Digitizer: A Synthetic Dataset for Well-Log Analysis

Computers & Geosciences

Brittleness is an important geomechanical property of reservoirs, which is usually estimated from cores or sonic logs that are expensive to acquire. In this study, we report data-driven, machine learning workflows to predict brittleness from less expensive, readily available conventional logs. We propose three strategies to predict brittleness using gamma ray, neutron porosity, density, caliper, sonic, and photoelectric factor logs by utilizing gradient boosting (GB), support vector regression (SVR), and neural networks (NN). The first strategy involves predicting brittleness directly from the logs while the second strategy predicts shear sonic logs used for the estimation of brittleness. The performance of the models given as R 2 on deployment on the testing set for the first strategy is: GB (0.87), SVR (0.73), and NN (0.82), while for the second strategy: GB (0.94), SVR (0.87), and NN (0.94). In the third strategy, we convert the prediction into a classification task by grouping the brittleness estimate into ductile, transition, and brittle. The accuracy of model deployment on the testing set for the third strategy is: GB (89.37%), SVM (89.06%), and NN (89.16%). We demonstrate that depending on the strategy adopted, the gradient boosting algorithm outperforms the other peer algorithms in terms of training and validation scores. Furthermore, we combine the three algorithms using a committee machine to improve the performance of the model. The workflow in this study can be adopted to predict other reservoir properties from available logs. The workflow can also be adopted to characterize reservoir heterogeneity from seismic traces trained by well logs.

Journal of Petroleum Science and Engineering

One of the first steps during the investigation of geological objects is the interwell correlation. It provides information on the structure of the objects under study, as it comprises the framework for constructing geological models and assessing hydrocarbon reserves. Today, the detailed interwell correlation relies on manual analysis of well-logging data. Thus, it is time-consuming and of a subjective nature. The essence of the interwell correlation constitutes an assessment of the similarities between geological profiles. There were many attempts to automate the process of interwell correlation by means of rule-based approaches, classic machine learning approaches, and deep learning approaches in the past. However, most approaches are of limited usage and inherent subjectivity of experts. We propose a novel framework to solve the geological profile similarity estimation based on a deep learning model. Our similarity model takes well-logging data as input and provides the similarity of wells as output. The developed framework enables (1) extracting patterns and essential characteristics of geological profiles within the wells and (2) model training following the unsupervised paradigm without the need for manual analysis and interpretation of well-logging data. For model testing, we used two open datasets originating in New Zealand and Norway. Our data-based similarity models provide high performance: the accuracy of our model is 0.926 compared to 0.787 for baselines based on the popular gradient boosting approach. With them, an oil&gas practitioner can improve interwell correlation quality and reduce operation time.

In unconventional reservoir sweet-spot identification, brittleness is an important parameter that is used as an easiness measure of production from low permeability reservoirs. In shaly reservoirs, production is realized from hydraulic fracturing, which depends on how brittle the rock is-as it opens natural fractures and also creates new fractures. A measure of brittleness, brittleness index, is obtained through elastic properties of the rock. In practice, problems arise using this method to predict brittleness because of the limited availability of elastic logs. To address this issue, machine learning techniques are adopted to predict brittleness at well locations from readily available geophysical logs and spatially using 3D seismic data. The geophysical logs available as input are gamma ray, neutron, sonic, photoelectric factor, and density logs while the seismic is a post-stack time migrated data of high quality. Support Vector Regression, Gradient Boosting, and Artificial Neural Network are used to predict the brittleness from the geophysical logs and Texture Model Regression to invert the brittleness from the seismic data. The Gradient Boosting outperformed the other algorithms in predicting brittleness. The result of this research further demonstrates the application of machine learning, and how these tools can be leveraged to create data-driven solutions to geophysical problems. Also, the seismic inversion of brittleness shows promising results that will be further investigated in the future.

Second International Meeting for Applied Geoscience & Energy, 2022

Determination of seismic lithology, porosity, and pore fluid requires detailed modeling of petrophysical logs to improve the correlation with a seismic AVO response. Unfortunately, acquiring a full set of logs for all wells in a seismic survey is unpractical, and estimating sonic, shear and density using empirical relationships is the standard approach. While these relationships have worked for recon analysis, they have generally not given the details needed for accurate geophysical analysis. Machine Learning has given us a new way of investigating these relationships, and this presentation will provide the insights we learned from analysis using over 138 wells with DT, Vs & RHOB logs from the North Sea, Australia, and Canada. This paper builds on the SEG Machine Learning workshop in the spring of 2022 on a lithological determination using machine learning (Emery et al. 2021).

Second International Meeting for Applied Geoscience & Energy

We show an integrated reservoir characterization study of a deepwater oil field in the Gulf of Mexico (GOM). Seismic data were acquired in our study area with both NATS and WATS (narrow-and wide-azimuth towed streamer) configurations. Data are of good quality apart from a large void in the WATS dataset, as these data were shot when a platform was in place over the crest of the field. NATS data were shot previously with no obstruction, however the two surveys have different azimuthal and angular coverage. To understand the impact and illumination of each dataset, we undertook a modeling study which supported merging the two datasets. Additionally, a poor data zone exists to the north of the field due to a salt overhang. We performed an extensive post-migration gather conditioning workflow to further improve image fidelity for better fault interpretation and reservoir definition. Additionally, we carried out a seismic rock property modeling study using wells in our study area to determine the expected seismic response of the reservoir. This demonstrated that there is an optimal projection for both lithology and fluid separation, but this separation is difficult to detect seismically. To help identify compartmenting faults across the field, we utilized a machine learning-based approach to quickly help predict the location of the faults. We also trained a probabilistic neural network to predict areas of high net-to-gross reservoir which we correlate to wells within our study area. This study demonstrates integrated applications in imaging, rock properties modeling and seismic interpretation using machine learning methods, where the approaches leveraged here could be used to further aid in reservoir interpretation and characterization elsewhere in the GOM.

SPE Journal, 2020

Petroleum reservoirs are often associated with multiple target zones or a single zone adjacent to nonproductive intervals. Real-time geo-steering therefore becomes important to remain in zone or to dynamically steer toward a target. This requires knowledge of the petro-physical/rock mechanical properties of the rock surrounding the bit. Although logging while drilling can provide this information, a cost-effective and almost-real-time solution is lacking. In general, there is a depth lag, and therefore, a time delay, between what the logging-while-drilling sub relays to the surface and the bit performance. This study focuses on relating drill-bit-and drillstring-performance data in a machine-learning (ML) workflow to predict the lithology at the bit while drilling. The method we are proposing offers several advantages in terms of cost and time savings for real-time geosteering applications, where going out of zone requires costly intervention. In this study, we have used a public data set from Volve Field on the Norwegian continental shelf. Within our proposed workflow, as a first step, logs sensitive to lithology [such as density, gamma ray (GR), and sonic] are grouped into three electrofacies. We also had access to core data, which helped us interpret the electrofacies in terms of mineralogy. The three electrofacies corresponded to quartz-rich (sandstone/siltstone), clay-rich (shale), and carbonate-rich (limestone) lithologies. The next step is to predict the electrofacies using various measurement-while-drilling (MWD) variables, such as rate of penetration (ROP), weight on bit (WOB), and several others that are monitored in real time. Supervised classification algorithms were used to relate real-time surface measurements to lithology. The algorithms were able to predict lithology in test wells with more than 80% accuracy. These results, although encouraging, constitute a small step toward drilling-automation/advisory systems. The development of such systems can prevent costly out-of-zone drilling and minimize rig time and equipment use, thereby potentially reducing capital expenditures. This study was specifically performed in Volve Field in the North Sea using petrophysical and surface drilling data from vertical wells. However, the workflow has a potential to be extended to other formations in other fields in different well types. Introduction ML has become increasingly important in exploration-and-production operations to gain insights from the vast amounts of heterogeneous and real-time data being collected from multiple sources in the air (drones), in the well (fiber optics), on the drillstring and casing, from artificial lift, from production monitoring, and others. The acceleration in the implementation of ML has largely been driven by the democratization of computer power, robust algorithms and workflows, and, most importantly, the thousands of wells being drilled and completed each year. The recent volatility in oil prices, specifically in 2008 and 2015, has also given a great impetus to minimizing costs and increasing the efficiency of drilling, completing, and producing these reservoirs. Saputelli et al. (2003) identified the need for a multilayer and multiscale integrated strategy for making decisions while drilling to optimize drilling performance, initial well-completion design, and well productivity. ML applications span all disciplines, including petrophysics, geophysics, reservoir, production, completions, and drilling. Kale et al. (2010) and Gupta et al. (2017a, 2017b) integrated core, log, and production data using techniques such as k-means, self-organizing maps (SOMs), and support-vector machines to predict sweet spots within the Eagle Ford, Wolfcamp, Woodford, and Barnett formations. Zhao et al. (2015) applied several classification algorithms, including generative topographic mapping, Gaussian-mixture models, and neural networks, for seismic-facies classification. Gupta et al. (2019) compared several tree-based regression techniques to predict synthetic sonic using other petrophysical properties. Other ML applications include the work of Bhatt and Helle (2001) and Bhatt (2002) to predict reservoir properties using MWD data and wireline logs, while Ahmadi (2016) used the support-vector-machines technique to predict drilling-fluid rheology under a variety of wellbore conditions and different types of drilling fluids such as oil-based muds, water-based muds, and gas aphrons. Unrau et al. (2017) developed a ML-based adaptive-alarm system to detect subtle changes in mud volume and flow rate out of the well for identifying potential well kicks and lost circulation. Pollock et al. (2018) used a combination of reinforcement learning and neural networks on horizontal wells in the Permian and Appalachian basins to minimize tortuosity and maximize ROP. Marx et al. (2014) developed a multilayer neural-network model and used real-time drilling parameters to forecast permeability, lithology, pore pressure, and bit wear. Kanfar (2012) applied neural-network models to predict lithofacies and rock permeability using logging-while-drilling logs in real time. Gu et al. (2018) used a probabilistic neural network for complex-lithofacies identification and obtained prediction accuracy of up to 88%. Han et al. (2019) used long-short-term-memory recurrent neural networks to predict the ROP using commonly recorded drilling data, such as WOB, rotary speed, bit wear, and bit size. They observed that the drilling rates at previous timesteps have a strong effect on ROP at any given current timestep. Noshi and Schubert (2019) also tested random forest, neural networks, support-vector regression, k-nearest neighbor, and gradient-boosting machine to maximize ROP by using drilling variables, such as WOB, rev/min, flow rate, mechanical specific energy (MSE), bit-run distance, and GR. Yu et al. (2019) used Kalman filters, convolutional neural networks

First International Meeting for Applied Geoscience & Energy Expanded Abstracts

In unconventional reservoir sweet-spot identification, brittleness is an important proxy used as a measure of easiness for oil and gas production. Production from this low permeability reservoir is realized by hydraulic fracturing, which depends on how brittle the rock is-as it opens natural fractures and also creates new fractures. An estimate of brittleness, brittleness index, is obtained at well locations through a mathematical combination of elastic logs. In practice, problems arise to predict brittleness because of the limited availability of elastic logs and sparsity of wells to understand the lateral variation of brittleness. To address this problem, machine learning techniques are adopted to predict brittleness at well locations from readily available geophysical logs and spatially using continuous 3D seismic data. This study tests machine learning algorithms to forecast reservoir brittleness throughout the entire reservoir interval using well logs and 3D seismic in a shale gas field of central Appalachian Basin. Our results show the effectiveness of using gradient boosting to predict brittleness from gamma ray, density, and neutron logs with a training and testing R 2 score of 0.95 and 0.85, respectively. We demonstrate a novel application of seismic texture as an indicator for brittleness through the qualitative agreement of the inversion output with the blind well and also the fracture attribute.

Journal of Petroleum Science and Engineering, 2019

Keypoint detection is critical in image recognitions. Deep learning such as convolutional neural network (CNN) has recently demonstrated its tremendous success in detecting image keypoints over conventional image processing methodologies. The deep learning solutions, however, heavily rely on labeling target images for their reliability and accuracy. Unfortunately, most image datasets do not have all labels marked. To address this problem, this paper presents an effective and novel deep learning solution, Masked Loss Residual Convolutional Neural Network (ML-ResNet), to facial keypoint detection on the datasets that have missing target labels. The core of ML-ResNet is a masked loss objective function that ignores the error in predicting the missing target keypoints in the output layer of a CNN. To compensate for the loss induced by the masked loss objective function that likely results in overfitting, ML-ResNet is designed of a data augmentation strategy to increase the number of training data. The performance of ML-ResNet has been evaluated on the image dataset from Kaggle Facial Keypoints Detection competition, which consists of 7049 training images, but with only 2140 images that have full target keypoints labeled. In the experiments, ML-ResNet is compared to a pioneer literature CNN facial keypoint detection work. The experiment results clearly show that the proposed ML-ResNet is robust and advantageous in training CNNs on datasets with missing target values. ML-ResNet can improve the learning time by 30% during the training and the detection accuracy by eight times in facial keypoint detection.

MEOS 2023, 2023

Well logs present a concise, detailed plot of formation parameters versus depth. From these plots, interpreters can identify lithologies, differentiate between porous and nonporous rock and quickly recognize pay zones in subsurface formations. The ability to interpret a log hugely relies on interpreter’s ability to recognize patterns, past experiences and knowledge of each measurement. The standard approach has been to manually correct logs for various anomalies and normalize them at the field scale, select different input parameters to calculate the formation and fluid volumes, a very time-consuming and often very subjective approach. The workload is even more time-consuming for mature fields where log data has been collected across multiple vintages of logging tools and from multiple vendors. The future of petrophysical evaluation workflows that are rapidly heading towards increased efficiency, accuracy, and objectivity through smart automation. In this paper, we show the application of unsupervised class-based machine learning algorithm to automate well log processing and interpretation. The algorithm classifies the provided input data into set of classes that are converted to zones to drive petrophysical interpretation. The main advantage is reducing the turnaround time of the interpretation and eliminating subjective inconsistencies often encountered with standard interpretation approaches. The algorithm uses cross-entropy clustering (CEC), Gaussian mixture model (GMM) and Hidden Markov Model (HMM) that identifies locally stationary zones sharing similar statistical properties in logs, and then propagates zonation information from training-wells to other wells. The training phase involves key wells which best represent the formation and associated heterogeneities to automatically generate classes (clusters), the resulted model is then used to reconstruct inputs and outputs with uncertainty and outlier flags for cross-check and validation. The model is then applied to predict the same set of zones in the new wells that require interpretation and predict output curves. We present examples of this algorithm applied on data from diagenetically altered carbonate & clastic reservoirs in the Middle East where the workflow speeded-up the petrophysical analysis process, reduced analyst bias and improved consistency result between one well to another within the same field.