Abstract A machine learning-driven workflow has been designed to provide drilling engineers and geosteering geologists with a high-resolution earth model populated with elastic, petrophysical or geomechanical properties to assist with safe and efficient drilling. The technology can be deployed pre-drill or in real-time as a well is being actively drilled. The real-time deployment affords ahead-of-the bit and away-from-the-bit awareness that may help to mitigate pre-drill sub-surface uncertainties. Examples include shallow water flow, shallow gas, depleted formations, and the presence of salt. These phenomena may all cause excessive non-productive time (NPT) or may cause the wellbore's possible loss in extreme circumstances. Introduction A subset of a 3D prestack depth imaged seismic data covering 51 square miles (132 km 2 ) acquired in Midland and Upton Counties, Texas, was used in this study. Wireline log data from seven wells were made available. Five of the seven wells were incorporated in the training model and the two remaining wells were retained as "blind" tests for validation purposes. Results were achieved in three weeks with a significant increase in resolution compared to seismic data (60 ft to 15 ft). In this study, a suite of pre-drill attributes was generated. In addition, a simulated drilling scenario demonstrating the real-time deployment of the method is also described.

Abstract Robust production forecasting allows for optimal resource recovery through efficient field management strategies. In hydraulically fractured unconventional reservoirs, the physics of fluid flow and transport processes is not well understood and the presence of, and transitions between multiple flow regimes further complicate forecasting. An important goal for field operators is to obtain a fast and reliable forecast with minimal historical production data. The abundance of wells drilled in fractured tight formations and continuous data acquisition effort motivate the use of data-driven forecast methods. However, traditional data-driven forecast methods require sufficient training data from an extended period of production for any target well and may have limited practical use. In this paper, a deep recurrent neural network (RNN) model is developed for robust long-term production forecasting in unconventional reservoirs. As input data, the model takes completion parameters, formation and fluid properties, operating controls, and early (i.e., 3-6 months) production response data. The model is trained on a collection of historical production data across multiple flow regimes, control settings, and the corresponding well properties from multiple shale plays. The proposed RNN model can predict oil, water, and gas production as multivariate time-series under varying operating controls. Once the forecast model is trained, it can be used to obtain a one-step forecast by feeding the model with input well properties, operating controls, and a short initial production. The long-term forecast is obtained by either recursively feeding the model with forecast results from the preceding timesteps or by training the model for multi-step ahead predictions. Unlike other applications of RNN that require a long history of production data for training, our model employs transfer learning by combining early production data from the target well with the long-term dynamics captured from historical production data in other wells. We illustrate our approach using synthetic datasets and a case study from Bakken Play in North Dakota.

Abstract A physics-assisted deep learning model is presented to facilitate transfer learning in unconventional reservoirs by integrating the complementary strengths of physics-based and data-driven predictive models. The developed model uses a deep learning architecture to map formation properties to their corresponding production responses using a low-dimensional feature space representation. Transfer learning is accomplished by first training the network weights using production data from a mature shale play and combining the learned weights with limited data from a new unconventional field to generate a predictive model. The simulated data provides approximate production predictions for input parameters of the target field, for which the source data may not provide a good prediction. The resulting model has a superior performance to simulation-based and data-driven predictions alone. The results indicate that (1) physics-based simulated data can facilitate production predictions when out-of-range (unseen) input parameters have to extrapolate from data, and (2) transferring the weights learned from the source field to the target field can add valuable information to enhance the prediction performance of the target field. Introduction There has been a significant increase in the development of unconventional reservoirs in recent decades, particularly tight oil reservoirs in the United States. These tight oil formations often have extremely low permeability and are not economically exploitable using traditional drilling and completion techniques used for conventional high permeability reservoirs. Fortunately, due to the advancement in key technologies, in particular hydraulic fracturing and horizontal drilling, these low permeability shale formations can be stimulated to induce the production of hydrocarbons (King, 2010). However, the traditional flow equations that are well documented and studied for conventional wells have not been applied to tight oil reservoirs. Traditionally, the development of conventional reservoirs relies heavily on numerical reservoir simulators (Altman et al., 2020, Aziz et al., 1979). These simulators provide a reliable prediction that assists in decision-making for resource development. They are built using well-studied and trusted physics-based analytical expressions that are solved during the simulation to provide prediction responses. Since these physics-based simulation models are based on analytical equations, they have the advantage of being able to produce a prediction response for any possible range of input variables, and this allows them to easily extrapolate to any range of data. Additionally, simulation models can be run for any combination of input variables and enable the collection of large amounts of simulated data for all possible scenarios. However, the equations used in simulation models are not able to provide dependable predictions for unconventional tight oil resources since the complex physical relation of flow from tight formations and fractures along with the fracture generation is poorly understood. Additionally, there are various components in the field that are not always accounted for when building simulation models (Fung et al., 2016). These limitations make simulation models unreliable to be used to develop unconventional reservoirs.

Abstract Bottomhole pressure is an essential parameter to measure well performance, design and optimize artificial lift and calculate well deliverability. However, it is rarely continuously measured in most unconventional wells. It is typically estimated from the wellhead pressure, using physics-based steady-state flow correlations, originally derived as empirical or mechanistic models. However, these methods do not scale adequately for large fields, where computational time increases proportionally with the well count. Another common drawback is accuracy. As flow correlations have been developed with limited datasets and simplifying assumptions, they imperfectly capture the physics, often failing to generalize and fully capture the flow behavior in a field under all flow patterns. For improving bottomhole pressure predictive accuracy, the flow correlations (with approximate physics) were combined with data-driven models (capturing the field specificity). Several hybrid modeling approaches were tested to provide a better pressure prediction, using physics to inform the machine learning. This not only enables robust pressure-based well performance analysis, but also offsets expensive pressure gauge data acquisition costs. To scale-up the bottomhole pressure calculation and efficiently process large well counts, a methodology was developed for systematically building proxy models, preserving predictive accuracy while achieving significant computational speedup. The flow correlations were sampled through experimental design and used to train various machine learning models, resulting in a proxy model that can be applied at field scale. For hybrid model development, four different approaches were tested using the physics-based correlations to inform the machine learning. The results were evaluated by predicting bottomhole pressure from surface data in a well with a permanent downhole pressure gauge, and two of the evaluated approaches showed a significant improvement in prediction accuracy compared to the traditional physics models. Several data-driven model classes were evaluated as proxy-model candidates, using two types of representations: a pressure point solution, and the whole pipe/wellbore pressure profile. A blind hold-out validation (comparing vertical lift performance curves) was performed to rank candidate model performance. Proxy models were successfully developed with large computational gains, while maintaining accuracy. The presented methodologies are practical ways to leverage existing models and field-specific data to improve accuracy and performance of bottomhole pressure estimation in unconventional reservoirs, by taking advantage of data-driven models, while still being rooted in physics-based principles. The proposed methods can be applied to large fields with thousands of wells, and situations where the physics-based models are not well calibrated or insufficient to adequately represent the flow patterns describing multiphase flow.

Abstract Forecasting the future performance of unconventional wells is essential for reserves estimation, production optimization and to support decision-making. In this paper, we present a physics-guided deep learning predictive model that forecasts production profiles given an input of completion, formation, and fluid properties. The predictive model accounts for prediction errors originating from potentially missing physical phenomena, simulation inputs, and incomplete reservoir description by correcting the production profiles obtained from a physical simulator. Using the multi-task learning concept, the predictive model creates a shared representation that can also be used to forecast production behavior, estimate reserves, and as a diagnostic tool to understand well performance under varying well properties. Compared to purely data-driven predictive models, it also allows for improved prediction for when well properties (as input) are outside the range of the original training data. The developed method is tested using field data from the Bakken Shale Play in North Dakota. A physics-based simulator is used to obtain simulated production responses using relevant completion, formation, and fluid properties from wells in Bakken. The deep learning architecture is jointly trained with the field data and simulated data. For any unseen input well properties (test data), the learned feature space (i.e., shared representation) provides a reliable diagnosis of potential reserves, oil rate decline trend, and other well performance metrics. The multi-task predictive model also provides robust production forecasts when compared to a purely physics-based/data-driven model or individual (single-task) predictive models. Introduction Hydrocarbon production from hydraulically fractured unconventional reservoirs involves complex flow and transport processes that are not well understood. Existing simulation methods are based on theories and mathematical models rooted in the flow and transport phenomenon of conventional reservoirs and may introduce substantial errors when directly applied for unconventional resources (Trangenstein et al., 1989; Aziz et al., 1979). Modeling flow behavior and physical processes that take place during production from extremely tight formations with complex fracture networks is an active research area. To accurately represent the physical relationship between well properties (i.e., formation, completion, and fluid properties) and production responses, a simulation model must be able to capture the wide range of uncertainty in model parameters, such as the interaction between natural and induced fractures, coupled with complexities in modeling geomechanical interaction that can adversely affect predictions obtained from a simulator.

Abstract Deep learning has become an integral part of image classification and segmentation, especially with the use of convolutional neural networks (CNN) and their variants. Although computationally expensive and time-consuming, there are several promising applications to classify or segment SEM images, images of core and thin sections. But we have not really questioned the need for really deep networks in these applications? Can shallower networks be competitive in relation to deeper networks? Can a shallower network with a wider diversity of convolutional filters (breadth) do better than a deeper network? What image resolution and filter complexity do we need to achieve a high degree of accurate classification? In this paper, we assess image classification using over 8000 SEM images acquired from 22 different unconventional plays to answer the questions posed above and provide guidelines to select an optimal depth and breadth for image classification. We evaluate several different CNN architectures systematically by changing the breadth (the number of filters within each layer) or the depth of the network (the number of layers) to relate classification accuracy and the complexity of the CNN. We also test the performance of the different CNN's against different image resolutions to determine if there is a specific field-of-view that is necessary to obtain satisfactory play classification. For all image resolutions considered, surprisingly, the simplest and shallowest one-convolutional layer model performs remarkably well with even 22 different classes (plays) to identify. Despite the simplicity of the network, we achieve over 80% accuracy in play identification (with correspondingly high recall and precision). A moderate increase in depth to 2 convolutional layers advances the accuracy to beyond 90%, even with a modest number of filters. Deeper networks that lack filter width perform poorly, indicating the significance of filter diversity in each of the convolutional layers of a CNN. The results from this study show that deeper networks are probably not necessary for image classification of SEM images/core or thin-section images. The microstructural features within the samples probably necessitate a wider diversity of filters. Finally, although several studies have relied on transfer learning of ‘published’ or open-source CNNs for play identification and image segmentation, this study shows that the level of complexity required is far less, making training more efficient and reducing the likelihood of overfitting.

Abstract In the past decade fiber-optic distributed sensing, including distributed acoustic sensing (DAS) and distributed temperature sensing (DTS), has been increasingly used in unconventional fields for intelligent completions, production monitoring, and optimization. Recent advancement in fiber-optic distributed sensing technology has made it possible to reveal critical operational information in situ and in real time. Despite the complexity in analyzing large volume of fiberoptic sensing data, the association of these measurement data with operation states remains mostly qualitative, correlative and after-fact descriptive. To address this challenge, we have developed deep learning based methods to directly predict operation states and variables from measured fiberoptic sensing data sets. This will enable an automated quantitative framework for intelligence completion and production monitoring, with minimal manual interpretation. Introduction Fiber optic DAS and DTS measures strain/strain rate and temperature along the fiber, respectively, as induced by vibrations or temperature fluctuation in the media where the fiber cable is coupled (Hartog 2017). Due to the distributed sensing capability and the recent technology advancement, DAS and DTS have enabled a large number of applications in oil/gas fields, for instance: downhole production and injection profiling (van der Horst et al. 2015, Carpenter 2016, Sanni et al. 2018), sand-ingress detection (Thiruvenkatanathan et al. 2016, Carpenter 2017), gas lift surveillance (Hemink et al. 2018), hydraulic fracture stimulation diagnosis (Molenaar et al. 2013), hydraulic stimulation and geomechanical deformation monitoring (Hull et al. 2019, Jayaram et al. 2019, Bohn et al. 2020), hydraulic fracture monitoring (Richter et al. 2019), early gas detection in wellbore ( Sharma et al. 2021), hydraulic frac-hit corridor monitoring (Wu et al. 2020), Plug-and-perf completion monitoring (Shen et al. 2017), and analyze cluster efficiency (Anifowoshe et al. 2016) and stage communication (Ghahfarokhi et al. 2018). Throughout these applications, time traces of DAS at each depth location are transformed into time-spectral domain, and magnitude of transformed data in certain frequency bands most correlated with operation variables of interest are identified and collated with operation states as the so-called frequency band extracted (FBE) signals. DAS measurements tend to accumulate significant amount of raw data. Even simple spectral analysis can be computationally costly and complex due to the large channel numbers, high sampling frequency and long-time duration over which the measurements are taken. Despite of that, in most of the examples above the approaches used to associate these processed DAS information with operation variables and states of interest are generally qualitative, correlative and after-fact descriptive, requiring significant amount of human intervention and interpretation. This can be cumbersome and even in-feasible for applications involving long duration monitoring. This compels the need for automated and quantitative processing frameworks capable of predicting operation states or variable values directly from fiber optic DAS and/or DTS data, with minimal manual intervention. When combined with efficient pre-processing of the raw measurement data, this will enable DAS/DTS based field monitoring to improve real-time operation decision making.

Abstract Core sedimentary structure identification and facies classification play a crucial role in reservoir characterization for oil & gas exploration and production. The task is traditionally done manually by geologists and can be inconsistent and time-consuming. In addition, it is challenging to standardize the approach due to human bias. Over the last two decades, Machine Learning (ML) has emerged as a new approach to solve real-world problems through the combination of computational technology and artificial intelligence (AI), and it has been increasingly adopted by the oil and gas industry, including the application for unconventional resource assessment. The new automated core image characterization and identification approach based on Transfer Learning (TL) and Convolutional Neural Network (CNN) has the potential to provide accurate real-time sedimentary structure identification and reservoir quality prediction. Following the breakthrough of unconventional resource development in North America, China has made a significant investment in its unconventional plays. This study investigated core sedimentary structure images of Bohai Bay Basin and Songliao Basin in China, which have been the key conventional oil & gas production bases for decades and seen increasing interest in their unconventional resource potential. The paper will discuss the findings and capabilities uniquely enabled by the new TL&CNN technology. Sedimentary structure automated recognition requires prior knowledge learned from the identified sedimentary structures. Three different sedimentary structures of 695 core images in seven wells were identified as wavy bedding, parallel bedding and carbonaceous foliation based on core description. The prototype image classification model, which has been successfully applied and reiterated in many fields such as commodity labeling, security monitoring, industrial quality inspection and medical diagnosis, was directly used to train the sedimentary structure identification model through TL method. The CNN method was then applied to optimize the model using the 695 annotated core images that have been edited by removing extraneous elements and backgrounds. Among them, 70% of the images were randomly selected to be used as training samples, while the other 30% for blind testing purpose. The ML investigation has 11 modeling and validation configurations, with the recognition accuracy ranging from 88.00% to 92.73% with an average of 90.01%. Subsequently, in order to further improve the model accuracy, the sample set was filtered to 604 core images by removing misidentified images. The accuracy of the new model of the ten versions was from 89.36% to 92.34% with an average of 91.11%. Although the accuracy and stability of the new model increased slightly, it created some uncertainty as some images that have been correctly identified before were actually misidentified this time. In the future, we plan to increase the number and coverage of core image samples in the ML study to evaluate the impact on the accuracy and consistency of the TL&CNN sedimentary structure identification model, and expand the approach to identify other sedimentary structure types. As this TL&CNN sedimentary structure identification model becomes mature, it can be applied to the unconventional plays.

Abstract Hydrocarbon production from shale reservoirs is inherently inefficient and challenging since these are low permeability plays. In addition, there is a limited understanding of the fundamentals and the controlling mechanisms, further complicating how to optimize these plays. Herein, we summarize our past and current efforts to reveal the shale fundamentals and devise development strategies to enhance extraction efficiency with a minimal environmental footprint. Integrating these fundamentals with machine learning, we outline a pathway to improve the predictive power of our models, enhancing the forecast quality of production, thereby improving the economics of operations in unconventional reservoirs. For instance, we have developed science-informed workflows and platforms for optimizing pressure-drawdown at a site, which allow operators to make reservoir-management decisions that optimize recovery in consideration of future production. Recently, our work relies on the hybridization of physics-based prediction and machine learning, whereby accurate synthetic data (combined with available site data) can enable the application of machine learning methods for rapid forecasting and optimization. Consequently, the workflow and platform are readily extendable to operations at other sites, plays, and basins. Introduction Shale reservoirs have redefined the energy landscape of the world (Melikoglu, 2014). Shale reservoirs currently contribute to 70% of U.S. natural gas production and 60% of U.S. oil production (Perrin & Geary, 2019). However, these reservoirs possess unique characteristics that entail custom-designed development plans (Mehana & El-monier, 2016). Currently, we recover less than 10% of the hydrocarbons in place from these low permeability plays with the most efficient development plan (Alharthy et al., 2015). This low efficiency is also due to the limited understanding of shale characteristics' effects on the fluid properties and transport through porous media. Therefore, a fundamental understanding and quantification of these effects are required for devising better development plans to improve the current recovery factors (Middleton et al., 2017).

Many analytical methods rely on complete datasets. However, data obtained from the field are often incomplete or inaccurate. An example of this would be well production data, where oil, water and gas rates are provided, but sometimes water or gas rates are missing. With current technology, generative machine learning methods, such as those used for Deep Fake, can generate images and data that are all but indistinguishable from reality. Using an adapted generative method known as the Generative Adversarial Imputation Network (GAIN), this paper evaluates these methods and capabilities for filling in missing data. Validation of these methods was done by creating missing data within complete data sets, comparing the generated values to that of the original. This work found that with initially large datasets and upwards of 45% of missing values, data can be "filled in" with surprising accuracy. Quantification of the GAIN process and association of missing data and variable importance was done with the use of probability distributions. The relationship between the amount of missing data and the accuracy and probability associated with the predictions has been further quantified and presented within the context of various types of datasets. This paper discusses how generative methods of machine learning were utilized to fill in the missing portion of existing data with great success. Using a GAIN model, the missing fields can be generated for use in statistical analysis, decision making, and the optimization of current and future projects. Introduction Many studies have been done using machine learning in order to find better and better ways to predict the unknowns. This is true for all fields, and when the data acquired does not always meet the criteria of the researcher, decisions on how to go forward tends to lead to some data just outright getting discarded. Missing data is a profound issue affecting a number of fields, from agricultural to medical and, of course, particularly data in the oil industry. These missing variables can be problematic when a researcher desires to use all the data they can when doing any sort of data analytics. Suddenly, the researcher is required to make decisions on how to, or whether to, make up for the missing information. That person must make a choice, to take an average of the other available data, come up with a control value to make up for any missing data in a specific variable, or determine if it is the best practice to simply drop that sample of data. For example, if there is a row of data (sample) with 30 data values, and three values are missing, a choice must be made on whether to disregard the entire row of data, or to find suitable values that can be used to replace the missing values.

Multiphase flow through fractures is common in many fields, yet our understanding of the process remains limited. In general, this is because some factors which separate multiphase flow from single-phase flow (interfacial tension, wettability, residual saturation) are difficult to characterize and control in a laboratory setting, and are also challenging to implement in traditional numerical simulators. Here, we present a series of lattice Boltzmann simulations of CO 2 displacing brine in rough fractures with heterogeneous wettability. This extended abstract focuses on the application of this technique to predict irreducible brine saturation within the fractures. We show that this irreducible brine saturation may be greater than 25%, which could have significant impacts on production estimates from unconventional reservoirs and is typically not accounted for in reservoir simulators. However, performing these simulations at the field scale is not possible due to their computational expense. Therefore, we present a machine learning technique based on deep neural networks to predict the fluid distribution within these fractures at steady state trained upon on the lattice Boltzmann simulations. To our knowledge, this is the first example of machine learning being used to predict the distribution of fluid within a subsurface media. Here we show that a trained network is able to accurately predict the fluid residual saturation and distribution based solely on the dry fracture characteristics. This proves that machine learning holds promise for upscaling these simulations to a relevant scale for application to the oil and gas industry. Introduction Multiphase flow in fractures has implications to many fields including nuclear waste disposal, CO 2 sequestration, geothermal energy, and the oil and gas industry. During multiphase flow, factors that do not play a role in single-phase flow become important. These factors include the interfacial tension between fluids, the fluids viscosity ratio, and the wettability between the solid surfaces and each fluid. Compared to porous media, where the effect of wettability has been extensively researched, the influence of wettability during fracture flow is relatively unstudied. This is partly due to the difficulty in characterizing the wettability of natural rock cores and conducting experiments as well as the difficulty in including wettability into numerical simulations. Therefore, the importance, or lack thereof, of wettability during multiphase fracture flow remains uncertain.

Predicting EUR in unconventional tight-shale reservoirs with prolonged transient behavior is a challenging task. Most methods used in predicting such long-term behavior have shown certain limitations. However, long short-term memory (LSTM) – an artificial recurrent neural network (RNN) architecture used in deep learning – has proven to be well-suited to classifying, processing, and making predictions based on time series data with lags of unknown duration between important events. This study compares LSTM and reservoir simulation forecasts. Available unconventional tight-shale reservoir data is analyzed by LSTM and predictions obtained. A reservoir simulation model based on the same data is used to compare the LSTM forecast with results from a physics-based model. In the LSTM forecasting, any operational interferences to the well are taken into account to make sure the machine learning model is not impacted by interferences that do not reflect the actual physics of the production mechanism on the behavior of the well. The forecasts from the LSTM machine learning model and the physics-based reservoir simulation model are compared. The LSTM model shows a good level of accuracy in predicting long-term unconventional tight-shale reservoir behavior using the physics-based reservoir simulation model as a benchmark. An analysis of the comparison shows that the LSTM machine learning model provides robust predictions with its long-term forecasting capability. This allows for better data-driven forecasting of EUR in unconventional tight-shale reservoirs. A detailed analysis is done using the forecast results from LSTM and the reservoir simulation model, and the key drivers of the EUR response are evaluated and outlined. Deep learning applications are limited in the oil and gas industry. However, it has key advantages over other conventional machine learning methods; especially where relationships are in time and space and not very clear to the modeler. This study provides a detailed insight into deep learning applications in the oil and gas industry by using LSTM for long-term behavior prediction in unconventional shale reservoirs.

In the petroleum industry, accurately predicting production potential of undrilled unconventional horizontal wells requires the use of highly complex models and is a constant research in progress. Deterministic approaches such as decline curve analysis (DCA) are used in most cases to estimate production potential due to ease of implementation. The main disadvantage of using DCA by itself is that it requires an existing well to forecast the production, which is very expensive to drill. This paper shows the procedures for building a flexible machine learning based decline curve-spatial method that can be easily used to predict the estimated ultimate recovery (EUR) of newly proposed wells without the requirement of costly data or other time consuming methods. A type of artificial neural network (ANN) called feed forward neural network (FFNN) was used as the machine learning method during this process. In order to achieve this goal, production and well data were collected from public domain sources. Power law exponential (PLE) DCA method was implemented on a portion of the existing wells with sufficient production history. The data then was divided into training and test sets. The training set was fed into the ANN model and the results were compared with the results obtained from other inherently spatial methods such as universal kriging, geographically weighted regression (GWR), and generalized additive model (GAM). Finally, the EUR of the new wells were compared to the original training and test data to observe any discrepancies in the prediction and necessary adjustments to the model hyperparameters were made when there were discrepancies. Analyzing and observing all the results from the various combinations of methods indicates that using ANN without any spatial correlation is a less reliable method to estimate the production potential of new wells in the Marcellus shale gas reservoir when compared to other inherently spatial methods. This study shows that it is necessary to include spatial correlations between wells in the EUR prediction of new wells in the ANN models. The predictive strength of other spatial methods were within a reasonable range. One of the main input parameters that may increase the accuracy of the model are completion parameters. However, most of these completion parameters that will affect the production are very difficult and expensive to obtain. These parameters which could not be included as inputs can increase the accuracy of these models when available. This experimental study indicates that there are many other possible variations of this method other than the methods discussed in this paper. It also shows that the spatial correlation of wells is highly important when predicting the new well production in unconventional reservoirs. This is a flexible method, which can be easily modified to find better predictive methods for other areas of interest.

During hydraulic fracturing treatments, screen-out has caused delay in placement of subsequent stages, stand-by of on-site teams, loss production and safety concerns. Current prediction methods rely on either post-minifrac diagnostic or pure physics-based approaches, which were not always appropriate for real-time applications. This study uses data-driven methods like deep neural networks along with physics-based approach (i.e., Inverse Slope Method) to provide advanced warning of screen-outs in real-time. With domain knowledge, 65 screen-out stages in the Niobrara-DJ Basin were labeled into output binary indicators with or without screen-out. Input datasets include surface treatment pressure, slurry rate, surface proppant concentration and 11 engineering features in high frequency 1-second time interval. 65 stages were first split into 53, 8, and 5 stages for training, validation and test, respectively. Then, modified inverse slope method was utilized to predict screen-out, and three parameters were optimized such as slope duration, slope change, minimum slope value. Then, a baseline model and another advanced deep neural network-based model (CNN-LSTM) were trained, and CNN-LSTM model architecture was optimized by tuning hyperparameters. The final ensemble model was created as weighted average of the modified inverse slope and CNN-LSTM models. The ensemble model showed high precision, recall, f1-score and Area Under the Curve (AUC) of ROC curves for the validation and test datasets, and confusion matrices indicated better performance than either model alone. The implemented workflow is applicable to detect anomalies in real-time for high-frequency time series. In addition, it can handle automatic feature engineering, long-term time dependency, and leverage synergic effect of physics and machine-learning approaches. Introduction During hydraulic fracturing, large amount of fracturing fluids is pumped downhole to create fractures which are filled with proppant particles. However, when the proppants carried in the fluid create a bridge across the perforations or similar restricted flow area, it creates a sudden and significant restriction to fluid flow that causes a rapid rise in pump pressure, i.e., screen-out (Schlumberger 2020). Screen-out causes stand-by status, delay in the placement of subsequent stages, wellbore cleanout operations, and lost production days. Therefore, advanced warning of screen-out is critical to improve operation safety and efficiency.

Machine and deep learning (ML and DL, respectively) have currently, and will remain so in the future, proven to be a successful approach to get an exponential insight for Oil & Gas seismic-based interpretation and production in conventional and shale developments. But, when their impact on regional exploration and reservoir characterization, as well as pre-drilling hazard analysis are pondered, a relatively lesser number of published cases are available for both conventional and unconventional shale reservoirs. This number can be minor if 3-D seismic data (and Seismic Quantitative Analysis, or QA) is integrated into physical-based deep learning to play one of its fundamental roles in Oil & Gas reservoir characterization during decades: regional mapping and forecasting, in a 3-D fashion, of rock properties and fluid distributions far from control wells. The present paper shows that there are just those intrinsic attributes of the 3-D seismic data for regional mapping and forecasting, even if it represents an indirect image of the subsurface and can be affected by quality and resolution issues in some cases, that stands seismic data as the most likely regional model to guide physical-based deep learning and get deeper insight of the subsurface. Locally, those seismic-based DL models can be also very valuable on supporting better making decisions to optimize field developments when are integrated with other source data (i.e., microseismic and borehole DAS data) by statistical methods or as augmented training and validation samples for deep neural classification and prediction. Some examples from Teapot Dome and Niobrara Shale are included to support the use of well-known QA techniques to improve supervised deep learning performance using 3-D seismic data. The core of the present approach is to include QA as part of the exploratory data analysis (EDA) and data-driven seismic features as training and validation samples for supervised classification of test data labeled by fluid type (i.e., brine, oil and gas) and prediction of fracture density (proxy) distributions by using LSTM deep recurrent and neural networks, respectively. The analysis of results shows that a significant validation accuracy using LSTM deep recurrent network for fluid classification and deep neural network for fractured density (proxy) prediction can be obtained if well log and seismic data are integrated through QA in the EDA analysis for physical-based DL classification and prediction. Particularly, in cases where the training set is very limited our approach let seismic data to play an important role on augmenting training and validation samples and improving the supervised neural performance.

Determining the closure pressure is crucial for optimal hydraulic fracturing design and successful execution of fracturing treatment. Historically, the use of diagnostic tests before the main fracturing treatment has significantly advanced to provide more information about the pattern of fracture propagation and fluid performance to optimize the designs. Many analytical methods such as the G-function, square root of time, etc. have been developed to determine the fracture closure pressure. In some cases, the difficulty in determining the fracture closure pressure, as well as personal bias and field experiences, make it challenging to interpret the changes in the pressure derivative slope and identify fracture closure. These conditions include: High permeability reservoirs where fracture closure occurs very fast due to the quick fluid leak-off. Extremely low permeability reservoirs, which require a long shut-in time for the fluid to leak-off and determine the fracture closure pressure. The non-ideal fluid leak-off behavior under complex conditions. The objective of the paper is to apply machine learning methods to implement predesigned algorithms to predict the fracture closure pressure and other minifrac parameters while minimizing the shortcomings in determining the closure pressure for non-ideal or subjective conditions. This study demonstrates training different supervised machine learning algorithms to help predict fracture closure pressure, fracture closure time, permeability and the time to reach late flow regimes. The workflow involves using the datasets to train and optimize the models, which are subsequently used to predict the minifrac parameters of testing data. The output results are then compared with actual results from more than 120 Diagnostic Fracture Injection Test (DFIT) data points. It is further proposed that an integrated approach helps feature selection, dataset processing and studying the effects of data processing on the success of the model prediction. The results from this study reduce the subjectivity and need for experience in interpreting DFIT data. We speculate that a linear regression and multilayer perceptron (MLP) neural network algorithms can yield high scores in the prediction of the minifrac parameters.

Decline Curve Analysis (DCA) has been widely applied in production forecasting of wells in unconventional hydrocarbon reservoirs. However, traditional curve-fit-based methods fall short of forecast accuracy due to three weaknesses: first, they cannot capture the reservoir signals not modeled by the underlying DCA model formulas; second, when predicting the production of a target well, the production history of other wells in the geologic formation (which is valuable information) are not considered; third, the wells' geographic, geologic, wellbore, well spacing, and completion properties, which are highly relevant to production capability, are not utilized. More recent approaches have begun replacing traditional DCA with machine-learning methods (e.g., Random Forest, Support Vector Regression, etc.) for production forecast. Nevertheless, these methods are still sub-optimal in detecting similar production trends in different wells, leading to large forecast error. A simple and novel method called Dynamic Production Rescaling (DPR) is developed to improve the accuracy of machine-learning DCA (ML-DCA). By combining DPR with common ML-DCA methods, we observe that the error mean, deviation, and skewness can be significantly reduced by 15% to 35% compared with ML-DCA without DPR. The error reduction is 30% to 60% compared with automatic curve fit of traditional Modified Arps DCA model. DPR has been tested successfully on monthly production data of over 20,000 unconventional horizontal wells in the Permian and Appalachian basins for both long- and short-term forecasts. The significant error reduction is consistent across different basins and formations. DPR is computationally efficient, so large number of wells can be analyzed automatically and quickly. Moreover, the effectiveness and efficiency of DPR is independent of the underlying machine-learning algorithm, further demonstrating its robustness. Introduction Operation and management of unconventional hydrocarbon reservoirs requires production forecast of each well. This enables better development planning, economic outlook, reserve estimates, and business decisions such as trading and pricing strategies. A methodology called Decline Curve Analysis (DCA) has been widely applied in production forecast of wells in unconventional reservoirs, and many analytical DCA models are available for use. They describe the production physics in analytical equations of flow rate versus time, and the coefficients of the equations are computed from curve-fitting the production history.

The focused-ion-beam scanning electron microscopy (FIB-SEM) technology allows imaging of nano-porous tight reservoir rock samples in 3D at a resolution up to 3 nm/voxel. However, the quality and efficient segmentation of FIB-SEM images is still a complicated and challenging task. Correct porosity determination from FIB-SEM images requires fast and robust segmentation. Typically, a trained operator spends days/week for subjective and semi-manual labeling of a single FIB-SEM dataset. The presence of FIB-SEM artifacts, such as pore backs, requires the development of a new methodology for efficient image segmentation. We developed a robust approach for an automated highly-efficient multimodal segmentation of FIB-SEM datasets using machine-learning (ML) based methods. A representative collection of rocks samples was formed based on the petrophysical interpretation of well logs for a complex tight gas reservoir rock of the Berezov formation (West Siberia, Russia). The core samples passed through a multiscale imaging workflow for pore space structure upscaling from nanometer to log scale. FIB-SEM imaging resolved the finest scale using FEI Versa 3D DualBeam analytical system. Image segmentation utilized an architecture based on a convolutional neural network (CNN) in the DeepUnet configuration. The implementation utilized the Pytorch framework in a Linux environment. Computation exploited high-performance computing system based on Intel(R) Core (TM) i7-6700 CPUs and NVIDIA GTX 1080i TitanBlack GPUs. The target data included three 3D FIB-SEM datasets with a physical size of around 20 × 15 × 25 μm with a voxel size of 5 nm. A professional geologist manually segmented (labeled) a fraction of slices. We split the labeled slices into training (TD) and validation data (VD). We then augmented training data to increase TD size. The developed CNN delivered promising results. The model performed automatic segmentation with the following average quality indicators based on VD: accuracy of 96.66%, the precision of 91.67%, recall of 67.57%, and F1 score of 77.00%. We achieved a significant boost in segmentation speed of 14.5 Mpx/min. compared to 0.18–1.45 Mpx/min. for manual labeling, yielding at least 10 times of efficiency increase. The presented research work improves the quality of quantitative petrophysical characterization of complex reservoir rocks using digital rock imaging. The development allows the multiphase segmentation of 3D FIB-SEM data complicated with artifacts. It delivers correct and precise pore space segmentation resulting in insignificant turn-around time saving, as well as increased quality of porosity data. Although image segmentation using CNNs is a mainstream in the modern ML world, it is an emerging novel approach for reservoir characterization tasks.

Production analysis for low permeability reservoirs is crucial in characterizing flow regimes and reservoir properties. However, outliers, aberrant segments, and systematic noise often disturb and can cause significant bias. Additionally, incomplete sampling during system offline or failure can truncate useful information and reduce our capability to interpret reservoir characteristics. Uniform or non-uniform sampling with different sampling frequencies and increasing bytes in the dataset can also complicate the analysis. In this paper, we present a Denoising Long Short-Term Memory (DeLSTM) approach for noise reduction in production rate surveillance, and additionally provide a physics-based, deep-learning workflow based on a traditional polynomial window-based least-square fitting, Savitzky-Golay (SG) filter, coupling with a Cascaded Long Short-Term Memory (CLSTM) Network, to facilitate surveillance of pressures and production rates. Recorded pressures and rates are time-series data pairs with spatial dependency. Two production surveillance scenarios are considered. One is to simply monitor the production rates and perform traditional decline curve analysis (DCA) by assuming constant bottom hole pressure with boundary dominated flow (BDF) being reached. The other is to incorporate pressure data, and the ensemble analyses could further improve the surveillance quality and production data analysis. We first present the DeLSTM algorithm solely to denoise the noisy rate data with existing outliers and aberrant segments and further reconstruct the missing rates in the first scenario. Then, to better preserve the characteristics of the original datasets with the presence of pressure information, we used SG-CLSTM with the superposition physics-based training features for rate reconstruction with high sampling frequency and incomplete sampling events occurred in transient flow regimes. The proposed workflow was first applied to a synthetic case with random outliers, aberrant segments, and artificial white Gaussian noise to approximate the measured signals collected in the field. The missing segment was obtained by hiding a period of production data. The formed incomplete samples were then input into DeLSTM and SG-CLSTM for rate, pressure-rate reconstruction respectively. Then, the incomplete histories were generated and back-tested with the hidden dataset.

Abstract Machine learning and deep learning, a subset of machine learning, have become successful predictive modeling tools. Machine learning applications include classifying DNA sequences and fraud detection. Deep learning applications include language processing and facial recognition. Both methods have shown to be useful in understanding unconventional reservoirs when traditional techniques used to understand conventional reservoirs do not readily transfer. In this work, machine learning and deep learning are used to predict geological facie classifications using computed tomography (CT). A combined approach is developed to predicted geological facies. The first aspect of this approach utilizes molecular weight, density, and porosity from CT scans on core samples to learn from existing user-defined geological facie classifications. Once the machine understands the relationship between geological facies and CT scan physics, the resulting model is used to predict subsequent geological facies using the CT scan parameters. The second aspect of this approach utilizes CT scan thin section images of core samples to learn from corresponding geological facie classifications. Once a deep learning model learns from these CT scan images, the resulting model is used to predict subsequent geological facies using CT scan images. Results show misclassification rates at approximately less than 30% which are more favorable to error rates associated with a single person classifying facies. In addition, unsupervised machine learning results show robust clustering of geological facies. Machine learning and deep learning show the value of utilizing CT scan parameters and images in unconventional oil and gas reservoir workflows. While typically used in other areas such as cancer detection, CT scans in conjunction with machine learning and deep learning can drastically reduce the error rate and time it takes to fully categorize an unconventional play. Current methodologies involve having a team of petrophysicists who characterize core samples at every specified CT scan resolution depth. These workflows can be cumbersome and time-consuming for a team of petrophysicists and at times result in facies classification governed by human bias. The proposed methodology in this work drastically reduces human error and the time to understand geological facies. As a result, business decisions can be made quicker. These models are expected to provide substantially more accurate and consistent geological facie predictions.