Academic literature on the topic 'AdaBoost (Adaptive Boosting)'

Abstract Lost circulation during drilling operations is a persistent challenge in the oil and gas industry, leading to significant financial losses and increased non-productive time. The common use of lost circulation materials (LCMs) in drilling fluids helps mitigate mud loss only to an extent. However, predicting the extent of mud loss before drilling specific formations would greatly benefit engineers. This study aims to predict mud loss using advanced boosting machine learning frameworks, addressing the need for more accurate forecasting tools. We evaluated three ensemble boosting algorithms—Adaptive Boosting (AdaBoost), Light Gradient Boosting Machine (LightGBM), and Extreme Gradient Boosting (XGBoost)—and compared them to Random Forest, a baseline bagging algorithm. Utilizing a dataset of over 7,000 data points with 27 features from drilling operations in Well MXY at the Utah FORGE field, we found that XGBoost and Random Forest were the most accurate models, with R2 scores of 0.935 and 0.934, respectively. These results indicate that while XGBoost is the top-performing framework, Random Forest remains a robust and reliable method for predicting lost circulation, providing valuable insights for drilling engineers.

In this paper, we investigate four key issues associated with data-driven approaches for fault classification using the Pratt and Whitney commercial dual-spool turbofan engine data as a test case. The four issues considered here include: (1) Can we characterize, a priori, the difficulty of fault classification via self-organizing maps? (2) Do data reduction techniques improve fault classification performance and enable the implementation of data-driven classification techniques in memory-constrained digital electronic control units (DECUs)?, (3) When does adaptive boosting, an incremental fusion method that successively combines moderately inaccurate classifiers into accurate ones, help improve classification performance?, and (4) How to synthesize classifier fusion architectures to improve the overall diagnostic accuracy? The classifiers studied in this paper are the support vector machine (SVM), probabilistic neural network (PNN), k-nearest neighbor (KNN), principal component analysis (PCA), Gaussian mixture models (GMM), and a physics-based single fault isolator (SFI). As these algorithms operate on large volumes of data and are generally computationally expensive, we reduce the dataset using the multi-way partial least squares (MPLS) method. This has the added benefits of improved diagnostic accuracy and smaller memory requirements. The performance of the moderately inaccurate classifiers is improved using adaptive boosting (AdaBoost). These results are compared to the results of the classifiers alone, as well as different fusion architectures. We show that fusion reduces the variability in diagnostic accuracy, and is most useful when combining moderately inaccurate classifiers.

This paper presented here deals with study of identification and verification approach of Diabetes based on human iris pattern. In the pre-processing of this work, region of interest according to color (ROI) concept is used for iris localization, Dougman's rubber sheet model is used for normalization and Circular Hough Transform can be used for pupil and boundary detection. To extract features, Gabor Filter, Histogram of Oriented Gradients, five level decomposition of wavelet transforms likeHaar, db2, db4, bior 2.2, bior6.8 waveletscan be used. Binary coding scheme binaries’ the feature vector coefficients and classifier like hamming distance, Support Vector Machine (SVM), Adaptive Boosting (AdaBoost), Neural Networks (NN), Random Forest (RF) and Linear Discriminative Analysis (LDA) with shrinkage parametercan be used for template matching. Performance parameters such as Computational time, Hamming distance variation, False Acceptance Rate (FAR), False Rejection Rate (FRR), Accuracy, and Match ratio can be calculated for the comparison purpose.

Abstract With the advancement in machine learning (ML) applications, some recent research has been conducted to optimize fracturing treatments. There are a variety of models available using various objective functions for optimization and different mathematical techniques. There is a need to extend the ML techniques to optimize the choice of algorithm. For fracturing treatment design, the literature for comparative algorithm performance is sparse. The research predominantly shows that compared to the most commonly used regressors and classifiers, some sort of boosting technique consistently outperforms on model testing and prediction accuracy. A database was constructed for a heterogeneous reservoir. Four widely used boosting algorithms were used on the database to predict the design only from the output of a short injection/falloff test. Feature importance analysis was done on eight output parameters from the falloff analysis, and six were finalized for the model construction. The outputs selected for prediction were fracturing fluid efficiency, proppant mass, maximum proppant concentration, and injection rate. Extreme gradient boost (XGBoost), categorical boost (CatBoost), adaptive boost (AdaBoost), and light gradient boosting machine (LGBM) were the algorithms finalized for the comparative study. The sensitivity was done for a different number of classes (four, five, and six) to establish a balance between accuracy and prediction granularity. The results showed that the best algorithm choice was between XGBoost and CatBoost for the predicted parameters under certain model construction conditions. The accuracy for all outputs for the holdout sets varied between 80 and 92%, showing robust significance for a wider utilization of these models. Data science has contributed to various oil and gas industry domains and has tremendous applications in the stimulation domain. The research and review conducted in this paper add a valuable resource for the user to build digital databases and use the appropriate algorithm without much trial and error. Implementing this model reduced the complexity of the proppant fracturing treatment redesign process, enhanced operational efficiency, and reduced fracture damage by eliminating minifrac steps with crosslinked gel.

Abstract Rock facies are typically identified either by core analysis to provide visually interpreted lithofacies, or determined indirectly based on suites of recorded well-log data, thereby generating electrofacies interpretations. Since the lithofacies cannot be obtained for all reservoir intervals, drilled section and/or wells, it is commonly essential to model the discrete lithofacies as a function of well-log data (electrofacies) to predict the poorly sampled or non-cored intervals. The process is called predictive lithofacies classification. In this study, measured discrete lithofacies distributions (based on core data) are comparatively modeled with well-log data using two tree-based ensemble algorithms: extreme gradient boosting (XGBoost) and adaptive boosting (AdaBoost) configured as classifiers. The predicted lithofacies are then combined with recorded well-log data for analysis by an XGBoost regression model to predict permeability. The input well-log variables are log porosity, gamma ray, water saturation, neutron porosity, deep resistivity, and bulk density. The data are derived from the Mishrif carbonate reservoir in a giant southern Iraqi oil field. For efficient lithofacies classification and permeability modelling, random sub-sampling cross-validation was applied to the well-log dataset to generate two subsets: training subset for model tuning; and testing subset for prediction of data points unseen during training of the model. Confusion matrices and the total correct percentage (TCP) of predictions are used to measure the prediction performance of each algorithm to identify the most realistic lithofacies classification. The TCPs for XGBoost and AdaBoost classifiers for the training subset were 98% and 100%, respectively. However, the TCPs achieved for the testing subsets were 97%, and 96%, respectively. The mismatch between the measured and predicted permeability from the XGBoost regressor was determined using root mean square error. The XGBoost model provides accurate lithofacies classification and permeability predictions of the cored data. The XGBoost model is therefore considered suitable for providing reliable predictions of lithofacies and permeability for the non-cored intervals of the same well and for non-cored wells in the studied reservoir. The workflow for lithofacies and permeability prediction was fully implemented and visualized using R open-source codes.

Abstract Rock facies are typically identified either by core analysis to provide visually interpreted lithofacies, or determined indirectly based on suites of recorded well-log data, thereby generating electrofacies interpretations. Since the lithofacies cannot be obtained for all reservoir intervals, drilled section and/or wells, it is commonly essential to model the discrete lithofacies as a function of well-log data (electrofacies) to predict the poorly sampled or non-cored intervals. The process is called predictive lithofacies classification. In this study, measured discrete lithofacies distributions (based on core data) are comparatively modeled with well-log data using two tree-based ensemble algorithms: extreme gradient boosting (XGBoost) and adaptive boosting (AdaBoost) configured as classifiers. The predicted lithofacies are then combined with recorded well-log data for analysis by an XGBoost regression model to predict permeability. The input well-log variables are log porosity, gamma ray, water saturation, neutron porosity, deep resistivity, and bulk density. The data are derived from the Mishrif carbonate reservoir in a giant southern Iraqi oil field. For efficient lithofacies classification and permeability modelling, random sub-sampling cross-validation was applied to the well-log dataset to generate two subsets: training subset for model tuning; and testing subset for prediction of data points unseen during training of the model. Confusion matrices and the total correct percentage (TCP) of predictions are used to measure the prediction performance of each algorithm to identify the most realistic lithofacies classification. The TCPs for XGBoost and AdaBoost classifiers for the training subset were 98% and 100%, respectively. However, the TCPs achieved for the testing subsets were 97%, and 96%, respectively. The mismatch between the measured and predicted permeability from the XGBoost regressor was determined using root mean square error. The XGBoost model provides accurate lithofacies classification and permeability predictions of the cored data. The XGBoost model is therefore considered suitable for providing reliable predictions of lithofacies and permeability for the non-cored intervals of the same well and for non-cored wells in the studied reservoir. The workflow for lithofacies and permeability prediction was fully implemented and visualized using R open-source codes.

Abstract There is a strong push for decarbonization in the Global North and South to combat climate change and avert its dire consequences including rising sea levels, severe weather events, biodiversity loss, and threats to food and water security. Hydrogen (H2) emerges as a sustainable energy carrier in this regard. Underground hydrogen storage (UHS) presents significant potential but requires a thorough understanding of H2 behavior in porous media. One of the crucial parameters is interfacial tension (IFT) which influences capillary entry pressure, H2 column height, and storage capacity. Accurate IFT measurement is therefore imperative but current laboratory methods are resource-intensive. To address the limitation, the current study employed computational techniques to estimate the IFT of aqueous H2/cushion gases systems using minimum resources. The novelty of this investigation lies in the comparative analysis of diverse techniques including ensemble learning, evolutionary algorithm, and molecular dynamics simulation. Specifically, this study employed Random Forest and Adaptive Boosting (AdaBoost) to predict the IFT of aqueous H2/cushion gases systems. We utilized a comprehensive database of 2400 experimental data points encompassing pressure (0.50–45.20 MPa), temperature (298–449 K), brine salinity (0–3.42 mol/kg), and gas mole fractions (0–100 mol%). Numerical statistical error measures and multiple data visualizations were used to evaluate the performance of the developed models. The best smart model was also compared to an evolutionary algorithm and molecular dynamics simulation to ascertain robustness and generalizability. Subsequently, sensitivity analysis and feature importance were performed to comprehend the impact of the individual input parameters on the IFT. Finally, the implication of model-predicted IFT on UHS was extensively investigated. The developed paradigms demonstrated remarkable predictive prowess, yielding a coefficient of determination>0.96, average absolute deviation (AAD)<1.50 mN/m, and root mean square error (RMSE)<2.1 mN/m, with AdaBoost emerging as the upper-echelon paradigm. Additionally, AdaBoost exhibited predictive dominance over genetic programming and molecular-dynamics-based correlations developed by other scholars. Sensitivity analysis identified carbon dioxide mole fraction as the main factor influencing IFT. Analysis also indicate that higher concentrations of the cushion gas, particularly CO2 significantly reduces IFT and this could compromise structural hydrogen trapping and lead to H2 loss. AdaBoost also provided precise results for capillary entry pressure (11.9–12.1 MPa), H2 storage height (1177–1217 m), and storage capacity (690–2226 kg/m2) of a Saudi basaltic formation comparable to experimental findings (2-3% deviation), affirming its potential for practical applications. AdaBoost can also be utilized in conjunction with Monte Carlo simulation for a swift and accurate assessment of the uncertainty associated with the influencing factors of UHS. The paradigms offer a promising approach to estimating IFT using minimal resources. Additionally, the findings facilitate UHS optimization, reduce H2 loss, and improve storage capacity.

Abstract Modeling the drill bit Rate of Penetration (ROP) is crucial for optimizing drilling operations as maximum ROP causes fast drilling, reflecting efficient rig performance and productivity. In this paper, four Ensemble machine learning (ML) algorithms were adopted to reconstruct ROP predictive models: Random Forest (RF), Gradient Boosting (GB), Extreme Gradient Boost (XGB), and Adaptive Boosting (AdaBoost). The research was implemented on well data for the entire stratigraphy column in a giant Southern Iraqi oil field. The drilling operations in the oil field pass through 19 formations (including 4 oil-bearing reservoirs) from Dibdibba to Zubair in a total depth of approximately 3200 m. From the stratigraphic column, various lithology types exist, such as carbonate and clastic with distinct thicknesses that range from (40-440) m. The ROP predictive models were built given 14 operating parameters: Total Vertical Depth (TVD), Weight on Bit (WOB), Rotation per Minute (RPM), Torque, Total RPM, flow rate, Standpipe Pressure (SPP), effective density, bit size, D exponent, Gamma Ray (GR), density, neutron, and caliper, and the discrete lithology distribution. For ROP modeling and validation, a dataset that combines information from three development wells was collected, randomly subsampled, and then subdivided into 85% for training and 15% for validation and testing. The root means square prediction error (RMSE) and coefficient of correlation (R-sq) were used as statistical mismatch quantification tools between the measured and predicted ROP given the test subset. Except for Adaboost, all the other three ML approaches have given acceptable accurate ROP predictions with good matching between the ROP to the measured and predicted for the testing subset in addition to the prediction for each well across the entire depth. This integrated modeling workflow with cross-validation of combining three wells together has resulted in more accurate prediction than using one well as a reference for prediction. In the ROP optimization, determining the optimal set of the 14 operational parameters leads to the fastest penetration rate and most economic drilling. The presented workflow is not only predicting the proper penetration rate but also optimizing the drilling parameters and reducing the drilling cost of future wells. Additionally, the resulting ROP ML-predictive models can be implemented for the prediction of the drilling rate of penetration in other areas of this oil field and also other nearby fields of the similar stratigraphic columns.

Abstract CO2 Interfacial Tension (IFT) and the reservoir rock-fluid interfacial interactions are critical parameters for successful CO2 geological sequestration, where the success relies significantly on the rock-CO2-brine interactions. IFT behaviors during storage dictate the CO2/brine distribution at pore scale and the residual/structural trapping potentials of storage/caprocks. Experimental assessment of CO2-Brine IFT as a function of pressure, temperature, and readily available organic contaminations on rock surfaces is arduous because of high CO2 reactivity and embrittlement damages. Data-driven machine learning (ML) modeling of CO2-brine IFT are less strenuous and more precise. They can be conducted at geo-storage conditions that are complex and hazardous to attain in the laboratory. In this study, we have applied three different machine learning techniques, including Random Forest (RF), XGBoost (XGB), and Adaptive Gradient Boosting (AGB), to predict the interfacial tension of the CO2 in brine system. The performance of the ML models was assessed through various assessment tests, such as cross-plots, average absolute percentage error (AAPE), root mean square error (RMSE), and coefficient of determination (R2). The outcomes of the predictions indicated that the XGB outperformed the RF, and AdaBoost. The XGB yielded remarkably low error rates. With optimal settings, the output was predicted with 97% accuracy. The proposed methodology can minimize the experimental cost of measuring rheological parameters and serve as a quick assessment tool.

Stroke is one of the leading causes of death and disability worldwide and requires an immediate attention as the longer the patient is left untreated, the more sever its outcomes are. Enhancing access to optimal treatment and reducing mortality rates require improving the accuracy of stroke characterization methods in prehospital settings. This study explores how video analysis and machine learning (ML) can be leveraged to identify stroke symptoms on the National Institute of Health Stroke Scale (NIHSS), with the goal of facilitating the prehospital management of patients with suspected stroke. A total of 888 videos were captured from the research group members, who mimicked stroke symptoms including facial palsy, leg and arm paresis, ataxia and dysarthria, following the criteria of the NIHSS. Multiple algorithms, utilized in earlier studies, were examined to predict these symptoms, and their performance was assessed using accuracy, sensitivity and specificity. The best method for detecting facial palsy was found using Histogram of Oriented Gradients (HOG) features in conjunction with Adaptive Boosting (AdaBoost), achieving an accuracy, sensitivity and specificity values of 97.8%, 98.0% and 97.0%, respectively. The identification of arm paresis reached 100% on all metrics using a combination of MediaPipe and SVM. For leg paresis, all algorithms had poor detection rates. The outcome for ataxia for both limbs varied. Google Cloud Speech-to-Text was used to detect dysarthria and reached 100% on all evaluation metrics. These findings suggest that video analysis and ML have the potential to assist early stroke diagnosis, but further research is needed to validate this.