Search

Rate of penetration (ROP) is a key metric used to monitor the success of drilling a well. It is directly affected by drilling vibrations since excessive vibrations result in a reduction of ROP. Vibration modeling and monitoring is a complex process often requiring many simplifying assumptions that may not always generalize to different BHAs, reservoirs, geology and formations. Therefore, it would be desirable to minimize drill string vibrations using data driven models using readily available drilling data. The hypothesis tested is the classification of stick slip severity due to drilling vibrations using open source machine learning algorithms. The stick slip index (SSI) – measuring the severity of stick slip due to drilling vibrations – is classified as low or high using machine learning classification algorithms such as logistic regression, support vector machines, random forests, gaussian mixture models and discriminant analysis. Each algorithm was evaluated based on classification accuracy, F-1 score and area under the receiver operating characteristic curve (AUC). The random forest algorithm outperforms other algorithms with an average accuracy of 90% (F-1 score of 0.91 and AUC score of 0.89). The classification model can then be used within a ROP optimization model (or framework) to determine optimal operation parameters which do not result in stick-slip conditions while drilling addressing a serious limitation of previously published ROP optimization papers. • Machine learning is used for classification of torsional drilling vibrations. • Stick-slip index has been classified using algorithms such as logistic regression, SVM, random forests, and GMMs. • Issues such as class imbalance addressed to improve accuracy of the classification model. • Strategies to incorporate a vibrations model into an ROP optimization framework have been discussed. • Results show that classification algorithms can be used to accurately model drilling vibrations. [ABSTRACT FROM AUTHOR]

Abstract The identification of lithology from well log data is an important task in petroleum exploration and development. However, due to the complexity of the sedimentary environment and reservoir heterogeneity, the traditional lithology identification methods can not meet the needs of real-time and accurate prediction and identification with logging while drilling (LWD) equipment. The basic data of this paper are derived from conventional wireline logging (CWL) data and the LWD data in Yan'an Gas Field. The main research goal is to compare and analyse three popular machine learning algorithms, which are one-versus-rest support vector machines (OVR SVMs), one-versus-one support vector machines (OVO SVMs) and random forest (RF), and to optimize a more practical method in the field for LWD systems. To reduce the dimensions of the input data, the characteristic parameters of the training data are obtained by a correlation analysis of the logging data. The optimal parameter values of each algorithm are determined by grid search method and 10-fold cross-validation method. On this basis, the lithology predictions of the actual LWD data are carried out by using three classifiers. Considering the time consumption of the model training and the lithology identification accuracy of the model, the best lithology identification model while drilling is selected. The results show that the characteristic parameters of the training data after the correlation analysis are AC, CAL, GR, K, RD and SP logs. The overall classification and recognition performance of the RF classifier is better than that of the other two classifiers, and its accuracy is even greater than 90%. The evaluation matrix shows that the OVR SVMs and RF classifiers yield lower prediction errors than the OVO SVMs classifier in each single lithology identification, but the RF classifier spends much less time in the training process. Based on the comprehensive comparative analysis, it is considered that the RF classifier has the characteristics of a short training time and high recognition accuracy in practical production applications, so it is an ideal optimization classifier for lithology identification while drilling. The research results provide not only a theoretical basis for the drilling geosteering of oilfield development but also valuable information for future basic research. Highlights • Machine learning algorithms can be used for lithology identification while drilling. • Model identification accuracy and training time need to be considered simultaneously. • Only one parameter optimization step is required if its search range is appropriate. • RF model is preferred to identify lithology while drilling. [ABSTRACT FROM AUTHOR]

Abstract With rapid development of unconventional tight and shale reservoirs, considerable amounts of data sets are increasing rapidly. Data mining techniques are becoming attractive alternatives for well performance evaluation and optimization. This paper develops a comprehensive data mining process to evaluate well production performance in Montney Formations in western Canadian sedimentary basin. The general data visualization and statistical data evaluation are used to qualitatively and quantitatively evaluate the relationships between the stimulation parameters and first-year oil production. Then, the recursive feature elimination with cross validation (RFECV) is used to identify the most important factors on the first-year oil production in unconventional reservoirs. In addition, four commonly used supervised learning approaches including random forest (RF), adaptive boosting (AdaBoost), support vector machine (SVM), and neural network (NN) are compared to estimate the first-year well production. The results show that 6 features are the most important variables for constructing an accurate prediction model: well latitude, longitude, well true vertical depth (TVD), proppant pumped per well, well lateral length, and fluid injected per well. Compared to other algorithms, RF has the best prediction performance for the first-year oil production. Furthermore, such data-driven models are found to be very useful for reservoir engineers when designing hydraulic fracture treatments in Montney tight reservoirs. Highlights • The comprehensive data mining process is developed to evaluate well performance in unconventional tight reservoirs. • Feature selection is performed using recursive feature elimination with cross validation. • Various machine learning algorithms are compared to predict the first-year oil production. • The usage of proppant and hydraulic fracturing fluid are optimized by data-drivel models. [ABSTRACT FROM AUTHOR]

Abstract Direct measurements of formation properties such as the shale volume, porosity, permeability, and fluid saturation are often accompanied by expensive cost and are time-consuming too. Well logging inversion provides an alternative way for the determination of formation properties. Compared to traditional theoretical models or formalized empirical fitted models, machine learning assisted logging regression modeling is more accurate and objective. Several machine learning regression algorithms such as neural networks, support vector regression, fuzzy logic, k nearest neighbors regression, multivariate adaptive regression spline, and random forest have already been applied. In this article, we present the Linear Random Forest algorithm and investigate its application in logging regression modeling. By systematic comparison with 8 other algorithms including least squared linear regression, neural networks, epsilon support vector regression, k nearest neighbors regression, regression tree, regression random forest, gradient descent boosted trees, and linear decision tree, the advantage of linear random forest in performance is confirmed by 24 real-world tasks from 7 different areas. Deeper discussions reveal that the advantages of linear random forest source from its strong learning ability, robustness, and feasibility of the hypothesis space. Through our study, the superiority of linear random forest for logging regression modeling is substantiated, which provides a more reasonable way for the further practices of logging regression modeling. Highlights • The Linear Random Forest (LRF) algorithm is presented for better logging regression modeling. • The advantages of LRF in logging regression modeling compared to 8 other algorithms are confirmed by 24 real-world tasks. • The advantages of LRF are explained from the perspective of learning ability, robustness, and feasibility by case studies. • The performance and feasibility of 8 other algorithms are also evaluated and discussed systematically. [ABSTRACT FROM AUTHOR]

Abstract Machine learning provides numerous data-driven tools for automatic pattern recognition. Even though various algorithms such as neural networks and support vector machines have been widely applied, it is still necessary to explore new paradigms and algorithms to improve the machine learning assisted seismic interpretation. Random Forest (RF) is a widely used ensemble algorithm, however, only limited studies of random forest in the seismic application were published. In this article, the methodology of random forest is introduced systematically. Meanwhile, to solve the problem of hyper-parameter determination, we propose an improved algorithm named Pruning Random Forest (PRF). To reveal the advantages of PRF in terms of predictive performance, robustness, and feature selection compared with support vector machine, neural network, and decision tree, several well-designed experiments are executed based on the seismic data of the western Bohai Sea of China. The potential and advantages of random forest in the present case are confirmed by various experiments, which substantiates that the proposed pruning random forest algorithm provides a reliable alternative way for further machine learning assisted seismic interpretation. Highlights • The superiority of applying random forests for seismic interpretation is elucidated. • The Pruning Random Forest (PRF) algorithm is proposed to simplify the hyper-parameter tuning of the ordinary random forest algorithm. • The additional advantages of PRF in predictive performance, robustness and feature selection are revealed by well-designed experiments. • Compared with SVM, NNET, DT and RF, the advantages of pruning random forest in seismic interpretation are confirmed. [ABSTRACT FROM AUTHOR]

Abstract Identification and prediction of facies and fractures are critical to subsurface geosystems analysis and hydrocarbon exploration. However, accurate prediction of facies and fractures is hard in the absence of high-resolution advanced geophysical logs (e.g. image logs) and core samples. This study demonstrates the application of machine learning algorithms, such as Bayesian Network Theory and Random Forest to predict the presence of different facies and fractures in sedimentary rocks using common well logs. We build and train supervised machine learning models using Bayesian Network Theory and Random Forest to classify facies and fractures in unconventional shale and conventional sandstone, and carbonate reservoirs. The trained machine-learning models are cross-validated 10-fold to check their robustness. Application of Bayesian Network theory and Random Forest shows that both facies and fractures can be predicted with high accuracy using limited common well logs. In addition, the Bayesian Network shows the complex causal relationship among the input petrophysical parameters and output (facies or fracture), which is its unique feature compared to other machine learning algorithms, such as Random Forest. Highlights • Application of machine learning algorithms for facies and fracture classification in conventional and unconventional reservoirs. • Bayesian Network provides an understanding of causality of relationships among petrophysical logs. • Random Forest provides a ranking of predictor variables. [ABSTRACT FROM AUTHOR]

Recently, the petroleum industry has focused on deeply buried reservoir discoveries and exploring potential CO 2 storage sites close to existing infrastructure to increase the life span of already operating installations to save time and cost. It is therefore essential for the petroleum industry to find an innovative approach that exploits the existing core- and well log data to be successful in their endeavor of effectively characterizing and predicting reservoir quality. Continuous data sources (e.g. wireline logs) have a huge potential compared with expensive, time inefficient and sporadic data from cores in determining reservoir quality for use in a regional context. However, whereas core analysis offers in-depth knowledge about rock properties and diagenetic processes, continuous data sources can be difficult to interpret without a formation-specific framework. Here, we demonstrated how the pre-existing core data could be effectively used by integrating petrographic- and facies data with a pure predictive machine learning (ML) based porosity predictor. The inclusion of detailed core analysis is important for determining which reservoir parameter(s) that should be modeled and for the interpretation of model outputs. By applying this methodology, a framework for deducing lithological and diagenetic attributes can be established to aid reservoir quality delineation from wireline logs that can be used in frontier areas. With the ML porosity model, a Random Forest Regressor, the square of the correlation was 0.84 between predicted- and helium porosity test data over a large dataset consisting of 38 wells within the Stø Formation across the SW Barents Sea. By integrating the continuous ML porosity logs and core data, it was possible to differentiate three distinct bed types on wireline log responses within the Stø Formation. Particularly, the relationship between Gamma ray (GR) and porosity was effective in separating high porosity clean sand-, low porosity cemented clean sand and more clay and silt rich intervals. Additionally, in the P-wave velocity (VP) - density domain, separation of high porosity clean sand- and heavily cemented low porosity clean sand intervals were possible. The results also show that the ML derived porosity curves coincide with previously published and independent facies data from a selection of the wells included in the study. This demonstrates the applicability of the model in the region, because the Stø Formation has been described to exhibit similar lithological- and mineralogical properties over large parts of the Western Barents Sea area. Even though, continuous porosity data could be estimated from other sources like VP, neutron or density logs, this would generally require matrix and fluid information. This study demonstrated the effectiveness of the ML model in generating continuous porosity logs that are useful for characterizing and predicting reservoir properties in new wells. This methodology offers a workflow for exploiting already acquired core and well log data for frontier exploration that can be adapted to other formations and exploration scenarios worldwide. • Machine learning derived porosity can aid regional reservoir quality delineation. • Efficient use of historical core data to establish predictive frameworks. • ML guided distinct well log responses can differentiate reservoir units of varying quality. [ABSTRACT FROM AUTHOR]

Comprehensive logging curves are necessary for the accurate characterisation of unconventional hydrocarbon formations. However, the logging curves for some wells remain unavailable. Traditional methods of generating the missing logging curves (e.g. multiple regression techniques) have low accuracy, and it is difficult to represent the complex nonlinear relationships between the logging curves of unconventional reservoir using them. The neural random forest (NRF) is a new robust and fault-tolerant machine learning algorithm with high precision. Therefore, we used the NRF model to generate the missing logging curves of a shale gas reservoir in China for the first time. Specifically, four models for generating compensated neutrons, compressional slowness, gamma ray, and density curves were developed based on the NRF framework, incorporating stratigraphic subdivision information. Subsequently, considering the subnetwork connectivity characteristics of the NRF model, the joint and independent methods were separately used to train the model. Finally, the performance of the NRF model was evaluated by comparing it with neural network (NN) and random forest (RF) models. Results revealed that the NRF model achieved superior performance, with R2 > 0.85 on average. Compared to the NN and RF models, the NRF model demonstrated higher prediction accuracy. In addition, the prediction performance of the jointly trained NRF model was slightly superior to that of the independently trained NRF model. Moreover, stratigraphic subdivision information was proved to be important in reducing the model prediction errors and improving the accuracy by almost 20%. In summary, the proposed model provides a cost-effective method for generating the missing logging curves of horizontal wells in the shale gas reservoirs, which will further facilitate the exploration and development of unconventional reservoirs. • Neural random forests (NRF) was proposed for generating the missing well-logging curves in shale gas reservoirs. • The stratigraphic subdivision added in dataset contributed great improvement in prediction accuracy. • Hyperparamters of the NRF model were optimized. • Comparion of traditional machine learning and the NRF model was developed. [ABSTRACT FROM AUTHOR]

Training complex machine learning and deep learning models has become straightforward with the advent of highly efficient, open-source machine learning libraries. Supervised classification techniques such as logistic regression, random forests, and neural networks have also gained popularity in the drilling industry on the back of promising results. As a result, these techniques have been increasingly researched, especially in the domain of drilling vibrations. However, much of this research interest has been limited to finding the best classification model for estimating severity of downhole vibrations. While the choice of classification model is important, we argue that the successful implementation and adoption of machine learning technologies is equally dependent on correctly studying, cleaning, pre-processing the vibration drilling data before applying machine learning techniques. We show that, in certain cases, correctly pre-processing the data guarantees competitive classification performance regardless of the choice of classification model. Specifically, we empirically investigate how factors such as data sampling frequency, data labeling technique, feature extraction technique, and class imbalance impact the performance of different popular classifiers, when dealing with drilling data. We make recommendations specific to vibration classification and highlight pitfalls of certain techniques in that context. Finally, we also develop a step-by-step workflow which enables users to select the correct parameters and techniques at every step, from data collection to model training. • A workflow to systematically address data quality issues for classification is prescribed. • High sampling rates enable higher detection rates for downhole drilling vibrations. • Choice of sampling rate affects classification performance of some classifiers more than others. • Extracted features yield significantly better classification performance than raw features. • F1 score, not test accuracy, must be considered when dealing with imbalanced data sets. [ABSTRACT FROM AUTHOR]

A large body of experimental research supports the effectiveness of Low Salinity Water Injection (LSWI) for enhanced oil recovery from carbonate reservoirs in laboratory scale. Development of robust predictive smart models connecting effective parameters controlling this complex process to Final Recovery Factor (RF f), as the target parameter, is of a paramount importance. The main objective of this research work is to develop intelligent models using Artificial Neural Network (ANN), Support Vector Machine (SVM), Decision Tree (DT), Random Forest (RF), and Committee Machine Intelligent System (CMIS) to forecast performance of LSWI in carbonates using experimental data reported in the literature. Random Search (RS) and Anneal (AL) algorithms were used for optimization of hyperparameters. After data collection from 47 reliable coreflooding studies (582 data points), a rigorous data preprocessing was conducted to ensure quality of the database. Features selection process was used to determine the main parameters controlling LSWI performance in carbonates: brine permeability (K b), core diameter (d), porosity (Φ), and residual water saturation (S wi) of the core, HCO 3 − concentration, and salinity (S) of the connate brine, the salinity (S) of the injected brine, and initial recovery factor (RF i) which were used for development of the models. We considered initial oil recovery (RF i) in this research work, which was not considered in previous works reported in the literature. The applicability domain analysis showed that training and testing response outliers were zero and 9, respectively, indicating acceptable quality of the database. Performance of the developed smart models was analyzed and compared using statistical and graphical error analysis methods. The best performance was obtained for the RF model with Root Mean Square Error (RMSE) of 2.497 and 5.757 for training and testing datasets, respectively, which exhibits a very good agreement with the experimental data. [Display omitted] • A database of 582 data points was assembled using 47 studies of LSWI in carbonates. The largest database ever reported. • A thorough data preprocessing was performed to ensure data quality. • Different artificial intelligent models coupled with optimization algorithms were used to develop predictive models. • Comprehensive analyses of the modeling work were performed. [ABSTRACT FROM AUTHOR]

Main aspect of this study is to explore the possible use of Machine learning based approaches in the petroleum industry. The fourth industrial revolution is being ushered in by advances in computer power. The oil and gas sector contains a huge amount of produced oil data from wells that can be utilised to create data-driven models which can validate empirical correlations. This validation is vital in the current context because the oil market fluctuates a lot, thus making efficient judgments based on insights supplied by data science and machine learning algorithms are valuable. Artificial Neural Network (ANN), Random Forest Regressor (RF) and Gradient Boosting Regressor (GB) were used in this study to forecast daily oil production using available production parameters. In this study, the production data from Equinor's Volve Field was utilised. The obtained data was filtered, and then it was applied to mentioned algorithms with fine tuned hyperparameters. The data was split into training data (70%) and test data (30%). The model was constructed on training data then it was validated through test data. Seven parameters were chosen on the basis of magnitude of correlation coefficient. The result showed that even with the cases of several shut-ins, the hydrocarbon forecasting method for well 159-F-1C predicted production rates with a coefficient of determination (R2score) value greater or equal to 0.90. For performance analysis, coefficient of determination (R2score), mean squared error (MSE) and mean absolute error (MAE) were used for the selected wells. The outcome showed that prediction of hydrocarbon production for well 15/9-F-1 C ANN model showed the best performance (R2 score = 0.9, MAE = 20.52 and MSE = 1268) and for well 15/9-F-12 RF model shows the best performance (R2 score = 0.98, MAE = 84.95 and MSE = 37356.07). This novel approach could be trained on any dataset to assist in oil forecasting. • If sufficient historical data of any field is available it is possible to construct a model that can predict hydrocarbon production. • Data-driven approaches can predict hydrocarbon production for unconventional resources. • Machine learning algorithms cannot replace the conventional methods, but can assist conventional methods in improving efficiency and accuracy. • Different machine learning algorithms should be trained and evaluated on the dataset for selecting the best algorithm. • There is no one specific model for production prediction as no machine learning algorithm is superior. [ABSTRACT FROM AUTHOR]

One of the problems plaguing many Oil and Gas operators is the formation of stable emulsions in their separation systems. Emulsions are highly specific colloidal mixtures of two immiscible fluids, such as water and oil, in the presence of some stabilizing agent. The aim of this work is to assess and predict the stability of emulsions observed in the North Coast Marine Acreage (NCMA) of Trinidad and Tobago. The specific objectives are to (1) Determine the stabilizing factors of emulsions observed in NCMA via laboratory work, (2) Establish a model for predicting emulsion stability expected of a surfactant stabilized Pickering emulsion system and (3) Compare the results of laboratory testing to model results using NCMA emulsion samples. Two types of emulsions were found in the NCMA samples: (1) a cloudy white emulsion and (2) a thick grey rag emulsion. A series of laboratory testing was conducted to determine the stabilizing factors. Two Machine Learning algorithms were attempted, k-Nearest Neighbor and Random Forest Classification, to predict the emulsion result. The Machine Learning models developed were able to predict an outcome of emulsion stability proving the concept that these algorithms can be used for practical assessment. However, the accuracy compared to the laboratory findings was significantly lower than expected at only 45%. This proves that the behaviour of the NCMA emulsion system is unique with few universally applicable trends and warrants further detailed assessment. • An assessment of the stability of Pickering O/W emulsion stabilized by micron-sized particles and surfactants. • Intending to extend the analysis to silica particles in the micron-sized ranges. • Presence of micron-sized silica particles' effect on the stability and morphology of the emulsion. • Machine learning modelling being applied to help predict the stability of the emulsion based on input features commonly measured or known within the Oil and Gas industry. • k-Nearest Neighbor (kNN) and Random Forest Classifier models being used for prediction. [ABSTRACT FROM AUTHOR]

Machine learning supports prediction and inference in multivariate and complex datasets where observations are spatially related to one another. Frequently, these datasets depict spatial autocorrelation that violates the assumption of identically and independently distributed data. Overlooking this correlation result in over-optimistic models that fail to account for the geographical configuration of data. Furthermore, although different data split methods account for spatial autocorrelation, these methods are inflexible, and the parameter training and hyperparameter tuning of the machine learning model is set with a different prediction difficulty than the planned real-world use of the model. In other words, it is an unfair training-testing process. We present a novel method that considers spatial autocorrelation and planned real-world use of the spatial prediction model to design a fair train-test split. Demonstrations include two examples of the planned real-world use of the model using a realistic multivariate synthetic dataset and the analysis of 148 wells from an undisclosed Equinor play. First, the workflow applies the semivariogram model of the target to compute the simple kriging variance as a proxy of spatial estimation difficulty based on the spatial data configuration. Second, the workflow employs a modified rejection sampling to generate a test set with similar prediction difficulty as the planned real-world use of the model. Third, we compare 100 test sets' realizations to the model's planned real-world use, using probability distributions and two divergence metrics: the Jensen-Shannon distance and the mean squared error. The analysis ranks the spatial fair train-test split method as the only one to replicate the difficulty (i.e., kriging variance) compared to the validation set approach and spatial cross-validation. Moreover, the proposed method outperforms the validation set approach, yielding a minor mean percentage error when predicting a target feature in an undisclosed Equinor play using a random forest model. The resulting outputs are training and test sets ready for model fit and assessment with any machine learning algorithm. Thus, the proposed workflow offers spatial aware sets ready for predictive machine learning problems with similar estimation difficulty as the planned real-world use of the model and compatible with any spatial data analysis task. • Our data split method handles spatial autocorrelation and imposes prediction fairness. • The sets impose fair algorithms with similar difficulty in all machine learning steps. • Kriging variance is a surrogate of spatial prediction difficulty. • The resulting training and test sets are compatible with any machine learning model. [ABSTRACT FROM AUTHOR]

Precise and fast estimation of porosity is a vital element of reservoir characterization. A new technology for fast and reliable porosity prediction of chalk samples is presented by applying machine learning methods and X-ray fluorescence (XRF) elemental analysis. Input parameters of prediction models are based on rapid and accurate elemental analysis of chalk samples obtained from Hand-held X-ray fluorescence (HH-XRF) measurements. The intelligent models, including Random Forest (RF), Multilayer perceptron (MLP), Random Forest integrated by Genetic Algorithm (GA-RF) and Multilayer Perceptron integrated by Genetic Algorithm (GA-MLP), are trained and tested based on samples consisting of outcrop chalk samples from Rørdal and Stevns Klint (ST) and core samples from Ekofisk Formation in the North Sea. Results are evaluated by sustainability index (SI), determination coefficient (R2), correlation coefficient (CC), and Willmott's Index of agreement (WI). Results indicate that the combination of GA-RF intelligent method with XRF elemental analysis successfully provides an accurate model by 0.99, 0.02, 0.995 and 0.99 respectively for CC, SI, WI and R2, respectively. • To predict the porosity of chalk samples using Machine learning techniques. • The highest prediction accuracy was owned by GA-RF for predicting the chalk porosity. • The Sensitivity analysis was performed by RReliefF algorithm for input variables. • Silicon, calcium and aluminum play the most significant role in chalk porosity prediction. [ABSTRACT FROM AUTHOR]

Lithology identification methods based on conventional logging data are essential in reservoir geological evaluation. Due to the highly non-linear relationship between lithology and various logging parameters, conventional methods cannot meet the requirements. In recent years, machine learning methods such as Neural Networks and decision tree have been applied to the field of lithology identification and achieved good effects. However, there is no obvious difference in logging parameters for various types of lithology, and at the same time, there is a large amount of information redundancy between each logging curve. Therefore, its uncertainty and fuzziness are high, which interferes with the result of lithology identification. Combining fuzzy theory, decision tree and K-means++ algorithm, this paper proposes a novel hybrid technique of lithology identification which can better overcome the ambiguity and uncertainty of logging data. In the actual data test, we select six logging parameters: density (RHOB), neutron porosity (NPHI), natural gamma (GR), longitudinal wave velocity (VP), shallow formation resistivity (LLS), and deep formation resistivity (LLD). Then K-means++ clustering algorithm was used for clustering analysis on logging data. Finally, the triangular membership function is selected to fuzz the logging data according to the obtained clustering center points, and a fuzzy decision tree lithology identification model is constructed. The prediction accuracy of the model reached 93.92%. The fuzzy decision tree algorithm was also compared with five machine learning algorithms, including decision tree, extremely randomized trees (ET), Adaboost, random forest (RF) and gradient boosting decision tree (GBDT). The results show that the modeling results of fuzzy decision tree algorithm outperform other algorithms. In summary, the fuzzy decision tree model developed in the study is a practical and effective model for complex lithology identification, providing a new idea for lithology identification. • A fuzzy decision tree model using logging data for lithology identification is proposed. • The k-means++ algorithm optimizes the selection of initial clustering centers. • Logging data is fuzzily processed by the triangular membership function. • Pre-pruning the fuzzy decision tree by controlling the values of significance level α and truth degree threshold β. • The accuracy of the model can reach 93.92%, which is better than other algorithms. [ABSTRACT FROM AUTHOR]

Rate of penetration (ROP) management is a matter of importance in drilling operations and it has been considered in different studies. Different machine learning methods such as simple and optimized artificial neural networks (ANNs), support vector machines (SVMs), fuzzy logic (FL) or adaptive neuro-fuzzy inference system (ANFIS), ensemble methods of machine learning and meta-heuristic algorithms have been used for this purpose so far. In this article, some of the studies by using these methods as the main approach for ROP management are reviewed to achieve a better understanding of this concept, its economic benefits and also its research capacities. Results indicate that ANNs are the most popular machine learning method in ROP management, while simple ANNs excel the modified types in this regard. Still, modified ANNs outperform simple ones in terms of prediction accuracy, but as ANNs fall short in superior prediction performance, other machine learning approaches of ROP management such as linear regression (LR), random forest (RF) and gradient boosting method (GBM) have compensated this shortcoming and proved their efficiency and applicability. • Applications of different machine learning methods and optimization algorithms for ROP management are reviewed. • Application of ANN, FL, ANFIS and SVM along with other data-driven methods are compared based on ROP prediction precision. • Ensemble methods of machine learning have found a wide verity of applications in ROP management projects and studies. • Application of meta-heuristic optimization methods are investigated based on the ROP increase from the base case. • In the literature, MLP-ANN and PSO algorithms have been used for ROP prediction and optimization more repeatedly. [ABSTRACT FROM AUTHOR]

Feature Selection, a critical data preprocessing step in machine learning, is an effective way in removing irrelevant variables, thus reducing the dimensionality of input features. Removing uninformative or, even worse, misinformative input columns helps train a machine learning model on a more generalised data with better performances on new and unseen data. In this paper, eight feature selection techniques paired with the gradient boosting regressor model were evaluated based on the statistical comparison of their prediction errors and computational efficiency in characterising a shallow marine reservoir. Analysis of the results shows that the best technique in selecting relevant logs for permeability, porosity and water saturation prediction was the Random Forest, SelectKBest and Lasso regularisation methods, respectively. These techniques did not only reduce the features of the high dimensional dataset but also achieved low prediction errors based on MAE and RMSE and improved computational efficiency. This indicates that the Random Forest, SelectKBest, and Lasso regularisation can identify the best input features for permeability, porosity and water saturation predictions, respectively. [Display omitted] • Feature selection methods are dependent on the distribution of data. • Lasso regularisation selects the best features for water saturation prediction. • Random Forest feature selection is the most effective for permeability prediction. • Feature selection for porosity prediction only improves computational efficiency. • Hybrid machine learning models can be developed using feature selection methods. [ABSTRACT FROM AUTHOR]

Natural gas plays a crucial role in sustaining the economic development. The production behavior of natural gas reservoir is significantly affected by the connecting aquifer. To investigate the aquifer and reservoir properties, the pressure buildup well test is conducted and the pressure transient data need to be analyzed. However, the pressure transient data analysis usually involves a procedure to build an analytical or numerical model to predict the pressure transient dynamics. With the existence of many historical records on the well test data and advancement of machine learning method, a data-driven method is proposed here to automate the well test data analysis of water-drive gas reservoir. The dataset generated from the previous well test records is used as the training set for the machine learning method. To identify the water invasion mode, the random forest classification method is introduced, and the discrete linear segment slopes are extracted from the pressure derivative curve as the feature. To predict the pressure transient dynamics, the random forest regression method is proposed to construct a projection between the aquifer/reservoir properties and the pressure transient curves. Once the water invasion mode is determined and the data-driven model to predict the pressure transient dynamic is constructed, the ensemble Kalman filter is used to estimate the aquifer/reservoir properties from the well testing data. The results show that the random forest classification method can accurately identify the water invasion mode and the random forest regression based ensemble Kalman filter method can estimate the aquifer/reservoir properties accurately with the reduced uncertainty. • Random forest classification method is introduced to identify the water invasion mode in the water-drive gas reservoir. • Random forest regression method is proposed to construct a projection between reservoir properties and the well test curves. • Coupled random forest and ensemble Kalman filter is used to estimate the reservoir properties from the well testing data. [ABSTRACT FROM AUTHOR]

This study aims to put a supervised learning method for automatically classifying lithofacies in well-logging dataset, where several machine learning algorithms were compared in this study that took place in the Tarakan Basin, Indonesia. The predicted lithofacies in this study including shale, shaly sandstone, sandstone, and coal, where coal is considered as the unique lithofacies in the study area. As training and testing datasets, we used two separate well log datasets from the Tarakan Basin. The first well, named Omnicron, was used to train the model, while the second well, named Kay, was used to test it. Random Forest and Gradient Boosting outperformed the other models in the experiment, with the accuracy of 87.49% and 87.01%, respectively. When it came to classifying coal, however, both approaches had issues. The Pr-Recall curve revealed that the coal score was under average precision in each facies, with values of roughly 0.52 and 0.38, respectively, which explaining why, even with high accuracy, the machine learning algorithm predicted poorly in one lithofacies class. In order to evaluate this coal misclassification, we used rock physics to analyze the machine learning prediction in this report. As result, we found that each facies is well-differentiated by physical properties, and the predicted lithofacies have a distribution that is close to the original facies however, coal may be potentially misclassified as other lithofacies as some of the coals have similar rock physical properties with the surrounding lithology (e.g. coal with a mixture of shale may have similar DT and GR responses). Based on this research, the use of machine learning in the Tarakan Basin effectively provides lithofacies data with a high degree of precision and accuracy in a much shorter time. • The Tarakan basin lithofacies categorization using random forest and gradient boosting. • Evaluation each multiclass of the supervised learning methods. • Coal misclassification in case of high accuracy result. • Log-facies classification analysis, and rock physics analysis based on machine learning models. [ABSTRACT FROM AUTHOR]

A reliable method for selecting the most suitable Enhanced Oil Recovery (EOR) method is of a high necessity in reservoir engineering. Machine Learning (ML) techniques have been widely adopted as an effective and fast approach in primary stage of EOR screening. This paper presents a step-by-step procedure of applying ML in EOR screening with focusing on two main goals. Firstly, the problem of EOR screening is investigated from ML point of view. A large database consisting of more than 1000 experiences of worldwide EOR projects has been gathered from various sources. On the basis of this database, several ML models have been developed for predicting the suitable category of EOR methods for a candidate reservoir. These models include: Shallow and Deep Artificial Neural Networks (ANN), Naïve Bayes (NB), Decision Tree (DT), Random Forest (RF) and Dimension Reduction using Principal Component Analysis (PCA). Almost all models performed well in predicting non-observed data with an average accuracy ranging from 80 to 90%. However, RF and the deep ANN models delivered the best performance with an average accuracy of 90%. The results demonstrated that RF model performed much better in comparison with the DT model and improved the classification accuracy from 85 to 91%. In case of ANNs, it was observed that the deeper ANN developed here, improved the classification accuracy from 83 to 90% in comparison with the shallow ANN with one hidden layer. The weakest performance was delivered by NB classifier with an average accuracy of 49%. As the second purpose of this research, reliability of screening results from applying ML techniques is discussed in details. Because, this approach suffers from some serious limitations and challenges, which could unfavorably affect the screening results and never had been considered in previous works. In this work, potential challenges and limitations associated with this well-known approach are introduced from database issues to modeling problems. Accordingly, an important conclusion is that application of ML, although providing a very good insight into primary EOR screening, should not be relied upon as the only method for predictions. It is recommended that the obtained results from this approach should be cross-checked with other EOR screening approaches or with methods based on expert's knowledge. • Application of machine learning (ML) is a well-known approach in EOR screening. • Developing several ML models for selecting the most appropriate EOR method. • Random Forest and Deep ANN categorize EOR methods with the highest accuracy. • EOR screening using ML is associated with some limitations and challenges. • Applying ML methods should not be used as the only method for EOR screening. [ABSTRACT FROM AUTHOR]

Fracture interference has been frequently observed between the parent wells (i.e., the pre-existing wells) and the child wells (i.e., infill wells) in the unconventional tight/shale reservoirs with massive hydraulic fractures. Such interference usually decreases the expected productivity and capital efficiency of the child wells due to the overlapped drainage areas of the parent and child wells. In this study, a new procedure using machine learning (ML) methods is proposed to maximize the productivity of the child wells in Montney formation. More specifically, three types of parent-child well patterns are first defined based on production and trajectory and a comprehensive dataset is developed to include location information on the wells, fracturing parameters, and production history. Then, various strategies are employed to eliminate the redundant and nonsignificant variables and four ML methods including neural networks, random forests, gradient boosting decision trees and linear regression are introduced to predict the first-year production of child wells. Performances of the four ML methods are evaluated with anti-overfitting strategy and the random forests (RF) is found to be the most suitable one for this issue. Finally, a fine-tuned random forest model is applied to optimize the usage of proppant, the volume of fluid injected to get higher productivity of parent-child mode. In addition, the correlation among well spacing and first-year barrels of oil equivalent (BOE) of child well is also analyzed. The calculation results indicated that fracture interference decreases as the spacing increases and the recommend one is above 200 m. The results are also consistent with some actual observation data reported, which further demonstrates the feasibility and credibility of the method adopted. • Three Parent-Child patterns are identified according to the trajectory and production data. • ML methods are used to forecast the productivity of child wells in the Montney formation. • Anti-overfitting methods are successfully employed to reduce the over-training problem. • Well interference is detected for well pairs with a spacing of around 100 m. [ABSTRACT FROM AUTHOR]

This papers deals with the automatic detection and classification of faulty events during the practical operation of oil and gas wells and lines. The events considered here are part of the publicly available 3W database developed by Petrobras, the Brazilian oil holding. Seven fault classes are considered, with distinct dynamics and patterns, as well as several instances of normal operation. A random forest classifier is employed with different statistical measures to identify each fault type. Three experiments are devised in order to evaluate the system performance in distinct classification scenarios. An accuracy rate of 94% indicates a successful performance for the proposed system in detecting real events. Also, the system's time of detection was on average 12% of the transient period that precedes the fault steady-state. • In-depth study of the recently published 3W dataset. • A machine learning-based system that detects faults in oil wells and production lines. • The system-development methodology is applied to three different configurations. • Study and analysis of the multiclass scenario. • The system can be parametrized to improve its performance regarding any metric. • The proposed methodology detects the faults at the beginning of the transient stage. [ABSTRACT FROM AUTHOR]

Geochemical logs are an essential tool for hydrocarbon reservoir characterization. The rock composition given by these logs is useful for porosity calculation, stratigraphic modeling, diagenesis estimation, and well acidification. In reduced logging operations performed in carbonate reservoirs of the Brazilian pre-salt, the geochemical tool is no longer acquired, aiming at optimization and cost reduction. This research aims the development of synthetic geochemical logs using machine learning algorithms. The database includes 22 wells with complete logging acquisition, and the input logs are natural gamma-ray, gamma-ray spectroscopy, density, photoelectric factor, neutron porosity, nuclear magnetic resonance, and sonic. The chemical elements chosen as output are Al, Ca, Fe, Mg, Na, Si, S, and Ti. Five models based on machine learning are trained using 19 wells: Support-vector machine, Multilayer perceptron, Random Forest, AdaBoost, and XGBoost. AdaBoost represents the best algorithm because, in addition to showing the best results, it allowed for a more simplified preprocessing and hyperparameter tuning. The evaluation of the models applies R2 and root mean squared error (RMSE) in validation data and cross-validation. Robust models are acquired for Al, Ca, Fe, Mg, Si, S, and Ti, with R2 above 0.80. Na shows R2 slightly above 0.70. All the models have RMSE between 10-2 to 10-4. The most important logs during training are density, photoelectric factor, K, Th, U, and neutron porosity. Synthetic geochemical logs created for three test wells show a good agreement with acquired logs, being able to reproduce general trends of the pre-salt formations. The machine learning model is capable of substituting the acquisition of geochemical logs with high confidence, representing cost reduction, and supplying engineers and geoscientists with quality data to be used in formation evaluation. • Machine learning capable of substituting the acquisition of geochemical logs. • Estimate reservoir chemical properties where information cannot be acquired. • Reduce wireline logging costs maintaining quality reservoir geochemical data. [ABSTRACT FROM AUTHOR]

Uncertainty is an important indicator that can provide the confidence in the predictions. However, most machine learning methods in classification problems are incapable of predicting an accurate probability that matches the observed distribution of each class. To overcome this problem and improve the uncertainty analysis in lithofacies classification, a novel calibration process is performed on the validation subset using scaling algorithms. The validation subset should be separated from the training data in order to reduce overfitting and avoid unwanted biases. Random forest is selected as the machine learning classifier owing to its relative ease of use, and can handle high-dimensional input features quite well. The proposed approach is applied to two real datasets to demonstrate the improvement in the uncertainty quantification. Reliability diagram shows that the calibrated probability is closer to the diagonal truth, compared with the uncalibrated one. In addition to the visual check, a multiclass Brier score of calibrated model is lower than that by uncalibrated system, which can be referred to as a loss function. For the classification results by calibrated and uncalibrated classifiers on the testing subset, an increase of the Matthews correlation coefficient also means that the calibrated model has a better performance. • Uncertainty analysis is improved in well log classification by random forest. • Platt scaling algorithm is applied to probability values predicted by a pre-fit classifier. • Reliability diagram and Brier score are used to show the improvement in probability prediction. [ABSTRACT FROM AUTHOR]

The identification of lithology while drilling with horizontal well bit is a difficult problem to solve in geosteering. However, due to the existence of "zero length" (the distance between the logging tool and the bit), the lithology at the horizontal well bit cannot be accurately interpreted in real time, which creates a lag in geosteering. Based on logging while drilling (LWD) technology, this paper uses supervised learning in machine learning methods, conventional logging interpretation technology and big data idea in modern computer science, through interdisciplinary theories and methods to research lithology prediction for horizontal well bits in reservoirs. First, a measurement point and vertical reservoir boundary distance (D-MP-VRB) database is built according to different step size categories. Second, based on the D-MP-VRB database, D-MP-VRB prediction models are established using one-versus-one support vector machines (OVO SVMs), random forest (RF), neural networks (NN), and extreme gradient boosting tree (XGBoost) algorithms. To reduce the dimensions of the input data, the feature parameters of the samples are obtained by a correlation analysis of the logging data. The optimal parameter values of each algorithm are determined by grid search and 10-fold cross-validation methods. Finally, the prediction formula of the bit and vertical reservoir boundary distance based on the D-MP-VRB prediction model is established. A case study is performed with data from a sandstone reservoir in Changqing oilfield, Ordos Basin, China. On this basis, the lithology predictions at the bit in real time are carried out by using four models. Considering the principle of model prediction accuracy, through 1320 experiments, only the XGBoost prediction model can be selected, and the step size of the target category is 2 m, however, this model takes the longest time to train. Therefore, the reliable prediction model trained by the sample data of the original training set is used to predict the reservoir information encountered by horizontal well. After the newly drilled reservoir information has accumulated to a certain amount and accurately explained, it is added to the original training set sample data, and the prediction model is retrained to improve the accuracy and adaptability of the model. Based on the prediction results of the XGBoost model and the prediction formula, the distance prediction between the horizontal well bit and the vertical reservoir boundary is realized, real-time lithology correction at the bit is realized, and the adverse effect of the "zero length" on the lithology prediction at the bit is reduced. The research results provide not only a new method for the real-time prediction of lithology in horizontal well bits but also a theoretical basis for the geosteering of oilfield development and valuable information for future research. • A new method for real-time prediction of measurement point and vertical reservoir boundary distance (D-MP-VRB) while drilling is proposed. Real-time prediction models for D-MP-VRB were established and optimized by combining LWD technology with machine learning method based on OVO SVMs, RF, NN and XGBoost algorithms. The simultaneous of efficient drilling and D-MP-VRB prediction is realized. • A prediction formula of the horizontal well bit and vertical reservoir boundary distance based on the D-MP-VRB prediction model is established. The distance prediction between the horizontal well bit and the vertical reservoir boundary is realized, real-time lithology correction at the bit is realized, and the adverse effect of "zero length" on the lithology prediction at the bit is reduced. [ABSTRACT FROM AUTHOR]

Recognition of oil well drilling activities is a crucial task as it allows for identification of nonproductive time. Activity recognition is also important as it can be part of a complete oil well monitoring system. In this paper, we propose a novel hierarchical classifier that consists of Fuzzy Rule-based (FRB) and Random Forest (RF) Classifiers to recognize distinct oil well drilling activities in an oil well drilling process. The novel hierarchical classifier is designed by stacking the FRB and RF classifiers to achieve a classification of the numerous oil well drilling activities with high accuracy. The evaluation of the proposed method is illustrated through a case study of identifying the drilling activities in a real drilling data from oil well rig. Further recommended classifier generates an accurate report of time spent to execute the numerous drilling activities in a complete cycle of the oil well drilling process. Empirical evaluation of the real drilling data shows the efficiency and stability of the proposed method. • This paper presents a novel hierarchical classifier that identifies ten oil well drilling activities. • The method combines a fuzzy rule-based classifier and random forest classifier in a stacked manner to model uncertainty and expert knowledge. • The efficacy of the proposed method is shown by test the model on the real-time oil well drilling data. • The method outperformed as compared to the traditional machine learning algorithm used to detect the oil well drilling activities. [ABSTRACT FROM AUTHOR]

Shale gas is an important unconventional natural gas resource existing in shale reservoir with huge reserves. Due to the ultralow porosity and permeability, it requires the horizontal well drilling and the multi-stage hydraulic fracturing technology to successfully produce the shale gas. The accurate prediction of shale gas production is crucial to the reasonable design of the development plan. However, due to the complex hydraulic fracture network and the gas flow mechanism, the physics-based shale gas production prediction model is still under way. The data-driven model provide an alternative way to deal with the production prediction problem. The multi-objective random forest method is proposed to predict the dynamic production data. The geological and hydraulic fracturing properties are used as input feature. Its prediction performance is evaluated based on the R squared values after determining the appropriate hyper-parameters. The ranking of variable importance can be helpful to improve the interpretability of the data-driven model. The initial peak production rate before declining can be also used as an additional input feature. With the initial peak production rate augmented into the feature set, it can greatly improve the prediction of shale gas production. The variable importance analysis results show that it can be the most influencing factor to prediction accuracy and the ranking of other factors can be altered significantly. The performance of multi-objective random forest (MORF) and multi-output regression chain (MORC) methods are compared, and the comparison result indicates MORC requires a relatively smaller random forest structure, but the prediction performance of MORF is better than MORC. More sample data with less measurement errors can increase the accuracy of the data-driven shale gas production model but there exists a threshold value to improve the accuracy gain. • A data-driven shale gas production model can be built through the multi-objective random forest method. • Variable importance analysis can improve the interpretability of the data-driven shale gas production model. • The multi-objective random forest method can deal with correlated shale gas production data. • It is helpful to collect more data with less measurement errors to build the data-driven model. [ABSTRACT FROM AUTHOR]

Accurate characterization of rock and fluid properties in porous media using X-ray imaging techniques depends on reliable identification and segmentation of the involved phases. Segmentation is critical for the estimation of porosity, fluid saturations, fluid and rock topology, and pore connectivity, among other pore-scale properties. Therefore, the purpose of this study is to compare the effectiveness of different image segmentation techniques when applied to image data analysis in porous media. Two machine learning based segmentation techniques – a supervised ML technique called Fast Random Forest, and an unsupervised method combining k-means and fuzzy c-means clustering algorithms – were compared using an experimental data set. Comparisons are also presented against traditional thresholding segmentation. In addition, we discuss the potential and limitations of applying deep learning based segmentation algorithms. The performance of the segmentation techniques were compared on estimates of porosity, saturation, and surface area, as well as pore-scale estimates such as fluid-fluid interfacial areas, and Euler characteristic. X-ray micro-computed tomography images for a sintered glass frit, saturated with two-phases (air and brine), were acquired at two different voxel resolutions. The high-resolution images (6 μm) were used as the benchmark case, while the low-resolution images (18 μm) were segmented by three segmentation techniques: Fast Random Forest, clustering, and thresholding. The results for porosity and phase saturation from thresholding and from the supervised ML method (i.e. Fast Random Forest) were found to be close to the benchmark case. Segmentation results from the unsupervised ML method (i.e. clustering) were largely unsatisfactory, except for total surface area measurements. The supervised ML segmentation results provided better measurements for air-brine interfacial areas by capturing three-phase interfacial regions. Also, all segmentation techniques resulted in similar measurements for air-phase Euler characteristic confirming poor connectivity of the trapped air phase, although the closest results were obtained by the supervised ML method. Finally, despite the supervised ML segmentation technique being more computationally intensive, it was found to require less user intervention and its implementation was more straightforward. In summary, this work provides insights into different segmentation techniques, their implementation, as well as advantages and limitations with regards to quantitative analysis of pore-scale properties in saturated porous media. • We provide insight into emerging image segmentation techniques, their implementation, and comparative advantages. • Capabilities and limitations of traditional and evolving image segmentation methods are compared on porous media properties. • Properties analyzed include porosity, saturations, surface areas, interfacial areas, and non-wetting phase Euler numbers. [ABSTRACT FROM AUTHOR]

Prediction of permeability is one of the most difficult aspects of reservoir characterization because permeability cannot be directly measured by current well logging technology. This is particularly challenging for carbonate rocks. Machine learning (ML) and robust multivariate methods have been developed that have been used in many fields of study to make accurate estimators for variables of interest from both large and small datasets. ML has been criticized for utilizing approaches that are typically not interpretable. That is, it is not clear how the answers are arrived at and what aspects of input data may be resulting in inaccurate results. The current study uses a number of the mathematical algorithms that operate inside ML modules. It applies them to developing porosity-permeability transforms, with or without rock types, to two well-characterized data sets for carbonate reservoirs. One data set is from Jerry Lucia's 1995 study of carbonate rock types, and the other is from a study of the Seminole, West Texas, San Andres Unit. This study of statistical analysis of porosity-permeability transforms includes: transforming the data to normal distributions; performing cross-validation blind testing; and detecting heteroscedasticity by creating plots of residuals. Heteroscedastic data (populations with variable variance) may have an adverse impact on ML algorithms such as Random Forests (RF). We find that including lithofacies information does not greatly improve porosity-permeability transforms. We also propose a number of strategies to make ML analyses of reservoir (and other geosciences) data sets more robust and accurate. • Lucia rock type transforms are analyzed using modern statistical approaches. • Robust porosity-permeability transforms require careful and extensive pre-processing of data. • A strategy is proposed for performing machine learning analysis of reservoir data to attain repeatable results. [ABSTRACT FROM AUTHOR]

Drilling optimization has been studied extensively given its impact on an oil and gas project – especially in a low-price environment. This is generally approached by optimizing the rate of penetration (ROP) of the well, which may not always be the best strategy. Additionally, optimization strategies rarely include the effect of torsional drilling vibrations (stick-slip) – the biggest ROP inhibitor. As such, most optimization studies have attempted to optimize only one drilling metric – ROP, mechanical specific energy (MSE) or vibrations – independent of the other, despite their interaction downhole. This does not represent bottom hole conditions accurately. This paper introduces a method of coupling several downhole parameters using machine learning algorithms. Individual models for ROP, torque on bit (TOB), MSE, and stick-slip are built using a data-driven modeling approach using the random forests algorithm. The random forest algorithm is used given its lower error rate in modeling ROP (11%), TOB (13%), and classifying drilling vibrations (F-1 score of 0.94) as compared to other machine learning algorithms evaluated in this paper (linear regression, k nearest neighbors, bagging, and neural networks). The models are coupled by building them conjointly during training. The model uses weight-on-bit (WOB), rotary speed, flowrate, and rock strength as inputs to provide an optimal set of these input parameters to use ahead of the bit for optimal drilling. The coupled model was evaluated to optimize ROP and MSE on validation data. ROP increased by 31% and MSE decreased by 49% on average with the use of this optimization model. A case study from the Williston Basin is discussed to illustrate the practical application of this model. • A fully coupled data-driven approach is used to optimize drilling. • Random forests algorithm is used to fit rate of penetration (ROP) and mechanical specific energy (MSE) data to develop models. • Drilling vibrations classification models are used to restrict the optimization space. • Models are coupled by building them conjointly during training. • ROP is increased by 31% and MSE by 49% on average using an end-to-end coupled model. [ABSTRACT FROM AUTHOR]