No full-text available
To read the full-text of this research,
you can request a copy directly from the authors.
Machine-learning-based object detection in images for reservoir characterization: A case study of fracture detection in shales
Abstract
Imaging tools are widely used in the petroleum industry to investigate structural features of reservoir rocks directly at multiple scales. Quantitative image analysis is often used to determine various rock properties, but it requires significant time and effort, particularly to analyze a large number of samples. Automated object detection represents a potential solution to this efficiency problem. This method uses computers to efficiently provide quantitative information for thousands of images. Automated fracture detection in scanning electron microscope (SEM) images is presented as an example to show the workflow of using advanced deep-learning tools for quantitative rock characterization. First, an automatic object-detection method is presented for fast identification and characterization of microfractures in shales. Using this approach, we analyzed 100 SEM images obtained from deformed and intact samples of a carbonate-rich shale and a siliceous shale with the goal of analyzing the abundance and characteristics of microfractures generated during hydraulic fracturing. Most of the fractures are detected with about 90% success rate relative to manual picking. Second, we obtained statistics of length and areal porosities of these fractures. The experimentally deformed samples had slightly more detectable microfractures (1.8 fractures/image on average compared to 1.6 fractures/image), and the microfractures induced by shear deformation tend to be short (<50 μm) in the Eagle Ford and long in the siliceous samples, presumably because of differences in rock fabric. In future work, this approach will be applied to characterize the shape and size of mineral grains and to analyze relationships between fractures and minerals.
... Reichstein et al. (2019) showed a perspective to implement deep learning for pattern recognition and statistical downscaling in data-driven earth systems. There are several related works published by Valera et al. (2018), Tian and Daigle (2018), and Srinivasan et al. (2019). For instance, Valera et al. (2018) utilized conventional machine learning using random forest and support vector machine to identify conductive fractures, which could reduce the simulation model size. ...
... Srinivasan et al. (2019) proposed an extended version by including the size and apertures of fractures. Tian and Daigle (2018) implemented deep learning to identify fractures using TensorFlow object detection API, which is based on convolutional neural network. The proposed algorithm, however, cannot detect the fracture objects but rather the zone containing the fractures. ...
... Tian and Daigle (2018) utilized convolutional neural network (MobileNet) to recognize area containing fractures. However, these proposed methods produce significant noise in the fracture patterns and therefore are not accurate enough to be used for DFN construction (Grangier et al. 2009, Farabet et al. 2012, Zhang et al. 2016, Badrinarayanan et al. 2017, Tian and Daigle 2018. ...
In modeling fractured reservoirs, outcrops may offer useful insights about the subsurface characterization of the heterogeneous rock formation. They provide analogs that could be replicated in the reservoir to capture the fracture and matrix characteristics, which are crucial to assess the governing recovery mechanisms. Constructing outcrop-based reservoir models is a labor-intensive process, which is subject to personal interpretation and error. In this work, we propose a novel workflow for modeling fractured reservoirs within a deep learning framework. The workflow consists of three main steps that include fracture network recognition to map explicitly the fractures from digital images, fracture characterization to provide an assessment of the fracture effective hydraulic apertures, and reservoir model construction to integrate the multi-scale data and construct the up-scaled simulation model. In this paper, we focus on the first step in the workflow.
The fracture network recognition starts with segmentation for the images of the fractured formation. The ultimate objective is to identify the fractures from RGB, greyscale, or hyperspectral images. We developed a U-Net-based algorithm to perform the segmentation using 64×64 pixel-resolution. This resolution is carefully selected to accelerate the fracture recognition process and to narrow down the variability in the training set. The inputs are images of the fractured medium with any resolution which are pre-processed before feeding it to the recognition process. The output is a list of the identified fractures, where each fracture is composed of a set of segments, and each segment is defined by the coordinates of its end-points. The output format could be readily processed by any fracture modeling software.
We demonstrate our workflow to recognize and identify fractures from different 2D images, where we discuss the machine-learning (ML) training and testing stages. The algorithm shows accurate predictions and identifications for the fractures. This workflow has the potential to be extended and applied at the field scale.
... In recent years, among all machine learning-based methods for object detection [7][8][9][10], especially deep convolutional neural networks (DCNNs) have drawn immense research attention [11][12][13][14][15][16][17][18]. Due to the high-quality detection accuracy and the powerful feature representation abilities, numerous DCNNsbased methods have been developed for the detection of defect objects [19][20][21]. ...
Steel defect detection is important for industry production as it is tied to the product quality and production efficiency. However, previous steel defect detection methods based on deep convolutional neural networks heavily rely on large‐scale data for training and tend to have poor generalization ability for a novel defect category. In this paper, a novel few‐shot steel defect detection model based on multi‐scale semantic enhancement representation and mask category information mapping is introduced, where only a few annotated samples are acquired for the novel defect category. More concretely, three main components are built: an information‐guidance enhanced multi‐head detector is proposed to improve the representation of information in meta‐feature maps, a mask category representation module is designed to enhance the category feature representation of the mask region in the support set, and a novel multi‐scale category edge loss function is designed to assist the generation of category reweighting vector. Extensive experiments on the North‐east University few‐shot steel defect data set demonstrate that the proposed method significantly outperforms the state‐of‐the‐art methods and verify its effectiveness through ablation studies.
... Additionally, some studies have been undertaken for fracture detection in the micro-CT and SEM images. Tian and Daigle (2018) have trained the CNN-based method of Single Shot Detection (SSD) with MobileNet for segmenting the fractures in the SEM images of the Eagle Ford and siliceous shale samples. The results indicated the success of the developed network even for the slightly damaged fracture. ...
X-ray imaging technology has seen immense progress in extracting the internal structure of geomaterials, but the segmentation of images into voids and solids has remained a challenge due to the stochastic nature of sediments. The traditional segmentation techniques suffer from operator bias; deep learning methods entail time/hardware expenses and a large number of segmented images as ground truth (GT). To address these issues, we introduce PoreSeg, a segmentation framework that leverages the convolutional layers and ensemble learning to automatically segment images via one-shot learning. The framework offers two approaches for implementation: (I) unsupervised-based (UB), in which the imported images are segmented automatically without user intervention, and (II) interactive-based (IB), in which the algorithm uses the manually selected regions of interest (ROI) by the user. Both approaches can be used for binarization and multi-mineral segmentation. To evaluate the performance of the algorithm, a comprehensive databank including the X-ray images of 15 sandstones and 5 carbonates was collected. Our findings indicated the promising performance of UB PoreSeg in the majority of cases with an average Intersection Over Union (IoU) score of 0.97. Moreover, it was construed that the IB approach can cover the UB bottlenecks for complex images and multi-mineral segmentation tasks. The IB PoreSeg outperformed trainable Weka and deep learning based on the visual examination. The PoreSeg is a resource saver, does not need GT masks, and minimizes user bias, making it a suitable tool for the segmentation of X-ray images.
... The segmentation's performance depends on the model type, configuration, and architecture. Many studies have assessed these factors for the segmentation of medical scans [45,[56][57][58][59][60] and X-ray images of subsurface porous deposits [61][62][63][64][65][66]. However, there have been few studies on applying DL models for X-ray image segmentation of GDLs due to lack of data and the complex structure of the wet scans [55]. ...
High-resolution X-ray computed tomography (micro-CT) has been widely used to characterise fluid flow in porous media for different applications, including in gas diffusion layers (GDLs) in fuel cells. In this study, we examine the performance of 2D and 3D U-Net deep learning models for multiphase segmentation of unfiltered X-ray tomograms of GDLs with different percentages of hydrophobic polytetrafluoroethylene (PTFE). The data is obtained by micro-CT imaging of GDLs after brine injection. We train deep learning models on base-case data prepared by the 3D Weka segmentation method and test them on the unfiltered unseen datasets. Our assessments highlight the effectiveness of the 2D and 3D U-Net models with test IoU values of 0.901 and 0.916 and f1-scores of 0.947 and 0.954, respectively. Most importantly, the U-Net models outperform conventional 3D trainable Weka and watershed segmentation based on various visual examinations. Lastly, flow simulation studies reveal segmentation errors associated with trainable Weka and watershed segmentation lead to significant errors in the calculated porous media properties, such as absolute permeability. Our findings show 43, 14, 14, and 3.9% deviations in computed permeabilities for GDLs coated by 5, 20, 40, and 60 w% of PTFE, respectively, compared to images segmented by the 3D Weka segmentation method.
... For instance, Katterbauer et al. [56] incorporated electrical and acoustic image logging data to classify fractures and estimate the fracture degree. With 100 shale scanning electron microscope images, Tian and Daigle [57] applied the automated object detection based on ML to identify and characterize microfractures. Several studies have focused on integrating various data sources to comprehensively characterize the reservoir, typically referred to as integration modelling [58,59]. ...
In the past few decades, the machine learning (or data-driven) approach has been broadly adopted as an alternative to scientific discovery, resulting in many opportunities and challenges. In the oil and gas sector, subsurface reservoirs are heterogeneous porous media involving a large number of complex phenomena, making their characterization and dynamic prediction a real challenge. This study provides a comprehensive overview of recent research that has employed machine learning in three key areas: reservoir characterization, production forecasting, and well test interpretation. The results show that machine learning can automate and accelerate many reservoirs engineering tasks with acceptable level of accuracy, resulting in more efficient and cost-effective decisions. Although machine learning presents promising results at this stage, there are still several crucial challenges that need to be addressed, such as data quality and data scarcity, the lack of physics nature of machine learning algorithms, and joint modelling of multiple data sources/formats. The significance of this research is that it demonstrates the potential of machine learning to revolutionize the oil and gas sector by providing more accurate and efficient solutions for challenging problems.
... Tian and Daigle presented a machine learning framework for the identification of fractures in scanning electron microscope (SEM) images from carbonate-rich shale and siliceous shale samples. The framework utilizes the tensor flow object-detection API to detect objects in images, and utilizes single-shot detector in combination with a mobile net model (Tian and Daigle, 2018). While the results are promising for the detection of fractures in SEM images, the same technique may not easily be applied to formation image logs. ...
Carbon capture and storage (CCS) has attracted strong interest from industry and the scientific community alike due to the ability of storing CO 2 in subsurface reservoirs. Deep saline aquifers may be well suited for the safe and long-term storage given their geological structure. The long term underground storage in saline aquifers depends on variety of interrelated trapping mechanisms in addition to the caprock sealing efficiency. Fractures are commonplace in many geological settings and represent a crucial role for hydrocarbon migrations and entrapment. Fracture impact fluid flow in variety of forms, particularly due to the complexity and varying natures of the fractures, which channel the injected CO 2 throughout the reservoir formation. This is especially important for tight gas reservoirs and low permeable cap rock structures whose permeability is primarily characterized by the fault and fractures. This outlines the importance of determining accurately fracture penetration in wellbores for CO 2 injection. We present a new deep learning framework for the detection of fractures in formation image logs for enhancing CO 2 storage. Fractures may represent high velocity gas flow channels which may make CO 2 storage a challenge. The novel deep learning framework incorporates both acoustic and electrical formation image logs for the detection of fractures in wellbores for CO 2 storage enhancement and injection optimization. The framework was evaluated on the Pohokura-1 well for the detection of fractures, with the framework exhibiting strong classification accuracy. The framework could accurately classify the fractures based on acoustic and electrical image logs in 98.1% for the training and 85.6% for the testing dataset. Furthermore, estimates of the fracture size are strong, indicating the ability of the framework to accurately quantify fracture sizes in order to optimize CO 2 injection and storage.
... al. presented a machine learning framework for the identification of fractures in scanning electron microscope (SEM) images from carbonaterich shale and siliceous shale samples. The framework utilizes the TensorFlow object-detection API to detect objects in images, and utilizes single-shot detector in combination with a mobileNet model (Tian & Daigle, 2018). While the results are promising for the detection of fractures in SEM images, the same technique may not easily be applied to formation image logs. ...
We have presented a new deep learning framework for the detection of fractures in formation image logs for enhancing CO2 storage. Fractures may represent high velocity gas flow channels which may make CO2 storage a challenge. The novel deep learning framework incorporates both acoustic and electrical formation image logs for the detection of fractures in wellbores for CO2 storage enhancement and injection optimization. The framework was evaluated on the Pohokura-1 well for the detection of fractures, with the framework exhibiting strong classification accuracy. The framework could accurately classify the fractures based on acoustic and electrical image logs in 98.1 % for the training and 85.6 % for the testing dataset. Furthermore, estimates of the fracture size are strong, indicating the ability of the framework to accurately quantify fracture sizes in order to optimize CO2 injection and storage.
... Owing to the development of artificial neural networks, neural network models have been applied to solve inverse problems. For example, scholars have used artificial neural networks to determine reservoir parameters to assess the production capacity of wells (Ahmadi et al., 2017;Xiao and Hugh, 2018). Computer vision and neural networks were used to acquire microscopy images, manually construct annotated big data samples, and train those samples to identify, extract, or count cracks, pores, or particles. ...
... Their workflow may suffer from human error and requires extensive labor works for fracture characterization. Tian and Daigle (2018) utilized combined Single Shot Detector (SSD) and MobileNet to recognize fracture in the Scanning Electron Microscope (SEM) images of shales. They use bounding box to locate the fractures and evaluate the length of the fractures by taking the diagonal length of the box. ...
Automatic fracture recognition from borehole images or outcrops is applicable for the construction of fractured reservoir models. Deep learning for fracture recognition is subject to uncertainty due to sparse and imbalanced training set, and random initialization. We present a new workflow to optimize a deep learning model under uncertainty using U-Net. We consider both epistemic and aleatoric uncertainty of the model. We propose a U-Net architecture by inserting dropout layer after every "weighting" layer. We vary the dropout probability to investigate its impact on the uncertainty response. We build the training set and assign uniform distribution for each training parameter, such as the number of epochs, batch size, and learning rate. We then perform uncertainty quantification by running the model multiple times for each realization, where we capture the aleatoric response. In this approach, which is based on Monte Carlo Dropout, the variance map and F1-scores are utilized to evaluate the need to craft additional augmentations or stop the process. This work demonstrates the existence of uncertainty within the deep learning caused by sparse and imbalanced training sets. This issue leads to unstable predictions. The overall responses are accommodated in the form of aleatoric uncertainty. Our workflow utilizes the uncertainty response (variance map) as a measure to craft additional augmentations in the training set. High variance in certain features denotes the need to add new augmented images containing the features, either through affine transformation (rotation, translation, and scaling) or utilizing similar images. The augmentation improves the accuracy of the prediction, reduces the variance prediction, and stabilizes the output. Architecture, number of epochs, batch size, and learning rate are optimized under a fixed-uncertain training set. We perform the optimization by searching the global maximum of accuracy after running multiple realizations. Besides the quality of the training set, the learning rate is the heavy-hitter in the optimization process. The selected learning rate controls the diffusion of information in the model. Under the imbalanced condition, fast learning rates cause the model to miss the main features. The other challenge in fracture recognition on a real outcrop is to optimally pick the parental images to generate the initial training set. We suggest picking images from multiple sides of the outcrop, which shows significant variations of the features. This technique is needed to avoid long iteration within the workflow. We introduce a new approach to address the uncertainties associated with the training process and with the physical problem. The proposed approach is general in concept and can be applied to various deep-learning problems in geoscience.
... However, human detection is a challenging task because everyone has their unique appearance, and the shape of humans can make thousands of gestures simultaneously [3]. In general, object detection and recognition technologies fall either within the machine learning-based approaches (or early methods) [4] or deep learning-based approaches (or modern methods) [5,6]. The early methods require relatively less computing power (No need for graphics processor units (GPUs) to work in real-time) than modern approaches [7]. ...
Human detection is a technology that detects human shapes in the image and ignores everything else. However, modern person detectors have some inefficiencies in detecting pedestrians during video surveillance at night, and the accuracy rate is still insufficient. Therefore, this paper aims to increase the accuracy rate for automatic human detection at night from thermal infrared (TIR) images and real-time video sequences. For this purpose, a new architecture is proposed to enhance the backbone of the Tiny-yolov3 network. The enhanced network used the YOLOv3 algorithm’s tasks with the K-means clustering method to extract more complex features of a person. This network was pre-trained on the MS. COCO dataset to obtain the initial weights. Through the comparison with other related methods showed that the experimental results have achieved the significantly improved performance of human detection from thermal imaging in terms of accuracy, speed, and detection time. The method has achieved a high accuracy rate (90%) compared with the TF-YOLOv3 (88%) trained on the DHU Night Dataset. Although the method has achieved an accuracy rate equal to the YOLOv3-Human (90%), the detection time (4.88 ms) is less, Furthermore, the method has a higher accuracy rate (49.8%) than the YOLO (29.36%) and TF-YOLOv3 (29.8%) with lower detection time (8 ms) on the FLIR Dataset. In addition, the model has achieved a good TP detection for multiple small size of person. By improving the performance of human detection in thermal imaging at night, the method will be able to detect intruders in the night surveillance system.
... 15 However, machine learning algorithms require a large amount of training data/labelled data and to generate these data sets it is very time consuming, even with a state-of-the-art manual labelling tool. 16,17 Results in Figure 2C and D compare the superpixels extracted by DT watershed as well as by the popular SLIC algorithm. 11 Both algorithms can capture most of the details in the background image. ...
Burning of clinker is the most influencing step of cement quality during the production process. Appropriate characterization for quality control and decision making is therefore the critical point to maintain a stable production but also for the development of alternative cements. Scanning electron microscopy (SEM) in combination with energy dispersive X-ray spectroscopy (EDX) delivers spatially resolved phase and chemical information for cement clinker. This data can be used to quantify phase fractions and chemical composition of identified phases.
The contribution aims to provide an overview of phase fraction quantification by semi-automatic phase segmentation using high-resolution backscattered electron (BSE) images and lower-resolved EDX element maps. Therefore, a tool for image analysis was developed that uses state-of-the-art algorithms for pixel-wise image segmentation and labeling in combination with a decision tree that allows searching for specific clinker phases. Results show that this tool can be applied to segment sub-micron scale clinker phases and to get a quantification of all phase fractions. In addition, statistical evaluation of the data is implemented within the tool to reveal whether the imaged area is representative for all clinker phases.
This article is protected by copyright. All rights reserved
... Wu et al. (2019b) provided a comparative study between different segmentation techniques and a random forest distribution bases ML algorithm to identify pores, organic matter matrix, and pyrite from SEM images. Tian and Daigle (2018) used TensorFlow object detection API for identifying fractures in Eagle Ford shale SEM images. On the other hand, Li et al. (2019a) made a finite element-based model using SEM images of shale, which were further analyzed using a convolutional neural network (CNN) to predict the mechanical property. ...
Imaging and image analysis of shale provides insights into the pore and fracture networks, disposition of compositional elements, and signatures of diagenesis and fluid flow. Extraction of features and reconstruction to get highly accurate models of shales using image analysis involves various techniques and tools. The present paper reviews different 3D and 2D imaging methods, their advantages, and limitations, along with recommendations on making the most appropriate choice for visualization of desired features and their applications. While fractures and other elements of deformation can be well resolved using X-ray CT, FIB-SEM is a better suited 3D imaging tool for resolving the nano-scale pore attributes of shale. Among 2D imaging methods, TEM can provide direct information on the micropores (<2 nm), which cannot be obtained using SEM imaging. The paper discusses guidelines on filtering, thresholding, and optimization, with a special emphasis on image segmentation and feature extraction that may be selected based on the problem definition. It was found that adopting multiple imaging methods can provide a plethora of complementary attributes; however, the scale of observation can vary up to three orders of magnitude, which may lead to inaccurate upscaling. Experimental methods such as low-pressure gas adsorption, NMR, and small-angle scattering can not only validate the results of imaging but also extend the range of identification of pore characteristics to deliver a highly comprehensive understanding of shales. Choice of imaging tools and their combinations, appropriate sample preparation protocols, application of machine learning and advanced numerical simulation, and validation with native sub-surface will remain key to the applicability of imaging in unlocking the full potential of shale as a reservoir.
... 최근 머신 러닝 알고리즘과 계산 자원의 발전 덕분에 많 은 분야에서 머신 러닝이 활발히 연구되고 있으며 탄성파 분야에도 머신 러닝을 적용하는 연구가 증가하는 추세이다 (Choi et al., 2020). 탄성파 탐사 분야에서는 현재까지 탄성 파 상 분석 (Wrona et al., 2018;Keynejad et al., 2019), 저류 층 특성화 (Chaki, 2015;Tian and Daigle, 2018;Smith et al., 2019), 단층해석 (Zhang et al., 2014;Xiong et al., 2018;Kumar and Sain, 2018), 암염구조 해석 (Guillen et al., 2015;Shi, et al., 2019;Sen et al., 2019), 탄성파 속성분석 (Zhao, 2018b;Griffith et al., (Zhao, 2018a). 합성곱 신경망은 특 징 추출과 토폴로지(topology)가 연산에 반영될 수 있도록 하는 합성곱층(convolutional layer)과 특징을 강화시키고 이미지의 크기를 조절하는 풀링층(pooling layer) 등이 사 용된다 (Yi, 2019). ...
... Trained computer models efficiently perform tasks that, by convention, are done by human experts with years of experience in specialized fields. Deep learning techniques have been successfully applied to many fields, such as computer vision, voice recognition, medical image analysis and many other domains including geosciences and geophysics, with results comparable to and in some cases superior to human experts alone (e.g., Aprile et al., 2014;Ciresan et al., 2012;Falk and Mai, 2019;Karpatne et al., 2018;LeCun et al., 2015;Noh et al., 2015;Ronneberger et al., 2015;Tian and Daigle, 2018;Zhang et al., 2015). Karimpouli and Tahmasebi (2019) proposed the use of a "deep convolutional autencoder" for their digital rock image segmentations. ...
The structural properties of the dust layer, including its thickness, porosity, and particle size distribution, play a critical role in ensuring the high precision and long-term stability of filter elements. However, observing these properties is challenging due to the weak adherence and cohesiveness of the layer. To address this issue, atomization thermosetting glue was used to achieve pre-curing, and the entire dust layer was cured with epoxy resin. After the sample was frozen and fractured using liquid nitrogen, the boundaries of the dust particles became plainly visible. Traditional binarization techniques were insufficient in identifying the edges of the dust particles since the grayscale values of particles and their environment partially overlap. As a result, a deep learning model based on the DeeplabV3+ network architecture was used to identify particles in the dust layer and achieved an accuracy of 90.99%. The research reveals that pulse-jet cleaning can double the thickness of the local dust layer on adjacent filter elements. Additionally, the surface morphology of the filter element significantly impacts the shape and thickness of the dust layer, causing it to change dramatically. Uneven thickness of the dust layer can result in a higher number of dust particles passing through the filter element membrane.
Understanding the temporospatial distribution of microseismic events is critical for the early warning of induced geohazards during deep rock excavation. However, the accuracy of microseismic source location can be influenced by many geological and engineering factors, such as rock discontinuities and excavated openings. This study addresses the application of fast marching method (FMM) to improve the location accuracy in faulted rock masses and takes a case of powerhouse excavation in the Shuangjiangkou hydropower project as an example. The FMM-based method includes five steps, such as inputs, gridding, velocity assignment, calculation, and output, and can be readily processed using a travel time database. Our results exhibit that the wave paths obtained from the FMM-based method are consistent with those from the analytical solution based on the Snell law and better than those from the commonly used uniform velocity model. The results also signify that the FMM-based method can improve the location accuracy for P-wave propagating across a low-velocity layer and an excavated opening. Moreover, this method can be further improved, such as reasonable determination of grid spacing and thorough detection of major discontinuities.
This chapter gives a brief overview of reservoir characterization, with a focus on how the emerging artificial intelligence, machine learning and data analytics techniques can improve the ability to do reservoir characterization (RC). Specifically, the ability to use AI to integrate different data types with differentScale, Uncertainty, Resolution, and Environment, that we refer to it as the SURE challenge is discussed. We show how AI offers a natural toolbox for reservoir property estimation, and their uncertainties. Machine/Deep Learning-based methods perform much like a human brain. They can receive variety of data from many different sources with drastically different characteristics, and undertake necessary evaluations and perception-based measures, and eventually make the right decisions and/or solve complicated problems. Human intelligence (engineers, geoscentists) will always have a superior performance with qualitative data than computers that are better in dealing with quantitative data. Several examples on how effective human-machine interfaces to create hybrid solutions to address different reservoir characterization problems are provided.
Segmentation and analysis of individual pores and grains of mudrocks from scanning electron microscope images is non-trivial because of imaging artifacts, variation in pixel grayscale values across images, and overlaps in grayscale values among different physical features such as silt grains, clay grains and pores, which make identifications difficult. Moreover, because grains and pores often have overlapping grayscale values, direct application of threshold-based segmentation techniques is not sufficient. Recent advances in the field of computer vision have made it easier and faster to segment images and identify multiple occurrences of such features in an image, provided that ground-truth data for training the algorithm are available. Here we propose a deep learning SEM image segmentation model, MudrockNet based on Google's DeepLab-v3+ architecture implemented with the TensorFlow library. The ground-truth data were obtained from an image-processing workflow applied to scanning electron microscope images of uncemented muds from the Kumano Basin offshore Japan at depths <1.1 km. The trained deep learning model obtained a pixel-accuracy > 90%, and predictions for the test data obtained a mean intersection over union (IoU) of 0.6663 for silt grains, 0.7797 for clay grains and 0.6751 for pores. We also compared our model with the random forest classifier using trainable Weka segmentation in ImageJ, and it was observed that MudrockNet gave better predictions for silt grains, clay grains and pores in most cases. The size, concentration, and spatial arrangement of the silt and clay grains can affect the petrophysical properties of a mudrock, and an automated method to accurately identify the different grains and pores in mudrocks can help improve reservoir and seal characterization for petroleum exploration and anthropogenic waste sequestration.
This chapter will attempt to provide an overview over some of the practical applications that machine learning has found in oil and gas. The aim of the chapter is twofold: First, it is to show that there are many applications that are realistic and have been carried out on real-world assets, that is, machine learning is not a dream. Second, the status of machine learning in oil and gas is in its early days as the applications are specialized and localized. It must be stated clearly that most of the studies done, have been done at universities and that the applications fully deployed in commercial companies are the exception. This chapter makes no attempt at being complete or even representative of the work done. It just provides many starting points for research on use cases and presents an overview. There are some use cases that attract a vast number of papers and this chapter will present such use cases with just one or a few exemplary papers chosen at random.
Segmentation and analysis of individual pores and grains of mudrocks from scanning electron microscope images is non-trivial because of noise, imaging artifacts, variation in pixel grayscale values across images, and overlaps in grayscale values among different physical features such as silt grains, clay grains and pores in an image, which make their identification difficult. Moreover, because grains and pores often have overlapping grayscale values, direct application of threshold-based segmentation techniques is not sufficient. Recent advances in the field of computer vision have made it easier and faster to segment images and identify multiple occurrences of such features in an image, provided that ground-truth data for training the algorithm is available. Here we propose a deep learning SEM image segmentation model, MudrockNet based on Google’s DeepLab-v3+ architecture implemented with the TensorFlow library. The ground-truth data was obtained from an image-processing workflow applied to scanning electron microscope images of uncemented muds from the Kumano Basin offshore Japan at depths < 1.1 km. The trained deep learning model obtained a pixel-accuracy about 90%, and predictions for the test data obtained a mean intersection over union (IoU) of 0.6591 for silt grains and 0.6642 for pores. We also compared our model with the random forest classifier using trainable Weka segmentation in ImageJ, and it was observed that MudrockNet gave better predictions for both silt grains and pores. The size, concentration, and spatial arrangement of the silt and clay grains can affect the petrophysical properties of a mudrock, and an automated method to accurately identify the different grains and pores in mudrocks can help improve reservoir and seal characterization for petroleum exploration and anthropogenic waste sequestration.
Inspired by image-to-image translation, we applied deep learning (DL) to regularly missing data reconstruction, aimed at translating incomplete data into their corresponding complete data. With this purpose in mind, we first construct a network architecture based on an end-to-end U-Net convolutional network, which is a generic DL solution for various tasks. We then meticulously prepare the training data with both synthetic and field seismic data. This article is implemented in Python based on Keras (a high-level DL library). We described the network architecture, the training data, and the training settings in detail. For training the network, we employed a mean-squared-error loss function and an Adam optimization algorithm. Next, we tested the trained network with several typical data sets, achieving good performances (even in the presence of big gaps) and validating the feasibility, effectiveness, and generalization capability of the assessed framework. The feature maps for a sample going through the well-trained network are uncovered. Compared with the f-x prediction interpolation method, DL performs better and is capable of avoiding several assumptions (e.g., linearity, sparsity, etc.) associated with conventional interpolation methods. We demonstrated the influences of the network depth, the kernel size of the convolution window, and the pooling function on the DL results. We applied the trained network to dense data reconstruction successfully. The proposed method can overcome noise to some extent. We finally discussed some practical aspects and extensions of the evaluated framework.
Fine‐grained sediment (mud) and lithified equivalents (mudrock, mudstone, and shale) contain components similar to ones in coarser sedimentary materials, albeit of such small size that high‐resolution imaging is required to observe them. Such imaging reveals that fine‐grained sedimentary rocks display diversities of grains, pores, and diagenetic features that actually exceed the variations of components in common sandstones and limestones. Mudrock diversity reflects the extraordinary range of grain and pore sizes, which extend from detrital grains and authigenic crystals in the <1–100 μm fraction to the nanomaterials (crystals and pores) in the matrix surrounding the silt‐size fraction. Prediction of bulk‐property evolution in fine‐grained materials lags current process understanding in coarse materials but a view is emerging that while there are similarities, there are also contrasts between the responses of coarse and fine materials to changing conditions in the subsurface. Mechanical and chemical processes that operate on submicrometer pores and crystals very likely proceed to different limits, at different rates, and even by entirely different mechanisms than do comparable processes in coarser materials. This paper reviews current knowledge about mudrock components, and explores some of the gaps that exist in our understanding of microscale properties and processes in Earth's most abundant sedimentary material.
Characterizing the microstructure of shales is challenging due to its extremely small scale, and generally involves manual interpretation of image data. Here, we present a completely automatic machine learning method for quantifying the preferential mineral-microfracture relationships in intact and deformed shales. This new method is innovative because it allows automated analysis of large volumes of image data, which is typically very time- and labor-intensive with existing techniques. Using automatic object detection algorithms, 225 images, including energy-dispersive X-ray spectroscopy images and backscattered electron images, were analyzed. These images were obtained from deformed and intact samples of a carbonate-rich organic shale and a siliceous organic shale. We quantitatively characterized the location and size of microfractures and their preferential association with particular minerals. The results show that compression created microfractures at the grain scale. More than 90% of microfractures developed within organic matter (OM), and that the microfractures tended to develop along grain/mineral boundaries. In addition, we found that the fabric of the rock plays an important role in microfracture generation, with laminated OM and clay tending to favor microfracture development while more massive minerals inhibited it. This quantitative analysis helps to improve understanding of the micromechanics of deformation during hydraulic fracturing. In addition, this approach is completely automatic, which could increase work efficiency and reduce the effects of subjective decision-making. The work presented here will greatly improve future studies of quantifying fracture and mineral properties, and can provide guidance for hydraulic fracturing and production strategies.
We performed a series of laboratory and image analysis on organic shale samples before and confined compressive strength tests. Following failure, we often observe an increase in pore volume in the sub-micron range, which appears to be related to the formation of microcracks that in some cases cross or terminate in organic matter, intersecting the organic-hosted pores. Samples with higher clay content tended not to display this behavior. The microcrack networks allow the hydrocarbons to migrate to the main induced tensile fractures. The disconnected nature of the microcracks causes only a slight increase in permeability, consistent with other observations.
We present a method for detecting objects in images using a single deep neural network. Our approach, named SSD, discretizes the output space of bounding boxes into a set of default boxes over different aspect ratios and scales per feature map location. At prediction time, the network generates scores for the presence of each object category in each default box and produces adjustments to the box to better match the object shape. Additionally, the network combines predictions from multiple feature maps with different resolutions to naturally handle objects of various sizes. SSD is simple relative to methods that require object proposals because it completely eliminates proposal generation and subsequent pixel or feature resampling stages and encapsulates all computation in a single network. This makes SSD easy to train and straightforward to integrate into systems that require a detection component. Experimental results on the PASCAL VOC, COCO, and ILSVRC datasets confirm that SSD has competitive accuracy to methods that utilize an additional object proposal step and is much faster, while providing a unified framework for both training and inference. For \(300 \times 300\) input, SSD achieves 74.3 % mAP on VOC2007 test at 59 FPS on a Nvidia Titan X and for \(512 \times 512\) input, SSD achieves 76.9 % mAP, outperforming a comparable state of the art Faster R-CNN model. Compared to other single stage methods, SSD has much better accuracy even with a smaller input image size. Code is available at https:// github. com/ weiliu89/ caffe/ tree/ ssd.
The term induced un-propped (IU) fractures refers to fractures created around the main propped fracture which are too small to accommodate any proppant. These could include natural fractures, and micro-fractures induced along bedding planes or along other planes of weakness. Based on production data, diagnostic methods and field observations it is becoming increasingly clear that induced un-propped fractures created during the hydraulic operation play a critical role in determining the success of fracture treatments.
In this paper five independent pieces of evidence are presented to prove the existence of induced un-propped fractures in most wells under downhole conditions and to demonstrate their importance in production. These include: micro-seismic data, production history matching, tracer data, pressure communication between wells and finally calculations on the fate of the injected fracturing fluids. Examples are provided to clearly demonstrate how this information indicates the presence of IU fractures.
Once the existence of these fractures has been demonstrated simulations are conducted to show what role these induced un-propped fractures play during fracturing and during short-term and long-term production. It is shown that shale properties and the rheology and rate of injection of fracturing fluids plays an important role in the spatial extent and width of IU fractures. In some shales these fractures play a dominant role while in others they may be less important.
Finally recommendations are made for fracture design to account for the presence of these IU fractures. These recommendations have a large impact on important decisions such as well spacing, fracture spacing, fluid rheology, proppant loading, proppant size and other fracture design considerations. In many instances these recommendations would have been quite different had the presence of IU fractures not been recognized.
Pore types, pore size, and pore abundance vary systematically across thermal maturity in the Eagle Ford Formation, Maverick Basin, southern Texas. Scanning electron imaging of 20 samples from four wells is used to assess the complex response of pores to chemical and mechanical processes, entailing both destruction of primary porosity and generation of secondary pores. Primary mineral-associated pores are destroyed by compaction, cementation, and infill of secondary organic matter, whereas secondary pores are generated within organic matter (OM). Destruction of primary pores during early burial (to Ro ∼0.5%) occurs by compaction of ductile detrital OM and clays and, to a lesser degree, as a result of cementation and infill of secondary OM. Larger pores are associated with coccolith debris. The dominant OM is spatially isolated detrital OM "stringers." Porosity is volumetrically dominated (average 6.2%) by relatively large, mostly interparticle mineral-associated pores (median size 51.6 nm [0.000002 in.]; detection limit near 3-4 nm [0.00000012-0.00000015 in.]). At low maturity, porosity and pore size correlate directly with calcite abundance and inversely with OM volumes. At higher maturity, further destruction of primary pores occurs through cementation, secondary OM infill, and greater compaction. Mineral-associated pores are present at high-maturity (Ro ∼1.2% - 1.3%), but are smaller (median size 30.2 nm [0.0000011 in.]) and less abundant (average of 2.5%) than at low maturity. A large portion of OM within high-maturity samples is diagenetic in origin and has pervaded into primary pore space, coating cement crystals, and filling intraparticle pores. Substantial mineral-associated porosity is locally present in samples where incursion of primary pore space by secondary OM has not occurred. Abundant secondary porosity is generated as OM matures into the wet-gas window. Porosity in most high-maturity samples is volumetrically dominated (average of 1.3%) by smaller, OM-hosted pores (median size 13.2 nm [0.00000051 in.]). Copyright © 2015. The American Association of Petroleum Geologists. All rights reserved.
Matrix-related pore networks in mudrocks are composed of nanometer- to micrometer-size pores. In shale-gas systems, these pores, along with natural fractures, form the flow-path (permeability) network that allows flow of gas from the mudrock to induced fractures during production. A pore classification consisting of three major matrix-related pore types is presented that can be used to quantify matrix-related pores and relate them to pore networks. Two pore types are associated with the mineral matrix; the third pore type is associated with organic matter (OM). Fracture pores are not controlled by individual matrix particles and are not part of this classification. Pores associated with mineral particles can be subdivided into interparticle (interP) pores that are found between particles and crystals and intraparticle (intraP) pores that are located within particles. Organic-matter pores are intraP pores located within OM. Interparticle mineral pores have a higher probability of being part of an effective pore network than intraP mineral pores because they are more likely to be interconnected. Although they are intraP, OM pores are also likely to be part of an interconnected network because of the interconnectivity of OM particles. In unlithifed near-surface muds, pores consist of interP and intraP pores, and as the muds are buried, they compact and lithify. During the compaction process, a large number of interP and intraP pores are destroyed, especially in ductile grain-rich muds. Compaction can decrease the pore volume up to 88% by several kilometers of burial. At the onset of hydrocarbon thermal maturation, OM pores are created in kerogen. At depth, dissolution of chemically unstable particles can create additional moldic intraP pores.
We present an integrated framework for using Convolutional Networks for
classification, localization and detection. We show how a multiscale and
sliding window approach can be efficiently implemented within a ConvNet. We
also introduce a novel deep learning approach to localization by learning to
predict object boundaries. Bounding boxes are then accumulated rather than
suppressed in order to increase detection confidence. We show that different
tasks can be learnt simultaneously using a single shared network. This
integrated framework is the winner of the localization task of the ImageNet
Large Scale Visual Recognition Challenge 2013 (ILSVRC2013), and produced near
state of the art results for the detection and classifications tasks. Finally,
we release a feature extractor from our best model called OverFeat.
The matrix permeability of shale is controlled by the microstructure of shale pore network. Therefore, a thorough understanding of shale pore structure is fundamental to the prediction of shale permeability. We constructed a physically representative pore network model for two Barnett Shale samples. We predicted a Darcy permeability of 7.19 nanodarcies (nd) and apparent permeability of 57.34 nd for these samples. We explored the pore structure of shale matrix with low-pressure nitrogen adsorption/desorption isotherms. The pore size distribution, the network connectivity, as well as pore spatial arrangement are determined for the organic matter and the whole matrix. We separated pores in the network model into two groups: the affiliated pores and the dominant pores. For these samples, the affiliated pores are pores with diameters smaller than 8 nm and the dominant pores are with diameters larger than 8 nm. The affiliated pores develop on the walls of the dominant pores. The cutoff value between affiliated pores and dominant pores may differ for different samples. The pore spatial arrangement is validated against previous literature resultsand a diagenetic explanation for development of shale pore system. The network model can be used to predict shale permeability and other petrophysical properties.
Can a large convolutional neural network trained for whole-image classification on ImageNet be coaxed into detecting objects in PASCAL? We show that the answer is yes, and that the resulting system is simple, scalable, and boosts mean average precision, relative to the venerable deformable part model, by more than 40% (achieving a final mAP of 48% on VOC 2007). Our framework combines powerful computer vision techniques for generating bottom-up region proposals with recent advances in learning high-capacity convolutional neural networks. We call the resulting system R-CNN: Regions with CNN features. The same framework is also competitive with state-of-the-art semantic segmentation methods, demonstrating its flexibility. Beyond these results, we execute a battery of experiments that provide insight into what the network learns to represent, revealing a rich hierarchy of discriminative and often semantically meaningful features.
Get started with MATLAB for deep learning and AI with this in-depth primer. In this book, you start with machine learning fundamentals, then move on to neural networks, deep learning, and then convolutional neural networks. In a blend of fundamentals and applications, MATLAB Deep Learning employs MATLAB as the underlying programming language and tool for the examples and case studies in this book.
With this book, you'll be able to tackle some of today's real world big data, smart bots, and other complex data problems. You’ll see how deep learning is a complex and more intelligent aspect of machine learning for modern smart data analysis and usage.
You will:
• Use MATLAB for deep learning
• Discover neural networks and multi-layer neural networks
• Work with convolution and pooling layers
• Build a MNIST example with these layers
Shales are heterogeneous media with porosity at many scales and in many microtextural positions, including within organic matter and clay aggregates. Because these materials have contrasting mechanical properties, it remains unclear how induced fractures manage to connect with this porosity whether during hydrocarbon production, wastewater injection, or carbon-capture-and-storage efforts. To explore porosity changes related to fracturing, we experimentally failed shale samples in a triaxial load apparatus and observed changes in microstructure through scanning electron microscopy, low-pressure nitrogen sorption, and nuclear magnetic resonance. We observed a system of microcracks, many of which were likely experimentally induced and localized on grain boundaries. In some cases these fractures propagated into regions of natural porosity in organic matter. In this subsurface this “pore capture” likely enhances pore connectivity, but only selectively depending upon mechanical conditions. Pore capture is one possible mechanism by which multi-scale compositional heterogeneity in shales may affect rheological heterogeneity.
Scanning electron microscopy (SEM) and its auxiliary technologies have been exploited by various industries for over 50 years. One of its main applications has been to characterize materials at very high magnification. Recent interest by the petroleum industry to better characterize and understand shale hydrocarbon reservoirs has led to an increased interest in utilizing SEM technology for shale reservoir studies. The purpose of this chapter is to review basic SEM operational principles and applications that are particularly useful for characterizing shale hydrocarbon reservoirs.
This paper proposes a Fast Region-based Convolutional Network method (Fast
R-CNN) for object detection. Fast R-CNN builds on previous work to efficiently
classify object proposals using deep convolutional networks. Compared to
previous work, Fast R-CNN employs several innovations to improve training and
testing speed while also increasing detection accuracy. Fast R-CNN trains the
very deep VGG16 network 9x faster than R-CNN, is 213x faster at test-time, and
achieves a higher mAP on PASCAL VOC 2012. Compared to SPPnet, Fast R-CNN trains
VGG16 3x faster, tests 10x faster, and is more accurate. Fast R-CNN is
implemented in Python and C++ (using Caffe) and is available under the
open-source MIT License at https://github.com/rbgirshick/fast-rcnn.
The Marcellus Formation of Pennsylvania represents an outstanding example of an organic matter (OM)-hosted pore system; most pores detectable by field-emission scanning electron microscopy (FE-SEM) are associated with OM instead of mineral matrix. In the two wells studied here, total organic carbon (TOC) content is a stronger control on OM-hosted porosity than is thermal maturity. The two study wells span a maturity from late wet gas (vitrinite reflectance [RoJ, -1.0%) to dry gas (Ro, -2.1%). Samples with a TOC less than 5.5 wt. % display a positive correlation between TOC and porosity, but samples with a TOC greater than 5.5 wt. % display little or no increase in porosity with a further increasing TOC. In a subset of samples (14) across a range of TOC (2.3-13.6 wt. %), the pore volume detectable by FE-SEM is a small fraction of total porosity, ranging from 2 to 32% of the helium porosity. Importantly, the FE-SEM-visible porosity in OM decreases significantly with increasing TOC, diminishing from 30% of OM volume to less than 1% of OM volume across the range of TOC. The morphology and size of OM-hosted pores also vary systematically with TOC. The interpretation of this anticorrelation between OM content and SEM-visible pores remains uncertain. Samples with the lowest OM porosity (higher TOC) may represent gas expulsion (pore collapse) that was more complete as a consequence of greater OM connectivity and framework compaction, whereas samples with higher OM porosity [lower TOC) correspond to rigid mineral frameworks that inhibited compactional expulsion of methane-filled bubbles. Alternatively, higher TOC samples may contain OM (low initial hydrogen index, relatively unreactive) that is less prone to development of FE-SEM-detectable pores. In this interpretation, OM type, controlled by sequence-stratigraphic position, is a factor in determining pore-size distribution. Copyright © 2013. The American Association of Petroleum Geologists. All rights reserved.
Can a large convolutional neural network trained for whole-image
classification on ImageNet be coaxed into detecting objects in PASCAL? We show
that the answer is yes, and that the resulting system is simple, scalable, and
boosts mean average precision, relative to the venerable deformable part model,
by more than 40% (achieving a final mAP of 48% on VOC 2007). Our framework
combines powerful computer vision techniques for generating bottom-up region
proposals with recent advances in learning high-capacity convolutional neural
networks. We call the resulting system R-CNN: Regions with CNN features. The
same framework is also competitive with state-of-the-art semantic segmentation
methods, demonstrating its flexibility. Beyond these results, we execute a
battery of experiments that provide insight into what the network learns to
represent, revealing a rich hierarchy of discriminative and often semantically
meaningful features.
To understand the deliverability of gas shales, one needs to understand the controls on porosity and permeability. The microstructure of shale is defined in part by the grain size which is typically less than 5 μm and composition. Gas shales unlike other lithologies contain significant quantities of organic matter in various stages of maturation. The geometry and nature of the mineralogical components and the organics would be easy to describe if the objects could be identified optically. Most investigations into shale microstructure have relied on technologies such as Scanning Electron Microscopy (SEM), X-ray imaging, Transmission Electron Microscopy (TEM) or Scanning Acoustic Microscopy (SAM). Each has advantages an limitations. Our focus here is limited to SEM studies we performed on gas shales. Our journey began with generic imaging of broken surfaces and has progress to imaging of ion-milled surfaces in dual beam SEM. The latest technology has provide three dimension images with maximum resolution of 4-5nm pores. Porosity is found on the microscale in organics, between grains, in pyrite framboids, fossils, within minerals and in the form of microcracks. The majority of pores in some shales an located in the organics. Other shales show the porosity to be largely associated with minerals. SEM resolved pore dimension agree well with "as received" NMR measurement. However, high pressure mercury injection measurements suggest that the paths connecting these pores are even smaller. Three dimensional reconstruction of sequentially ion-milled surfaces using a dual beam SEM provides controls on the volumetric distribution of pores and organics and their connectivity. Initial and limited analyses indicate that the shales investigated are dominated by smaller pores. Keep in mind that any such study only samples an extremely small portion of any reservoir and that while generalizations are tempting, statistical studies are required to establish the universality of such observations.
The paper presents the results of laboratory direct shear tests on the hydro-mechanical behaviour of an extensional fracture in shale. The tests were conducted on fracture specimens obtained by mechanically splitting blocks of a naturally fractured shale along the cemented fracture and dissolving the calcite fracture filling with a strong acid. The experiments consisted of radial fluid flow tests from a single fluid source along the fracture surface. Fracture samples were tested under different levels of effective normal stress and shear displacement. The experimental results show significant reduction of fracture permeability during increasing contact normal stress across the fracture and after shearing of the fracture under constant high normal stress. However, the tested fractures never completely closed even under normal stresses close to or higher than the unconfined compressive strength of the intact shale. The fracture permeability remained much higher than the intact shale (or matrix) permeability even when the fracture surface has undergone some gouge formation from local asperity failure during shearing. The results indicate that fractures once created are difficult to close by mechanical loading. In tight formations like shales, fractures will always be conduits for fluid flow unless closed by cementation. Currently, it is widely assumed that fractures in sedimentary basins stay open only when there is an overpressure which could keep the fractures hydraulically parted. However, the results presented here indicate that given a sufficient hydraulic gradient, fluid can still flow along a fracture even in the absence of large overpressures.
During the past decade there has been an explosion in computation and information technology. With it have come vast amounts of data in a variety of fields such as medicine, biology, finance, and marketing. The challenge of understanding these data has led to the development of new tools in the field of statistics, and spawned new areas such as data mining, machine learning, and bioinformatics. Many of these tools have common underpinnings but are often expressed with different terminology. This book describes the important ideas in these areas in a common conceptual framework. While the approach is statistical, the emphasis is on concepts rather than mathematics. Many examples are given, with a liberal use of color graphics. It is a valuable resource for statisticians and anyone interested in data mining in science or industry. The book's coverage is broad, from supervised learning (prediction) to unsupervised learning. The many topics include neural networks, support vector machines, classification trees and boosting---the first comprehensive treatment of this topic in any book.
This major new edition features many topics not covered in the original, including graphical models, random forests, ensemble methods, least angle regression and path algorithms for the lasso, non-negative matrix factorization, and spectral clustering. There is also a chapter on methods for ``wide'' data (p bigger than n), including multiple testing and false discovery rates.
Trevor Hastie, Robert Tibshirani, and Jerome Friedman are professors of statistics at Stanford University. They are prominent researchers in this area: Hastie and Tibshirani developed generalized additive models and wrote a popular book of that title. Hastie co-developed much of the statistical modeling software and environment in R/S-PLUS and invented principal curves and surfaces. Tibshirani proposed the lasso and is co-author of the very successful An Introduction to the Bootstrap. Friedman is the co-inventor of many data-mining tools including CART, MARS, projection pursuit and gradient boosting.
TensorFlow: A system for large-scale machine learning
- M Abadi
- P Barham
- J Chen
- Z Chen
- A Davis
- J Dean
- M Devin
Abadi, M., P. Barham, J. Chen, Z. Chen, A. Davis, J. Dean, M.
Devin, et al., 2016, TensorFlow: A system for large-scale machine
learning: Proceedings of 12 th USENIX Conference on Operating
Systems Design and Implementation, 265-283.
Gas flow tightly coupled to elastoplastic geomechanics for tight-and shale-gas reservoirs: Material failure and enhanced permeability
- J Kim
- G J Moridis
Kim, J., and G. J. Moridis, 2014, Gas flow tightly coupled to elastoplastic geomechanics for tight-and shale-gas reservoirs: Material failure and enhanced permeability: SPE Journal, 19, no. 6,
https://doi.org/10.2118/155640-PA.