- Home
- Conferences
- Conference Proceedings
- Conferences
First EAGE/PESGB Workshop Machine Learning
- Conference date: November 29-30, 2018
- Location: London, UK
- Published: 30 November 2018
25 results
-
-
Data-Driven Signal Recognition- A Machine Learning Application For The Real-Time Microseismic Monitoring
Authors A. Shamsa and M. PaydayeshSummaryA simple and robust machine learning technique is applied to automate signal detection and analyse recorded microseismic data. The method’s performance is tested and evaluated on real data. The fracture signals were well-detected using the proposed workflow and techniques when more data were introduced. In contrast to conventional methods, the techniques implemented herein described work on training the model prediction with additional data without restarting from the beginning, making them viable for continuous online learning. This method attempts to remove the burden of labour-intensive processing of microseismic data and replace it with a faster, cheaper, and more accurate way of achieving signal detection.
-
-
-
Pre-Stack Seismic Inversion With Deep Learning
More LessSummaryWe present a study of seismic inversion using deep learning tools. The purpose is to investigate the feasibility of using neural networks to construct acoustic and elastic earth models directly from pre-stack seismic data. Training and testing of the neural networks are done using thousands of synthetic 1D earth models and seismic gathers. We use 2 different types of neural network architectures in our numerical experiments to investigate seismic inversion in different geological scenarios. In both cases, the quality of the prediction is comparable with that obtained from conventional model building processes as such as travel-time and waveform inversion methods. The predicted earth models contain abundant low and medium wave-number information. In terms of performance, training took only less than 30 minutes on 4 GPUs whilst prediction adds negligible cost.
-
-
-
Comparative Study Of Deep Feed Forward Neural Network Application For Seismic Reservoir Characterization
Authors T. Colwell and Ø. KjøsnesSummaryMachine learning has been gaining momentum thanks to a new powerful technique called deep learning ( Bengio, 2016 ). These improvements are due to increasing the depth of neural networks to more than one hidden layer. This study uses a Deep Feed-forward Neural Network (DFNN) to predict reservoir properties from seismic attributes similar to Hampson et al. (2001) . These are shale, porosity and water saturation volumes, ultimately allowing the estimation of the net pay volume. We compare the results of the DFNN to other forms of machining learning such as multi-linear regression (MLR), Probabilistic Neural Network (PNN).
-
-
-
Automatic Facies Classification And Horizon Tracking In 3D Seismic Data
Authors A.J. Bugge, J.E. Lie and S. ClarkSummaryWe present an automatic method that first classify seismic facies and then interpret seismic horizons through four steps; local binary pattern segmentation, unsupervised clustering, supervised classification and dynamic time warping. Our approach avoids the need to manually label data, reducing the need for specialist geological knowledge. We test our method on a structurally complex seismic cube acquired in the SW Barents Sea, targeting rotated Mesozoic fault blocks.
-
-
-
How Machine Learning Is Replacing Conventional Interpretation
More LessSummaryThis presentation shows severaed classification process in successful case histories of the sample-basully finding hydrocarbons and delineating reservoir limits. This type of machine learning is especially good for thin bed exploration as it allows for stratigraphic pattern recognition below conventional seismic tuning.
-
-
-
Machine Learning To Support Technical Document Indexing, How To Measure The Accuracy?
Authors H. Blondelle and J. MicaelliSummaryUsing a machine learning systems, a set of seismic documents has been automatically indexed on 25 metadata. The hold-out methodology has been used to evaluate the accuracy of the models. Results and lessons learnt are discussed.
-
-
-
Semi-Supervised Deep-Learning Applied To UK North Sea Well And Seismic Data
Authors Y. Nishitsuji, R. Exley and J. NasseriSummarySemi-supervised deep-learning architectures provide a multi-layer, pattern recognition, approach that is powerful and ideally suited to the data rich environment that exists at the heart of the oil and gas industry. In this study we apply this technology in order to classify facies using elastic impedances from UK North Sea well and seismic data. The semi-supervised deep-learning method in this study uses a self-training strategy that combines both labelled and unlabelled data during the training phase so that classified data subsequently becomes part of the training dataset in the next iteration. This approach is ideal when the availability of labelled data is limited by practical constraints, which is often the case in subsurface geoscience. The resulting outputs of classified facies were visualised using elastic impedance cross-plots after application to a single training well from a North Sea oil discovery. To validate the result we upscaled the classification model to equivalent seismic data in order to compare the learning from the training well with two blind wells. The results indicate that semi-supervised deep-learning has the potential to accurately determine facies, including hydrocarbon distributions, in subsurface data at a field scale.
-
-
-
Gaussian Mixture Models For Robust Unsupervised Scanning-Electron Microscopy Image Segmentation Of North Sea Chalk
Authors J.S. Dramsch, F. Amour and M. LüthjeSummaryScanning-Electron images from North Sea Chalk are studied for important rock properties. To relieve this manual labor, we investigated several standard image processing methods that underperformed on complicated chalk. Due to the lack of manually labeled data, deep neural networks could not be adequately applied. Gaussian Mixture Models learnt a two-fold representation that separated the background well from the rock. Subsequent morphological filtering cleans up the prediction and enables automatic analysis.
-
-
-
DNN Application For Pseudo-Spectral FWI
By C. ZerafaSummaryFull-waveform inversion (FWI) is a widely used technique in seismic processing to produce high resolution earth models iteratively improved an earth model using a sequence of linearized local inversions to solve a fully non-linear problem. Deep Neural Networks (DNN) are a subset of machine learning algorithms that are efficient in learning non-linear functionals between input and output pairs. The learning process within DNN involves iteratively updating network neuron weights to best approximate input-to-output mapping. There is clearly a similarity between FWI and DNN for optimization applications. I propose casting FWI as a DNN problem and implement a novel approach which learns pseudo-spectral data-driven FWI. I test this methodology by training a DNN on 1D data and then apply this to previously unseen data. Initial results achieved promising levels of accuracy, although not fully reconstructing the model. Future work will investigate deeper DNNs for better generalization and the application to real data.
-
-
-
Deep Learning History Matching For Real Time Production Forecasting
Authors K. Loh, P. Shoeibi Omrani and R. van der LindenSummaryThe forecasting of gas production from mature gas wells, due to their complex end-of-life behaviour, is challenging and often associated with uncertainties (both measurements and modelling uncertainties). Yet, having good forecasts are crucial for operational decision making. In this paper, we present a purely black-box based approach, which combines the use of a data assimilation method, the Ensemble Kalman Filter (EnKF) and a modified deep LSTM model as the prediction model within the approach. This approach is tested on two mature gas wells in the North Sea which were suffering from salt precipitation. Results showed that the approach of combining a deep LSTM model within EnKF can be effective when deployed in a real-time production optimization environment. We observed that having the EnKF increases the robustness of the forecasts by the black box prediction model while reducing computational cost of retraining the black-box models for every individual well.
-
-
-
Stochastic Seismic Waveform Inversion Using Generative Adversarial Networks As A Geological Prior
Authors L. Mosser, O. Dubrule and M. BluntSummarySetting the seismic inversion problem in a Bayesian framework, we seek to obtain the posterior of acoustic rock properties given a set of seismic observations and a prior distribution of the acoustic properties. We use a generative adversarial network (GAN) based on a deep convolutional neural network to represent the prior distribution of acoustic properties. This prior distribution is derived by applying a neural network to a set of Gaussian latent vectors. Samples of the posterior of these latent vectors are obtained using a Metropolis-sampling method that combines gradients obtained from full waveform inversion with back-propagation through the neural network. We apply the proposed method to a synthetic reservoir-scale dataset of channel bodies.
-
-
-
Using Machine Learning Techniques To QC Log Data Before A Study
By J. JohnstonSummaryA large part of a petrophysics project lies in sorting and tidying up the input data, trying to fix the logs where they are bad or missing. Another step is identifying where the log response is not as expected. Typically this is done by looking at log plots and crossplots and making judgements on the fly, often in individual wells. The answers are often people-dependent. The advent of machine learning techniques has the potential to change this by enabling users to incorporate large quantities of data and view differences in a more holistic way. This project involved a set of wells from the Barents Sea with the objective of calibrating the logs with geological observed depositional facies from cored wells, and then using just the logs to propagate those to uncored wells.
-
-
-
Information Theory Considerations In Patch-Based Training Of Deep Neural Networks On Seismic Time-Series
Authors J.S. Dramsch and M. LüthjeSummaryRecent advances in machine learning relies on convolutional deep neural networks. These are often trained on cropped image patches. Pertaining to non-stationary seismic signals this may introduce low frequency noise and non-generalizability.
-
-
-
Seismic Data Interpolation With Conditional Generative Adversarial Networks (cGANs)
Authors D. Oliveira, R. Silva Ferreira, R. Silva and E. Vital BrazilSummaryIn seismic acquisition and processing, several factors may cause missing data or data issues. Primarily, the physical constraints of the method, such as limitation on the available length of the streamer or receiver cable, instrumental and recording problems, and target illumination, e.g., when a geo body shadows the waves, are some of the significant sources of issues in the survey. Many works have tackled this problem using pre-stack data and can be classified into three main categories: wave-equation, domain transform, and prediction-error-filter methods. In this work, we assess the performance of a cGAN (Conditional Generative Adversarial Network) for the interpolation problem in post-stack seismic datasets. To the best of our knowledge, this is the first work to evaluate a deep learning approach in this context. Quantitative and qualitative evaluations of our experiments indicate that deep-networks may present a compelling alternative to classical methods.
-
-
-
Deep Learning Approach For Automatic Detection Of Oil Slicks
Authors Z. Huang, P. Xie and V. MiegebielleSummaryThe aim of this study is to propose a deep learning approach for automatic oil slicks detection over surface of ocean based on Synthetic Aperture Radar (SAR) images. Deep networks such as U-Net is a kind of imagesegmentation- based algorithm which is proved to be effective for varies of image segmentation problems. Here we introduce an U-Net framework for our oil slicks segmentation task. Our database comes from SAR images of 5 differents regions over the world and is divided into training set and test set. With this U-Net structure, we have achieved an overall precision of 93% and a recall rate of 71% with our test set. The algorithm is able to distinguish between oil slicks and other object known as “lookalike”.
-
-
-
Automatic Seismic First Arrival Picking With Deep-Learning
More LessSummaryThis work implements a fully-convolutional neuron network to pick first arrival in difficult field land seismic data. Compared to traditional methods, it greatly improves the productivity. Current work is limited to 2D seismic shot gather and can be extended to 3D without much difficulty. In our test dataset, its picking takes few second per shot and has a credible precision.
-
-
-
Analysis Of Gas Production Data Via An Intelligent Model: Application Natural Gas Production
Authors M.A. Ahmadi and Z. ChenSummaryEstimation of natural gas reserves and forecasting future gas production throughout gas reservoirs is a critical issue for upstream experts. One of the practical approaches for defeating the aforementioned obstacle is decline curve analysis (DCA) which is a mathematical based approach to coordinate actual gas production rates of group of wells, individual wells, or reservoirs with proper function in order to forecast the efficiency of the production in future with the aim of extrapolation of the fitted decline function. Accordingly, applying robust predictive models in this area is of great interest in a gas production system. The current study demonstrates the framework for applying the predictive approach based on coupling artificial neural network and swarm optimization to estimate initial decline rate and cumulative gas production. Particle swarm optimization (PSO) was employed to choose and optimize weights and biases of a neural network which are embedded in PSO-ANN model. Utilization of this model showed high competence of the applied model in terms of coefficient of determination (R2) of 0.9865 and 0.9955, mean squared error (MSE) of 0.00013 and 2.4618 from experimental values for forecasted cumulative gas production and initial decline rate, correspondingly. Executing the suggested model is quite precise and user-friendly to determine the initial decline rate and cumulative gas production with negligible uncertainty. Petroleum experts can easily evolve their own software or program to determine gas reserves and production efficiency in reservoirs.
-
-
-
Implementation Of Meta Heuristic Algorithm And Pressure Match Method To Observe Aquifer Constant In Retrograde Gas Condensate Reservoirs
Authors M.A. Ahmadi and Z. ChenSummaryThe motivation of doing this research was applying the hybrid of pressure match method and genetic algorithm (GA) to optimize the general material balance equation (GMBE) for a condensate gas reservoir with an almost strong aquifer to Figure out its 3 coefficients which are Nfoi, Gfgi and C. The advantage of implementing genetic algorithm (GA) is that the number of parameters which are supposed to be determined is not a concern. There is no doubt that calculating the aquifer constant without taking the reservoir parameters such as viscosity, porosity, net thickness and absolute permeability through making the observer wells much deeper is the most important, beneficial and technical vantage of the mentioned method. The comparison between obtained results from running the method and acquired outputs from the simulator unmask this fact that the pressure match-GA method has highly been successful of determining the coefficients by generating well matched pressures. As a demerit, the method has some problems with lower pressures based on the nature of general material balance equation (GMBE), being rooted in uncertainty, which defeating this obstacle can be considered as a topic for future studies as well as examining the compatibility of the suggested methodology for the heterogeneous reservoirs.
-
-
-
Can Machines Learn To Pick First Breaks As Humans Do?
Authors L. Yalcinoglu and C. StotterSummaryMachine learning is a well-suited tool for first break picking since the process relies on detecting similar features between the seismic traces and thus is a kind of pattern recognition problem. The method we present in this paper applies support vector machine (SVM) as a machine learning algorithm for first break picking which achieve high accuracy.
-
-
-
Streamlining Petrophysical Workflows With Machine Learning
Authors L. MacGregor, N. Brown, A. Roubickova, I. Lampaki, J. Berrizbeitia and M. EllisSummaryThe oil and gas industry is not short of data, in the form of wells, seismic and other geophysical information. However, often because of the complexity of workflows and the time taken to execute them, only a fraction of this information is utilized. Making better use of information, using modern data analytics techniques, and presenting this information in a way that is immediately useful to geologists and decision makers has the potential to dramatically reduce time to decision and the quality of the decision that is made. Here we concentrate on using machine learning approaches to streamline petrophysical workflows. However, to do this requires a rich and diverse training dataset of wells that have been consistently processed for geophysical analysis. The work discussed in this paper has focused on the estimation of clay volume, determination of mineral volumes and determination of porosity and water saturation. A variety of machine learning techniques and algorithms have been tested to find the one most suited to this application. Initial analysis is regionally focused, but we plan to investigate whether the approaches and models developed can be generalized across regions, basins and geological settings.
-
-
-
Functional Estimator For Reservoir Proxy Models Made Scalable Through A Big Data Platform
Authors M. Piantanida, A. Amendola and G. FormatoSummaryThe abstract documents how a Big Data Analytics platform allowed to implement a complex functional estimator of a reservoir proxy model, involving complex machine learning operations on dynamic reservoir models, so that it can scale up to the size of realistic reservoir models.
-
-
-
An Extension For The RA Methodology: Stability Analysis
Authors E. Vital Brazil, R. Silva and L. FariasSummaryWe present an extension for a methodology proposed by Perez-Valiente et al (2014) , known as Reservoirs Analogues (RA). This method finds analogues using machine learning to complete a dataset. Our concern is this methodology does not track error carried from the imputation of missing values until ranking lists of analogues. This study aims to analyze the inherent uncertainty of this step discussing how it can be beneficial to obtain accurate information for reservoirs with limited information.
-
-
-
Building A Robust, Company-Wide Data Science Pipeline Using Programming Abstraction And Virtualization
Authors N. Jones and K. TorbertSummaryThe oil and gas industry presents a challenging and exciting environment for data projects due to the size, complexity, and variability in formatting, type, and quality of the data collected. This environment makes delivering and maintaining a data science pipeline from source systems through to the end user an enormous challenge in many companies ( Scully et al. 2014 ). Many projects fail before any analytics can even applied to the data due to difficulties handling legacy systems, data silos, complex dependencies between data sources, and more. In other cases, data science projects can only advance in one area or division of a company because of differences in data handling despite having broad applicability through the company’s assets. This presentation will discuss California Resources Corporation’s new company-wide data analytics effort as a case study of how we have used technologies like data virtualization ( Van Der Lans, 2018 ) and programming architectural principles such as abstraction to tackle difficult data integration and data quality problems to construct a data science pipeline capable of delivering results company-wide. Many of these problems have frustrated multimillion dollar attempts to address them in the recent past.
-
-
-
An Automated Information Retrieval Platform For Unstructured Well Data Utilizing Smart Machine Learning Algorithms Within A Hybrid Cloud Container
Authors N.M. Hernandez, P.J. Lucañas, J.C. Graciosa, C. Mamador, L. Caezar, I. Panganiban, C. Yu, K.G. Maver and M.G. MaverSummaryThere is a large amount of historic and valuable well information available stored either on paper and more recently as digital documents and reports in the oil and gas industry especially by national data management systems and oil companies. These technical documents contain valuable information from disciplines like geoscience and engineering and are in general stored in a unstructured format. To extract and utilize all this well data, a machine learning-enabled platform, consisting of a carefully selected sequence of algorithms, has been developed as a hybrid cloud container that automatically reads and understands the technical documents with little human supervision. The user can upload raw data to the platform, which are stored on a private local server. The machine learning algorithms are activated and implement the necessary processing and workflows. Structured data is generated as output, which are pushed through to a search engine that is accessible to the user in the cloud. The aim of the platform is to ease the identification of important parts of the technical documents, automatically extract relevant information and visualize it for the user, so they can easily do further analysis, share it with colleagues or agnostically port it to other platforms as input.
-
-
-
Input Data Quality Influence On Lithoclass Predictions In Relation To Supervised Machine Learning
Authors H.W. Bøe, K.B. Brandsegg, L. Marello and A.E. ČrneSummaryWe assess the importance of data availability and consistency prior to applying supervised machine learning for predicting lithoclasses from wireline logs. A dataset is pre-processed and used as training data by three machine learning models in order to investigate the sensitivity of the lithoclasses predictions. The first model uses the quality assured dataset without any modifications. The second model standardizes log signatures, whereas the third model uses the dataset in combination with additional features that dampens extreme outliers. The three models are evaluated against lithofacies interpretations based on CPI’s to show the varying predicting power of the models. The method is applied on a quality-controlled Jurassic interval dataset of ~100 exploration wells within a quadrant of the Norwegian part of the North Sea. The results shows that the number of wireline logs available has a direct influence on the prediction accuracy. For an acceptable prediction accuracy the wells should contain at least the gamma ray, density and neutron log. To distinguish between water-bearing and hydrocarbonbearing intervals in sandstones the resistivity logs should also be present. When implementing machine learning on a regional scale we should consider varying burial depth and depositional environment in order to gain optimal predicting power.
-