Long horizontal wellbore sections are now a key requirement of oil and gas drilling, particularly for tight reservoirs. However, such sections pose a unique set of borehole-cleaning challenges which are quite distinct from those associated with less inclined wellbores. Experimental studies provide essential insight into the downhole variables that influence borehole cleaning in horizontal sections, typically expressing their results in multivariate empirical relationships with dimensionless cuttings bed thickness/concentration (H%). This study demonstrates how complementary empirical H% relationships focused on pairs of influential variables can be obtained from published experimental data using interpolated trends and optimizers. It also applies five machine learning algorithms to a compiled multivariate (10-variable) interpolated dataset to illustrate how reliable H% predictions can be derived based on such information. Seven optimizer-derived empirical relationships are derived using pairs of influential variables which are capable of predicting H% with root mean squared errors of less than 1.8%. The extreme gradient boosting model provides the lowest H% prediction errors from the 10-variable dataset. The results suggest that in drilling situations where sufficient, locally-specific, information for multiple influential variables is available, machine learning methods are likely to be more effective and reliable at predicting H% than empirical relationships. On the other hand, in drilling conditions where information is only available for a limited number of influential variables, empirical relationships involving pairs of influential variables can provide valuable information to assist with drilling decisions.

The manual picking of reservoir formation boundaries using limited available well-log data in multiple wells across gas and oil reservoirs tends to be subjective and unreliable. The reasons for this are typically caused by the combined effects of spatial boundary complexity and limited well-log data availability. Formation boundary characterization and classification can be improved when treated as a binary classification task based on two or three recorded well logs assisted by their calculated derivative and volatility attributes assessed by machine learning. Two example wellbores penetrating a complex reservoir boundary,one with gamma-ray,compressional-sonic,and bulk-density logs recorded,the other with just gamma-ray and bulk-density logs recorded,are used to illustrate a more rigorous proposed methodology. By combining attribute calculation,optimized feature selection,multi-k-fold cross validation,confusion matrices,feature-influence analysis,and machine learning models it is possible to improve the classification of the formation boundary. With just gamma-ray and bulk-density recorded well logs plus selected attributes. K-nearest neighbour,support vector classification,and extreme gradient boosting machine learning models are able to achieve high binary classification accuracy: greater than 0.97 for training/validation in one well; and greater than 0.94 for testing in another well. extreme gradient boosting feature-influence analysis reveals the attributes that are the most important in the formation boundary predictions but these are likely to vary from reservoir to reservoir. The results of the study suggest that well-log attribute analysis,combined with machine learning has the potential to provide a more systematic formation boundary definition than relying only on a few recorded well-log curves.

A technique is proposed that calculates derivative and volatility attributes from just a few well log curves to assist in brittleness index predictions from sparse well-log datasets with machine learning methods. Six well-log attributes are calculated for selected recorded well logs: the first derivative, the moving average of the first derivative, the second derivative, the logarithm of the instantaneous volatility, the standard deviation of volatility, and the moving average of volatility. These attributes make it possible to extrapolate brittleness index calibrations from the few cored and comprehensively logged wells to surrounding wells in which only minimal well-log suites are recorded. Data from two cored wells penetrating the lower Barnett Shale with distinct lithology and five well logs recorded are used to demonstrate the technique. Based on multi-K-fold cross validation analysis, the data matching K-nearest neighbour machine learning model provides the most accurate brittleness index predictions, closely followed by tree-ensemble models. For this dataset, recorded data from three well logs plus calculated attributes matches the brittleness index prediction accuracy that is achieved by the five recorded logs. Moreover, any one of the logs plus their calculated attributes yields better brittleness index prediction performance than that achieved by a combination of just those three recorded well logs. Analysis of the Gini indices of the tree-ensemble models reveals the relative influences of the recorded logs and their attributes on the brittleness index prediction solutions. Such information is used to perform feature selection to optimize the well-log attributes involved to generate reliable brittleness index predictions.

Unconventional reservoir resources are important to supplement energy consumption and maintain the balance of supply and demand in the oil and gas market. However, due to the complex geological conditions, it is a significant challenge to develop unconventional reservoirs efficiently and economically. At present, unconventional reservoirs are extensively studied, covering a wide range of areas, with special attention to the multiscale characterization of pore structures and fracture networks, description of complex fluid transport mechanisms, mathematical modeling of flow properties, and coupled analysis with multiphysics fields. This work briefly describes the multiscale and multiphysics influences on fluids in unconventional reservoirs, and the modeling and simulation work conducted to analyze them, with the aim to provide some theoretical basis for enhanced recovery from these geo-energy resources. The present article also aims to enhance the community's knowledge of other potential utilizations associated with some unconventional reservoirs, specially related to environmentally-driven projects, including permanent greenhouse gas storage and cyclic underground energy storage.

Rate of change, second derivative and volatility of gamma-ray (GR) well-log curves provide useful indicators with which to characterize lithofacies in clastic sedimentary sequences. Rolling averages of these variables, as they change with depth, are also able to distinguish certain lithofacies features. These attributes make it possible to accurately distinguish key facies by using only gamma-ray data, both with formulaic calculations and employing machine-learning (ML) algorithms. This is useful in the many wellbores for which only basic logging suites are available. As well as enhancing lithofacies classification more generally using well-log variables, these GR attributes can be used to forecast facies in real time based on logging-while-drilling data. The application is demonstrated with simple formula using synthetic GR logs featuring common clastic lithofacies and their transitions. Seven widely used ML methods are each trained and validated with a synthetic GR curve (1450 data points) displaying six distinct facies. The ability of the ML model to distinguish those facies using seven GR attributes is compared and further tested with an independent GR data set (800 data points). The random forest algorithm outperforms the other ML models in this facies prediction task, achieving a mean absolute error of 0.25 (on a facies class range of 1 to 6) for the independent testing dataset. The results highlight the benefit of this technique in providing reliable facies analysis based only on GR data. Random forest, support vector classification and eXtreme Gradient Boost are the ML models that provide the most reliable facies classification from the GR attributes defined. Annotated confusion matrices assist in revealing the details of facies class prediction accuracy and precision achieved by the ML and models and classification formulas.

Controlling reservoir fluid flow is important for maximizing petroleum production through wellbores. A major challenge that reduces the production of oil is early breakthrough of secondary fluids to the wellbore perforations. This occurs due to the low viscosity of gas and water relative to oil, and the heterogeneity of reservoir permeability. Autonomous inflow control devices represent a new self-regulating technology that helps to increase petroleum production, particularly oil, by restricting the production of unwanted fluids like gas and water into the wellbores. This study develops smart systems based on machine learning models to predict the performance of autonomous inflow control devices. Several machine learning models are evaluated including adaptive neuro fuzzy inference system, hybrid adaptive neuro-fuzzy inference system genetic algorithm, artificial neural network and support vector machine and their prediction performance is compared to that of linear regression, full quadratic regression model and the mathematical autonomous inflow control device performance model. Each model is developed to estimate the differential pressure of Equiflow autonomous inflow control devices based on ninety experimentally recorded data records. The range of equiflow autonomous inflow control device, viscosity, density and flow rate are the input variables and differential pressure is the output dependent variable of each model. The prediction accuracy of the models is assessed in terms of several standard statistical accuracy performance measures. These performance indicators confirm that the machine-learning models provide superior prediction accuracy for autonomous inflow control device differential pressure. Overall, the support vector machine achieves the most accurate predictions of all the models evaluated recording root mean square error of 0.14 Mpa and coefficient of determination of 0.98. On the other hand, the linear regression model records the lowest prediction performance, highlighting the non-linearity of the autonomous inflow control device processes.

Surface roughness of shales has a key influence on the petroleum resources they are able to store and the fraction of them that can be recovered. The fractal dimension quantifies the degree of roughness and is influenced primarily by the pore surfaces within the shale that typically include micro-, meso- and macro-pores. Isotherms generated by gas adsorption experiments are the common data source used to derive estimates of fractal dimension. The Frenkel-Halsey-Hill fractal technique is the most widely applied fractal dimension estimation method. Other methods can derive fractal dimension from isotherm data but typically the values they generate are different from the Frenkel-Halsey-Hill derived fractal dimension values. Moreover, those differences can vary significantly depending on the type of shales involved. Those shales displaying more complex pore-scale distributions including extensive micro-porosity components tend to be associated with the greatest discrepancies. A comparison of three fractal dimension calculation methods applied to shales reveals aspects of their calculation and interpretation methods that explain the differences in the fractal dimension values they generate. This study identifies the uncertainties that should be taken into account when applying the methods and the appropriate curve fitting optimization configurations that should be evaluated. Taking these factors into account leads to more realistic selections of appropriate fractal dimension values from gas adsorption isotherms of organic-rich shales.

In-situ stresses are highly important for wellbore stability studies during drilling, completion and production. Different methods are available to estimate the horizontal stresses especially maximum horizontal stress. Typically, Circumferential Borehole Image Logs can be run to determine the direction and width of breakouts and then stresses at different depths based on the equation developed by Barton et al. (1988). This research focuses on image logs from Harvey-1 well located in the Southern Perth basin to compare the maximum horizontal stresses obtained by various methods. The magnitudes of stresses from the breakout width approach (Barton's method) exhibit a considerable offset in comparison with elastic methods. Further investigations show that the likely reason for the offset relates to the fundamental assumption of the breakout width approach in which shear failures are considered to be constrained to horizontal planes. Failures within the wellbore are not necessarily horizontal and can be developed in different non-planar trajectories with various angles to the horizontal plane. Furthermore, the possible in-situ stresses from regional studies are constrained by means of stress polygons against which the reliability of results from breakout methods can be checked. Results indicate that due diligence and special care must be exercised for determination of maximum stresses from breakouts and more reliable methods are required than those currently used.

In Algeria, wells drilled in the Illizi Basin suggest the presence of a significant areal trend of Devonian coal seams with the thickest coal seams penetrated in the Lower Devonian stratigraphic unit F6. This makes them some of the oldest thick coal seams encountered. These coals exist between approximately 1500 and 4000 meters below surface. In particular, numerous coals in these formations drilled in the Oudoume field have recorded gas shows while drilling. A study of basic well log data from five wells penetrating Illizi Basin coals is conducted to characterize their distribution and provisionally evaluate their gas-bearing potential using petrophysical analysis coupled with machine learning. A simple multi-layer perceptron model (one hidden layer with four nodes) is used in a novel way to replicate estimates of gas saturation in the coal samples calculated approximately with the modified Kim equation. It does so by considering three commonly measured well-log variables: gamma ray, sonic travel time, deep resistivity (307 data records from the five wells studied). The log-calculated approximations (modified Kim equation) can be matched to better than plus or minus 1 scf/ton by the multi-layer perceptron model. The results and analysis presented provide preliminary encouragement that suggests the presence of a potentially extensive gas-bearing Devonian coal trend in the Illizi Basin that is worthy of further exploration. Future work is required to integrate data from additional wells and laboratory analysis of core samples to verify the extent of that coal trend and to quantify its gas concentrations.

Directional drilling is an excellent option to extend the limited reservoir reach and contact offered by vertical wells. Pressure transient responses (PTR) of horizontal wells provide key information about the reservoirs drilled. In this study multilayer perceptron (MLP) neural networks are used to correctly identify reservoir models from pressure derivative curves derived from horizontal wells. To this end, 2560 pressure derivative curves for six distinct reservoir models are generated and used to design a machine-learning classifier. A single hidden layer MLP network with 5 neurons, trained with a scaled conjugate gradient algorithm, is selected as the best classifier. This smart classifier provides total classification accuracy of 98.3%, mean square error of 0.00725, and coefficient of determination of 0.97332 over the whole dataset. Performance accuracy of the proposed classifier is verified with real field data, synthetically generated noisy PTR, and some signals outside the range initially assessed by the training plus testing data subsets. The developed network can correctly identify the reservoir-flow model with a probability of close to 0.9. The novelty of this work is that it employs a large dataset of horizontal (not vertical) well tests applied to six reservoir-flow models and includes noisy data to train and verify a neural network model to reliably achieve a high-level of prediction accuracy.