Connectivity of material constituents govern the transport, mechanical, chemical, thermal, and electromagnetic properties. Energy storage, recovery and conversion depends on connectivity of material constituents. High-resolution microscopy image of a material captures the microstructural aspects describing the distribution, topology and morphology of various material constituents. In this study, six metrics are developed and tested for quantifying the connectivity of material constituents as captured in the high-resolution microscopy images. The six metrics are as follows: geobody connectivity metric based on percolation theory, Euler number based on integral geometry, indicator variogram based on geostatistics, two-point cluster function, connectivity function, and travel-time histogram based on fast marching method. The performances of these metrics are tested on 3000 images representing six levels of connectivity. The metrics are also evaluated on the organic constituent captured in the scanning electron microscopy (SEM) images of organic-rich shale samples. The connectivity function and travel-time histogram based on fast-marching method are the most robust and reliable metrics. Material constituents exhibiting high connectivity result in large values of average travel time computed using fast-marching method and average connected distance computed using connectivity function. The proposed metrics will standardize and speed-up the analysis of connectivity to facilitate the characterization of properties and processes of energy-relevant materials.

Subsurface sequestration of carbon dioxide (CO2) requires long-term monitoring of the injected CO2 plume to prevent CO2 leakage along the wellbore or across the caprock. Accurate knowledge of the location and movement of the injected CO2 is crucial for risk management at a geological CO2-storage complex. Conventional methods for locating/assessing the injected CO2 plume in the subsurface assume a geophysical model, which is specific and may not be applicable to all types of CO2-injection reservoirs and scenarios. We developed an unsupervised-learning-based visualization of the subsurface CO2 plume that adapts and scales based on the data without requiring an assumption of the geophysical model. The data-processing workflow was applied to the cross-well tomography data from the SECARB Cranfield carbon geo-sequestration project. A multi-level clustering approach was developed to account for data imbalance due to the absence of CO2 in the large portion of the imaged reservoir. The first level of clustering differentiated CO2-bearing regions from the non-CO2 bearing regions and achieved a silhouette score of 0.85, a Calinski-Harabasz index of 160666, and a Davies-Bouldin index of 0.43, which are indicative of high quality, reliable clustering. The second level of clustering further differentiated the CO2-bearing regions into regions containing low, medium, and high CO2 content. Overall, the multi-level clustering achieved a silhouette score, Calinski-Harabasz index, and Davies-Bouldin index of 0.74, 59656, and 0.32, which confirm the high quality and reliability of the newly proposed unsupervised-learning-based visualization. Three distinct clustering techniques, namely k-means, mean-shift, and agglomerative, generated similar visualizations. In terms of the adjusted Rand index, the similarity of clusters identified by the three distinct clustering techniques is around 0.98, which indicates the robustness of the cluster labels assigned to various regions of the CO2-injection reservoir. Further, we find certain geophysical signatures, such as Fourier transform and wavelet transform, to be highly relevant and informative indicators of the spatial distribution of CO2 content.

Ultrasonic transmission is sensitive to the variation in mechanical properties of materials. Wave propagation through fractured media introduces changes in the frequency content, travel time and transmission coefficient of the wave. A workflow based on physics-driven unsupervised learning is developed to process the transmitted ultrasonic-shear waveforms to non-invasively visualize the geomechanical alterations due to hydraulic fracturing of a tight sandstone. Novelty of the work involves the assignment of physically consistent clusters to the measurements of shear waveforms across the axial and frontal planes by incorporating the travel time of the peak of spectral energy and transmission coefficient. The proposed workflow generates maps of geomechanical alterations across the frontal and axial planes of the sample. The outputs of the workflow are in good agreement with independent techniques viz. acoustic emission and X-ray computed tomography. The proposed workflow can be adapted for improved fracture characterization in the subsurface when processing sonic-logging, cross-wellbore seismic or surface seismic waveform data.

High water cut has been an issue in the Delaware basin for many years now. Volume of produced water continue to increase, resulting in adverse environmental impacts and higher reservoir-management costs. To address these problems, a data-driven workflow has been developed to pre-emptively identify the high water-cut wells. The workflow uses unsupervised pseudo-rock typing followed by supervised classification trained on well logs from 17 wells in the Delaware basin. The workflow requires a suite of 5 well logs from a 500-ft depth interval surrounding the kick-off points of these wells, which includes 200 ft above and 300 ft below the KOP. First, the well logs are clustered into 5 pseudo-rock types using multi-level clustering. Using statistical features extracted from these 5 pseudo-rock types, 3 supervised classifiers, namely K-nearest neighbor, support vector machine, and logistic regression, are trained and tested to detect the high water-cut wells. Over 100 cross validations, the 3 classifiers perform at a median Matthew’s Correlation Coefficient (MCC) of 0.90. The kurtosis of the neutron porosity log response of the pseudo-rock type A0, interpreted as a shale lithology, is the most The submitted paper is currently under review. Dr. Sid Misra is the lead investigator on this topic. informative/relevant signature associated with high water cut. Next, the presence of pseudo-rock type A1, interpreted as high-permeability lithology, is an informative signature of low water-cut wells. The kurtosis of the density porosity log response of the pseudo-rock type B0, interpreted as carbonate lithology, and the presence of pseudo-rock type B1, interpreted as a tight sandstone lithology, serve as informative signatures for differentiating high water cut wells from low water cut wells.

1. Abstract Wave propagation and diffusive transport phenomena could work as evidence of the mechanical discontinuities in material. For the problem of poor efficiency of the existing fracture simulation methods, this paper proposes crack-bearing material characterization approach by processing wave travel-time using seven data-driven classification techniques. To that end, we perform classification models to predict discontinuities orientation, dispersion, and spatial distribution prediction by learning from the different-waves simulation model. The travel-time measured by multiple sensors placed around the material perform as our input data of machine learning method. As a result, this work found that machine learning models exhibit best classification performance on classifying crack dominant orientations. Combination of compressional wave and shear wave are enough to capture the crack information in the material, however, the pressure diffusion also able to optimize our algorithms. Voting classifier and gradient boosting classifier perform the best for purposes of characterization. When compare the performance of different mechanical discontinuities, embedded closed discontinuities shows high accuracy than open discontinuities on the classification models.

Machine learning has led to improvements in the efficiency and efficacy of subsurface engineering and characterization efforts that benefits the hydrocarbon and geothermal exploration and production as well as in carbon geo-sequestration. There have been rapid increases in sensor deployment, data acquisition, data storage, and data processing for purposes of geothermal/fossil energy development and exploration along with carbon geo-sequestration. This has promoted large-scale development of data-driven methods, machine learning and data analytics workflows to find and extract energy and material resources from the subsurface earth. Subsurface data ranges from nano-scale to kilometer-scale passive as well as active measurements in the form of physical fluid/solid samples, images, 3D scans, time-series data, waveforms, and depth-based multi-modal signals representing various physical phenomena, ranging from transport, chemical, mechanical, electrical, and thermal properties, to name a few. Integration of such varied multimodal, multipoint, time-varying data sources being acquired at varying scales, rates, resolutions, and volumes mandates robust machine learning methods to better characterize and engineer the subsurface earth. This review paper lays out popular machine learning applications in exploration, extraction, and recovery of subsurface energy resources, primarily in hydrocarbon exploration and production industry with potential applications in geothermal energy production and carbon geo-sequestration.

Mechanical discontinuity embedded in a material plays an essential role in determining the bulk mechanical, physical, and chemical properties. The ability to control mechanical discontinuity is relevant for industries dependent on natural, synthetic and composite materials, e.g. construction, aerospace, oil and gas, ceramics, metal, and geothermal industries, to name a few. The paper is a proof-of-concept development and deployment of a reinforcement learning framework to control the propagation of mechanical discontinuity. The reinforcement learning framework is coupled with an OpenAI-Gym-based environment that uses the mechanistic equation governing the propagation of mechanical discontinuity. Learning agent does not explicitly know about the underlying physics of propagation of discontinuity; nonetheless, the learning agent can infer the control strategy by continuously interacting the simulation environment. The Markov decision process, which includes state, action and reward, had to be carefully designed to obtain a good control policy. The deep deterministic policy gradient (DDPG) algorithm is implemented for learning continuous actions for the desired reinforcement learning. It is also observed that the training efficiency is strongly determined by the formulation of reward function. An adaptive reward function involving reward shaping improves the training. The reward function that forces the learning agent to stay on the shortest linear path between crack tip and goal point performs much better than the reward function that aims to reach closest to the goal point in minimum number of steps. After close to 500 training episodes, the reinforcement learning framework successfully controlled the propagation of discontinuity in a material despite the complexity of the propagation pathway determined by multiple goal points.

Fracture characterization is essential to hydrocarbon and geothermal exploration and production as well as carbon geo-sequestration. Active and passive measurements have been extensively applied to map fractures across various scales. Active and passive measurements have inherent advantages and limitations that complement each other. In this study, to improve the mapping of embedded fractures and the surrounding geomechanically altered regions, we integrate active shear-wave transmission measurement with passive acoustic-emission measurement collected during a lab-scale hydraulic-fracturing experiment. The proposed approach leverages the advantages of the two modalities of measurements, while minimizing the limitations. Polarized shear-wave transmission (active) measurement was collected before and after the hydraulic fracturing, and the acoustic emission (passive) measurement was collected during hydraulic fracturing. Two sets of two-dimensional maps of fracture and fracture-induced damage in axial, median and frontal planes were obtained by separately processing the physics-informed transformations of active and passive measurements. These 2D maps are then fed to wavelet-based image-fusion technique to integrate the two sources of information for the reliable mapping/imaging of the embedded fractures and the surrounding geomechanically altered regions. Plain word summary Fractures and geomechanical alterations embedded in a solid material can be characterized using active or passive measurements. Signal type, timing, duration and location of the source is controlled during an active measurement to allow optimal interactions of the propagating wave and the fractured material, whereas the signals due to waves generated by the processes of fracturing and geomechanical alterations are captured during a passive measurement. Active and passive measurements have advantages and limitations that complement each other. In this study, physics-informed transformations and wavelet-based image fusion are used to integrate the active and passive measurements to get the best of the two measurement modalities for purposes of reliable imaging of fractures and the surrounding geomechanically altered regions. Key points  Integration of active and passive measurements will leverage their strengths and minimize the limitations of individual measurement techniques.  To the best of authors’ knowledge, no reference exists where the two modalities are combined to characterize the fractures either in the field or laboratory scale.  We integrate active ultrasonic measurements with passive acoustic-emission measurements using physics-informed transformations and data fusion techniques.  The proposed method improves the imaging of embedded fractures and the surrounding geomechanically altered regions.  Two-dimensional maps of fracture and fracturing-induced damage is generated in axial, median and frontal planes.

A robust machine learning (ML) workflow was developed to identify kerogen and hydrocarbon storage and transport pathways in the scanning electron microscopy (SEM) maps of organic-rich shale samples from two different formations, namely Wolfcamp Shale and Barnett Shale. Kerogen is most often wrongly identified as pores/cracks are poor when the machine learning model is developed on one shale formation and applied on the other shale formation. ML workflow developed only on Barnett shale cannot detect cracks in the Wolfcamp samples. The best performing machine learning approach learnt from both the formations and exhibits an average F1 scores of 0.99 and 0.91 on the inner-region and outer-region pixels, respectively. The machine learning workflow performs better on Barnett as compared to Wolfcamp. Barnett shale, in comparison to Wolfcamp shale, provides better generalizable features and more complex microstructural aspects that are harder to 2 identify. Overall, it is easier to identify kerogen as compared to pores and cracks due to their distribution, availability, connectivity, and pixel intensity. By learning from merely 50 thousand pixels with corresponding labels, the proposed machine learning workflow can successfully identify the hydrocarbon storage and transport pathways and kerogen in a large 20-GB dataset containing approximately 10 billion pixels.

Enhanced geothermal systems can provide a substantial share of the global energy demand if certain hurdles are overcome. One such hurdle is the accurate imaging of the fracture networks created in subsurface through hydraulic stimulation of these systems. Microseismicity associated with the stimulation is the primary means to locate the event hypocenters for estimating the stimulated rock volume. Using the data from a single three-component accelerometer in a monitoring well, the polarization features viz. azimuth, incidence, rectilinearity, and planarity are used as inputs for the unsupervised manifold approximation followed by clustering. We show that density-based clusters in the projected 3D space correspond to distinct types of hydraulically fractured zones around the injection point, thereby refining the interpretation of the microseismic cloud. The temporal evolution of these clusters can be used to track fracture creation and propagation for the various types of fracture zones.

Mechanical wave transmission through a material is influenced by the mechanical discontinuity in the material. The propagation of embedded discontinuities can be monitored by analyzing the wave-transmission measurements recorded by a multipoint sensor system placed on the surface of the material. In our study, robust monitoring of the propagation of a mechanical discontinuity is achieved by using supervised learning followed by data-driven causal discovery to process the multipoint waveform measurements resulting from a single impulse source. The new data-driven causal-discovery workflow jointly processes the nine 25-µs waveforms measured by the multipoint sensor system comprising 9 sensors. The proposed workflow can monitor the propagation of mechanical discontinuity through three stages, namely initial, intermediate, and final stages. The workflow considers the wave attenuation, dispersion and multiple wave-propagation modes. Among various feature reduction techniques ranging from decomposition methods to manifold approximation methods, the features derived based on statistical parameterizations of the measured waveforms lead to reliable monitoring that is robust to changes in precision, resolution, and signal-to-noise ratio of the multipoint sensor measurements. Causal signatures have been successfully identified in the multipoint waveform measurements. The numbers of zero-crossing, negative-turning, and positive in the waveforms are the strongest causal signatures of crack propagation. Higher order moments of the waveforms, such as variance, skewness and kurtosis, are also strong causal signatures of crack propagation. The newly discovered causal signatures confirm that the statistical correlations and conventional feature rankings are not always statistically significant indicators of causality.

The continuous rise in global energy demand requires the production of oil and gas from unconventional shale resources. One major concern has been the large volumes of produced water associated with the production of hydrocarbon from the shale resources. We developed a data-driven workflow for identifying potentially high water-producing wells drilled in unconventional shale formation. To that end, we applied unsupervised learning followed by supervised learning to process five conventional well logs, namely shallow and deep resistivity logs, density porosity logs, neutron porosity logs, and gamma ray logs, from a well drilled in an unconventional shale formation. A novelty of our study is the use of clustering methods to generate pseudo-lithology that is fed into a classifier for the desired identification of the excess water producing wells. The data-driven workflow was tested on 23 wells in Gulf coast basin and 29 wells in Fort Worth basin. Fort Worth and Gulf Coast basins in the U.S. are highly productive shale basins that produce 380 million cubic feet of gas and 1.74 million barrels of crude oil every day. Additionally, we identified geophysical signatures that explain the excess water production from the wells drilled in unconventional shale reservoirs. For future work, molecular simulation, core analysis, and advanced well logs studies need to be incorporated for a better explanation of the causes of excess water production in unconventional reservoirs.