Exploring Convolutional Neural Networks and Transfer Learning for Oil Sands Drill Core Image Analysis
Date
2021-08-24
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
An accurate permeability estimate is crucial for effectively characterizing the McMurray oil sands for in situ recovery. Such an estimate is critical to inform the best locations for placing wells and pads and accurately forecast future oil production rates. This fact is becoming significantly important as in situ development moves to areas of increasingly complex geology. The traditional methods of estimating permeability largely do not work well in oil sands because of the core disturbance or the fact that the core is filled with immobile bitumen. Moreover, it is expensive to get physical samples from many different depths at many wells, and the experiments carried out in the labs to measure permeability sometimes are not representative. However, permeability can be estimated from different parameters such as mean grain size (MGS), median grain size, and particle size distribution (PSD). This thesis investigates how convolutional neural networks (CNNs) and transfer learning perform when estimating MGS from the oil sands drill core photos. Three preliminary approaches are explored for classifying core photos based on the facies, including (1) the application of transfer learning on the pre-trained VGG-16 CNN model, (2) fine-tuning a few top layers of VGG-16, and (3) the combination of VGG-16 and traditional machine learning (ML) algorithms. Experimental results achieved by these classification models reveal opportunities to extend these approaches for predicting MGS from core photos. Therefore, the three approaches are then investigated using a library of core photographs with known MGS calculated from PSD to see which one works best. Experimental results exhibit good performance in estimating MGS from core photos using the explored approaches. Overall, the investigation supports that the application of CNNs, and transfer learning is feasible in different oil sands drill core image analysis workflows and more advanced research outcomes can be achieved by further exploration of these techniques in the oil sands research domain.
Description
Keywords
Convolutional Neural Networks (CNNs), Transfer learning, VGG-16, Traditional Machine Learning, Random Forest, Oil Sands, Drill Core Image Analysis, Particle Size Distribution, Permeability, Mean Grain Size, Facies, McMurray oil sands
Citation
Anzum, F. (2021). Exploring Convolutional Neural Networks and Transfer Learning for Oil Sands Drill Core Image Analysis (Master's thesis, University of Calgary, Calgary, Canada). Retrieved from https://prism.ucalgary.ca.