bit, torque, rate of penetration, mud properties, etc.), but such data could have some bias, a "depth shift", which makes inaccurate correspondence between images from recorded depth, and depth of drilling data ([5]). Acquiring well log or laboratory data tends to be very costly, whereas data collected during drilling is relatively inexpensive. Drill cuttings contain a lot of information as they cover a wider stratigraphic range compared to cores. Analyzing drill cuttings nearly in real-time is a cost-effective approach to characterizing reservoirs, which includes evaluations of mineralogy, petrophysical properties, and mechanical properties. Drill cuttings are especially advantageous due to the cost-effectiveness of obtaining them and the comprehensive depth of the stratigraphic section they represent. Consequently, automating on-site lithology detection based on data collected during drilling is highly desirable. There are several methods to prepare and analyze drill cutting samples, such as using whole cuttings (unprocessed), creating thin sections (where rock samples are ground down to a specific thickness, usually around 30 microns), or making polished sections (which involves preparing a flat, smooth surface on a rock sample). However, the latter two methods necessitate laboratory equipment, making them impractical for field use. In our study, we explore the feasibility of detecting lithology solely through images taken under white light, which represents the most economical data acquisition method during drilling. In our research, we propose a method to determine the lithological composition of drill cuttings by utilizing digital photographs. This method is based on high-resolution images of cuttings taken from specific depths and historical lithological data from previously drilled and examined wells. We created a deep learning model, more specifically a convolutional model, that can predict the probability of a sample being classified into a particular rock category. This model was trained using wells with known lithological data, which were crucial for testing and verifying the model's effectiveness. Looking ahead, we anticipate that this machine learning model will be able to predict the lithological composition of a sample from cleaned cuttings images with a certain level of probability.