Search

This study utilizes machine learning to quantify CO2 plume extents by analyzing microseismic data from the Illinois Basin Decatur Project (IBDP). Leveraging a unique dataset of well logs, microseismic records, and CO2 injection metrics, this work aims to predict the temporal evolution of subsurface CO2 saturation plumes. The findings illustrate that machine learning can predict plume dynamics, revealing vertical clustering of microseismic events over distinct time periods within certain proximities to the injection well, consistent with an invasion percolation model. The buoyant CO2 plume partially trapped within sandstone intervals periodically breaches localized barriers or baffles, which act as leaky seals and impede vertical migration until buoyancy overcomes gravity and capillary forces, leading to breakthroughs along vertical zones of weakness. Between different unsupervised clustering techniques, K-Means and DBSCAN were applied and analyzed in detail, where K-means outperformed DBSCAN in this specific study by indicating the combination of the highest Silhouette Score and the lowest Davies–Bouldin Index. The predictive capability of machine learning models in quantifying CO2 saturation plume extension is significant for real-time monitoring and management of CO2 sequestration sites. The models exhibit high accuracy, validated against physical models and injection data from the IBDP, reinforcing the viability of CO2 geological sequestration as a climate change mitigation strategy and enhancing advanced tools for safe management of these operations. [ABSTRACT FROM AUTHOR]

Abstract Integrated petrophysical analysis is used in conventional and unconventional reservoirs to visualize the distribution pattern of different lithofacies and interpret depositional environments. In this study, core data from 17 wells and well logs from 517 wells are used to construct 3D data-driven lithofacies models for the upper and lower shale members in the Bakken Formation of the Williston basin in North Dakota, United States. The principal objective of this multi-scale (core, well log, and regional) study is to use the petrophysical response of lithofacies defined in core from a very limited number of wells to identify different shale lithofacies in the Bakken Formation in the numerous and geographically extensive wells with wireline log suites. Shale lithofacies are defined in the core using quantitative mineralogy, Total Organic Carbon content, and petrophysical properties. These lithofacies are calibrated to advanced geochemical spectroscopy logs and conventional well logs at well scales. A machine learning algorithm, Support Vector Machine, is used to recognize the pattern of different shale lithofacies, associated with basic petrophysical parameters from ubiquitous conventional well log suites. Sequential Indicator Simulation is used to populate all lithofacies in a 3D grid, covering a large portion of the Williston basin in North Dakota. The results show that the Bakken shale members are vertically and laterally heterogeneous, but are successfully classified into five different lithofacies. Organic-rich shale lithofacies outweigh the proportion of organic-lean shale lithofacies. It appears several factors influenced the pattern of shale lithofacies distribution of the Bakken Formation. Highlights • Application of machine learning for quantitative shale facies classification. • Implementation of Support Vector Machine algorithm to predict the presence of complex shale facies. • Vertical and lateral heterogeneity of the upper and lower Bakken shale members. [ABSTRACT FROM AUTHOR]

Quantitative lithofacies modeling is important to understand the depositional and diagenetic history, and hydrocarbon potential of unconventional resources at a regional scale. The complex heterogeneous nature and large data dimensionality of unconventional mudstone reservoirs increase the challenge of lithofacies interpretation by conventional qualitative methods. Quantitative shale lithofacies, which are meaningful, mappable, and predictable at core, well log, and regional scales, can be defined based on mineralogy and Total Organic Carbon (TOC) derived from core analysis and advanced geochemical spectroscopy logs (e.g. Pulsed Neutron Spectroscopy, PNS). However, access to numerous and widespread core samples and geochemical log responses is typically limited by cost and time. We apply different mathematical techniques to ubiquitous conventional well log suites calibrated to rock types, defined by the limited number of wells with high-quality core and geochemical logs. The documented interrelationships between lithofacies and conventional logs are propagated with a quantified degree of accuracy in wells without advanced log or core data. Our study addresses issues of different approaches of quantitative lithofacies classification and prediction techniques from well logs. Various data-driven supervised and unsupervised computational approaches, such as Support Vector Machine (SVM), Artificial Neural Network (ANN), Self-Organizing Map (SOM) and Multi-Resolution Graph-based Clustering (MRGC), are applied and compared to reduce uncertainty of propagating single-well based lithofacies analysis, and efficiently understand geological trends. Two different dataset from the Devonian Bakken and Mahantango-Marcellus Shale formations in North America are used, in order to undertake a comparative assessment of computational techniques for lithofacies characterization. Original shale lithofacies, defined from geochemical logs and core data, are used to compare the results of selected supervised and unsupervised computational approaches. The results show that both Bakken and Mahantango-Marcellus shale members are vertically and laterally heterogeneous, but can be classified into at least five mudstone lithofacies, along with calcareous siltstone and limestone lithofacies. SVM works better than other techniques for lithofacies classification and prediction in reduced computational time, no iteration, and with highly repeatable results. Accuracy of lithofacies prediction increases if the algorithms are supervised with geological rules. [ABSTRACT FROM AUTHOR]

Abstract Predicting gas production from stimulated unconventional reservoirs has always been a challenge to oil and gas companies because the physics of fluid flow through such reservoirs is not well understood. On the other hand, modeling gas production from shale reservoirs is a complex multi-variate problem that requires proper integration of data from multiple disciplines, such as geology, petrophysics, geomechanics, completions, and surface-based measurements, etc. This study demonstrates the application of data-driven machine learning algorithms, integrating geoscientific, distributed acoustic sensing (DAS), distributed temperature sensing (DTS) fiber-optic, completions, flow scanner production log, and surface data to model daily gas production from a 28-stage stimulated horizontal well drilled in the Marcellus Shale of the Appalachian basin in North America. In addition, this study aims to utilize the data from a fiber-optic monitoring system, such as DAS and DTS to evaluate the well performance. We build supervised data-driven machine learning models using Random Forest (RF), Artificial Neural Network (ANN), and Support Vector Machine (SVM) algorithms to understand the well performance and forecast the daily gas production. The study compares the forecasted results, tests the level of accuracy, and addresses different issues with these machine learning algorithms. A spatio-temporal database is constructed and used to perform sensitivity analyses to identify the key drivers for gas production. The results show that RF and ANN algorithms can be used to predict daily gas production with significantly high accuracy; however, RF algorithm is the best predictor in terms of highest accuracy (96%), less computational time and cost. Based on sensitivity analysis, it appears that Poisson's ratio, minimum horizontal stress, DAS, casing pressure, gamma log are the most important parameters to predict gas production. Highlights • Application of machine learning algorithms to forecast daily gas production from unconventional reservoirs using large multi-sensor dataset. • Random Forest and Artificial Neural Network can model daily gas production from each stage efficiently. [ABSTRACT FROM AUTHOR]