Introduction to the virtual special issue on AI application in petroleum science
Published 24 January, 2022
When the computer program AlphaGo defeated the human World Champion of the Go game, it symbolised the enormous potential of artificial intelligence (AI), especially deep learning, to problem solve. AI is aiding scientists and engineers from all fields, including petroleum science, to push back the frontiers of research and application. That is why Petroleum Science has organised this virtual special issue to re-introduce our previously published AI-related papers to readers.
Future AI-related papers will be shown in the Article Collection “AI application in Petroleum Science” on ScienceDirect. We invite you to monitor this collection as it develops.
Petroleum Geology and Geophysics
Wang et al. (2021) proposed a novel seismic impedance inversion method based on a cycle-consistent generative adversarial network (Cycle-GAN), which mitigates the shortage of labeled data. The proposed Cycle-GAN model includes two generative subnets and two discriminative subnets. Three kinds of losses, i.e. cycle-consistent loss, adversarial loss, and estimation loss, are adopted to guide the training process. Experiments on synthetic and real datasets show that the predicted impedance curve of the proposed method maintains a better correlation with the true impedance curve.
Zhang et al. (2021) proposed a supervised machine learning method for predicting the spatial distribution of laminated shale with great vertical heterogeneity in case of limited geological data available. Machine learning was used to construct the mapping relationship between the log data and the main mineral components. The results show the approach of 'conventional log data - mineral composition prediction - lamina combination type identification' works well in identifying the types of shale lamina combinations.
Sang et al. (2021) developed a new machine learning reservoir prediction method based on virtual sample generation, which mitigates the small sample issue for neuro network training. The generated virtual samples are consistent with the original data characteristics. Virtual samples screening and model iterative optimization are used to eliminate noise samples and ensure the rationality of virtual samples. The results show that this method can improve the prediction accuracy of machine learning significantly.
Gao et al. (2021) proposed a multi-layer perceptron (MLP) based method to identify low-resistivity-low-contrast (LRLC) pay zones that are easily overlooked due to the resistivity similarity to the water zones. The input of the MLP is a feature space with 49 dimensions, which is generated by clustering the conventional well-log data using the density-based spatial clustering algorithm with noise (DBSCAN), while oil testing results are the output of the MLP. A total of 3192 samples are used for 8-fold cross-validation, which indicates the accuracy of the MLP is 85.53%.
Huang et al. (2021) presented a deep learning-based technique to automatically estimate the local slope of seismic events. Three convolution layers are used to extract structural features in a local window and three fully connected layers serve as a classifier to predict the slope. Although the network is trained using only synthetic seismic data, it can accurately estimate local slopes within real seismic data.
Liu et al. (2020) proposed a lithofacies classification method based on a local deep multi-kernel learning support vector machine (LDMKL-SVM) that can consider low-dimensional global features and high-dimensional local features. The method can automatically learn parameters of kernel function and SVM to build a relationship between lithofacies and seismic elastic information. The results certify the advantages of the method.
Zhang et al. (2019) proposed a novel approach for generating 3-dimensional complex geological facies models based on deep generative models. The networks are trained in an adversarial manner until the generator can create "fake" images that the discriminator cannot distinguish from "real" images. The results demonstrated that the deep learning-driven modeling approach can capture more realistic facies architectures and associations.
Qian et al. (2018) proposed an effective workflow to allow intelligent prediction of sweet spots and comprehensive quantitative characterization of shale oil and gas reservoirs. The workflow is based on mathematical algorithms including fuzzy mathematics, machine learning, and multiple regression analysis. The practical application shows the high accuracy of sweet spot prediction, providing a reliable evaluation of shale reservoir potential.
Guo et al. (2014) proposed a methodology called intelligent prediction and identification system (IPIS) in identifying hydrocarbons. Six parameters (shale content , numbered rock type (RN), porosity , permeability , true resistivity (RT) and spontaneous-potential (SP)) are input into a neuro-fuzzy inference machine (NFIM) to get fluid identification results. The results show the high accuracy and effectiveness of IPIS.
Li et al. (2011) adopted the decision tree, support vector machine, and rough set methods to establish a predictive model of low gas-saturation reservoirs to classify a mass of multi-dimensional and fuzzy data. The predictive model was revised by absorbing the actual reservoir characteristics according to transparency of learning processes and the understandability of learning results. Practical applications indicate that the predictive model is effective in identifying low gas-saturation reservoirs in the study area.
Wang and Zhang (2008) proposed an improved BP model based on the conventional backpropagation (BP) model to predict the lithology of formations to be drilled in the Kela-2 gas field. The prediction model is with main modifications of backpropagation of error, self-adapting algorithm. Examples demonstrate this improved BP neural network is capable to predict formation lithology from sonic and gamma-ray log data, compared with conventional methods.
Petroleum Engineering and Mechanics
Zhang et al. (2021) proposed a multi-source information fused generative adversarial network (MSIGAN) model for parameterization of complex geologies. Various information, e.g. facies distribution, micro-seismic, and inter-well connectivity, can be integrated to learn the geological features of reservoirs. The experimental results show that the proposed MSIGAN model can effectively learn the complex non-Gaussian geological features, which can promote the accuracy of history matching.
Liu et al. (2021) proposed a deep-learning-based algorithm for the estimated ultimate recovery (EUR) evaluation of shale gas wells. They used the EUR evaluation results of 282 wells in the WY shale gas field with geological evaluation data, hydraulic fracturing data, and production data to train the network. Several different cases from the WY shale gas field verified the accuracy of the proposed approach.
Kim et al. (2021) proposed an innovative data integration that uses an iterative-learning method, i.e. a deep neural network (DNN) coupled with a stacked autoencoder (SAE), to solve issues encountered with many objective history matching. The proposed workflow shows an error below 4% for all objective functions, while a typical multi-objective evolutionary algorithm fails to significantly reduce the initial population uncertainty. The workflow would be a reliable alternative to overcome the less-convergent Pareto-based multi-objective evolutionary algorithm in the presence of geological uncertainty and varying objective functions.
Kang et al. (2021) proposed a U-Net deep learning model of the full convolution neural network to segment the pores, fractures, and matrix from the complex rock core with natural fractures, which is more accurate than conventional digital-core construction. A pore fracture network model is then established based on the equivalent radius, which lays the foundation for fluid seepage simulation. The accuracy of the established digital core is verified by the porosity measured via nuclear magnetic resonance (NMR) scanning.
Dong et al. (2021) Dong et al. (2021) proposed a new deep reinforcement learning (DRL) approach based on the double deep Q-network (DDQN) for automatic parameter tuning in three conventional well-testing models. An asynchronous parameter adjustment strategy is used to alleviate the dimensional disaster problem of parameter space. Results show that DDQN requires fewer steps to complete the curve matching than the naive deep Q-network (naive DQN) and deep Q-network (DQN), as well as, DDQN is more robust in curve matching than supervised machine learning algorithms.
Gao et al. (2021) proposed a novel complex network-based deep learning method for characterizing gas-liquid flow. A dual-input convolutional neural network to fuse the raw signals and the graph representation of LPVGs for the classification of flow structures and measurement of gas void fraction. The method provides a promising solution for characterizing gas-liquid flow and measuring flow parameters.
Yu et al. (2020) developed a predictive model for the pipeline friction in the 520-730 m³/h transmission range using the multi-layer. In view of the shortcomings of the multi-layer perceptron–back-propagation (MLP-BP) model, two optimization methods, the genetic algorithm (GA) and mind evolutionary algorithm (MEA), were used to optimize the MLP-BP model. The analysis results showed that the model can effectively guide pipe pigging and optimization.
Yan et al. (2019) proposed an artificial neural network to investigate drag coefficient correlation in the dense gas-solid two-phase flow. The proposed method applied the drag coefficient correlation based on an artificial neural network in the simulations of a bubbling fluidized bed filled with non-spherical particles. Good agreement between the experimental data and the simulation results reveals that the modified drag model can accurately capture the interaction between the gas phase and solid phase.
Wang et al. (2017) presented a virtual sensing technique based on artificial intelligence for vibration and oil debris analysis. They use a new nonlinear feature selection and fusion method named kernel factor analysis, aiming to reduce nonlinearity and uncertainty in the machinery degradation process. The results show that the developed kernel factor analysis method outperforms the state-of-the-art feature selection techniques in terms of virtual sensing model accuracy.
Li et al. (2015) proposed an automatic clustering algorithm called fast black hole-spectral clustering (FBH-SC) and applied it to diagnose down-hole conditions of pumping systems. The FBH algorithm is used to replace the K-mean method in SC, and a critic index function is used as the target function to automatically choose the best scale parameter and clustering number in the clustering process. This algorithm is not limited by any data distribution shapes and is insensitive to the initial cluster centers.
Lashkenari et al. (2013) developed a robust artificial neural network（ANN）code in the MATLAB environment to predict the viscosity of Iranian crude oils. The model was developed using 57 points for saturated oil viscosity, 376 points for viscosity below the bubble point pressure, and 287 points for viscosity above the bubble-point pressure to train a multi-layer perceptron network. It is confirmed that the ANN model has better accuracy and performance in predicting the viscosity of Iranian crudes.
Li et al. (2013) proposed a method based on the curve moment and PSO-SVM for fault diagnosis of sucker rod pumping units. The SVM method is used in this paper for pattern classification, which uses the curve moment parameters of different typical dynamometer cards as learning samples. The simulation results show that the curve moment and the PSO-SVM method have better classification ability to diagnose working conditions of sucker rod pumping units based on the dynamometer card analysis.
Yu et al. (2013) proposed a diagnostic model based on the support vector machine (SVM) method for identifying the working condition of the submersible pumping unit. The inputs of the SVM classifier were the characteristic quantities, the performance, and the misjudgment rate of this method were analyzed and validated by the data acquired from an experimental simulation platform. The SVM classifier had higher diagnostic accuracy than the learning vector quantization (LVQ) classifier.
Ye et al. (2011) used Multivariate regression analysis and artificial neural network (ANN) methods to establish permeability prediction models. Using the proposed models, permeability can be estimated more accurately when relevant data can be obtained from cores, cuttings, and well logs. The results show that the unique characteristics of neural networks enable them to be more successful in predicting permeability than the multivariate regression models.
Li et al. (2008) proposed a new method based on the combination of a neural network and a genetic algorithm to rank the order of exploitation priority of coalbed methane reservoirs. The genetic algorithm was used to optimize the initial connection weights of nerve cells in case the neural network fell into a local minimum. The method can ensure the truthfulness and credibility of the weights of parameters and reduce the probability of falling into local minima.
Petrochemistry and Chemical Engineering
Vijayan et al. (2021) developed adaptive inferential sensors with linear and non-linear local models based on recursive just in time learning (JITL) approach for prediction of naphtha initial boiling point (IBP) and end boiling point (EBP) in the crude distillation unit. Results show that the JITL model based on support vector regression with iterative single data algorithm optimization (ISDA) local model (JITL-SVR: ISDA) yielded the best prediction accuracy in reasonable computation time among the other different types of local models, including locally weighted regression (LWR), multiple linear regression (MLR), partial least squares regression (PLS) and support vector regression (SVR).
Ahmadi et al. (2018) presented the predictive model based on the least-squares support vector machine (LS-SVM) to calculate the CO2–oil swelling factor. A genetic algorithm is used to optimize hyperparameters and ) of the LS-SVM model. LS-SVM is a straightforward and accurate method to be incorporated in commercial reservoir simulators to include the effect of the -oil swelling factor when adequate experimental data are not available.
Dong et al. (2010) developed a new approach for the auto-design of artificial neural networks (ANN) based on a genetic algorithm (GA) for predicting vacuum gas oil (VGO) saturates. The number of neurons in the hidden layer, the momentum and the learning rates are determined by using the genetic algorithm. The artificial neural networks model based on a genetic algorithm can improve the prediction accuracy compared with conventional artificial neural networks models.
Petroleum Economics and Management
Shang et al. (2019) set up a steady-state simulation model of the high-sulfur natural gas purification process. The process applies the uniform design method and the table () to design the experiment points for training and testing the BP model, and then the nondominated sorting genetic algorithm-II has been developed to optimize the two objectives. The results demonstrate that the total comprehensive energy consumption is reduced by 13.4% and the production rate of purified gas is improved by 0.2% under the optimized operating conditions.
Bildirici and Ersin (2015) evaluated the forecasting capabilities of a new class of nonlinear econometric models, namely, the LSTAR-LST-GARCH-RBF and MLP models. The proposed models are the traditional LSTAR-LST-GARCH models further augmented with MLP and RBF type neural networks. The LSTAR-LST-GARCH-RBF and MLP models provide more important gains than traditional models in terms of modeling and forecasting volatility in crude oil prices.
Hou et al. (2015) gave a general review of the two basic techniques in closed-loop reservoir management. The paper summarizes the applications of gradient-based algorithms, gradient-free algorithms, and artificial intelligence algorithms. They discussed the emphases and directions of future research on both automatic matching and reservoir production optimization.
