- Home
- Conferences
- Conference Proceedings
- Conferences
Second EAGE Workshop on Machine Learning
- Conference date: March 8-9, 2021
- Location: Online
- Published: 08 March 2021
-
-
Assisted interpretation of core images with Deep Learning workflows: lessons learnt from a practical use case
Authors A. Lechevallier, A. Bouziat and S. DesroziersSummaryDeep Learning technologies are increasingly used to automate the processing of unstructured data, such as images, in many industrial and scientific areas.
They have notably proven their efficiency to optimize routine tasks and let expert staff focus on activities of higher added value.
Some applications on geological objects have produced promising results, and systems have been designed for assisted lithofacies characterization on core images.
However only few practical use cases have been documented so far.
In this paper, we confront Deep Learning workflows for image classification with actual core data.
To do so, we use a dataset from an IODP expedition in the Gulf of Corinth, consisting in core images from 3 drilling sites in the Gulf, and an expert interpretation in terms of 17 facies associations.
From this experience, we highlight the main challenges to expect in the assisted interpretation of core images with Artificial Intelligence and share some good practices.
Notably, we describe potential solutions to handle situations where only little training data is available and techniques to choose and tune a model through Transfer Learning.
-
-
-
Application of simple machine learning algorithms to critical stress analysis
By P. WilsonSummaryCritical stress analysis is a geomechanical technique used to understand whether faults are likely to reactivate, given an input effective stress field and fault mechanical properties, and is an important component of fault seal workflows for hydrocarbon exploration and carbon capture and storage applications. As the inputs to the analysis are all typically highly uncertain, it is difficult to cover the range of likely values for all the input parameters. In this contribution we outline a workflow for applying simple machine learning techniques to the results of a large number of stochastic critical stress analysis realizations. The aim is to allow us to better summarize the results of those realizations, gaining insights that can be used in our decision-making process.
-
-
-
Appraisal of several Deep Learning models for microfossil identification on thin section images
Authors A. Koroko, A. Lechevallier, M. Feraille, J. Lecomte, A. Bouziat and S. DesroziersSummaryRecent advances in Artificial Intelligence have led to impressive results allowing the development of reliable predictive systems. The use of Deep Learning and more precisely convolutional neural networks has opened up wide prospects and numerous applications have been proposed. Thanks to these techniques, it is now possible for a computer to recognize the different objects appearing in an image or a film, to identify people precisely or even to analyze documents. However, in comparison with other fields such as medicine, robotics or autonomous driving, these new technologies are still under-exploited in geoscience. In this work, we propose to study the use of Deep Learning and computer vision approaches to automatically detect microfossils of very small sizes in geological images. We train four state-of-the-art Deep Learning methods for object detection with a limited data set of 15 annotated images. The results on 130 other thin section images were qualitatively assessed by expert geologists, and precisions and inference times were quantitatively measured. This work serves as a proof of concept for fully automated microfossil identification, as the four models showed good capabilities in detecting and categorizing the microfossils. However differences in accuracy and performance were underlined, leading to recommendations for similar projects.
-
-
-
A Dynamic Well Classifier in a Heterogeneous Compartmentalized Reservoir Using Advanced Supervised Machine Learning Regression Methods
By M. TranSummaryThis study will guide how to integrate notable regression supervised machine learning approaches, namely multiple linear regression, decision tree regression, and random forest regression in the existing workflow of predicting future performance of drilling prospects. Expert geological insight, methodical statistical refinement, and sound data processing are synergized in a seamless workflow.
The primary objective of this project is to construct a robust workflow to predict critical responses of thousands of planned new development well in a gas condensate project. The responses of interest here are net pay and reserves per foot. These are parameters to evaluate reserves per well, which is extensively used in the prospect ranking workflow and field development plan optimization. Currently, there are thousands of existing producing wells in nearby areas with similar geologic and stratigraphic features. This well collection will serve as a training data block but requires methodical exploratory analysis and strategic statistical refinement. Key predictor variables will be deduced for each response variable
-
-
-
Neural Convolutional Seismic Amplitude Attribute Conv4_4-A: enhancing structural resolution
By A. PiasentinSummaryThis abstract presents a new method to combine seismic attributes and improve the resolution in the new produced seismic attribute map. Specifically in this paper two non-linear amplitude attribute maps of the same area/slice are processed within a deep convolutional neural network to produce an amplitude attribute map with enhanced resolution and with the potential of implementing parametrization for quality control during the seismic processing and inversion process.
-
-
-
Generating synthetic CPTs from marine seismic reflection data using a neural network approach
Authors S. Carpentier, J. Peuchen, B. Paap, B. Boullenger, B. Meijninger, V. Vandeweijer, W. Van Kesteren and F. Van ErpSummaryFor the development of an offshore wind farm, understanding the geological and geotechnical conditions in the upper 100 m below seafloor is crucial when reducing ground risk and designing (cost-effective) wind turbine foundations. We developed and tested a neural network approach to derive predictive (i.e. synthetic) values for CPT parameters– specifically net cone resistance (qn*) – from seismic reflection data. The synthetic parameter values come with a uncertainty bandwidth. Subsequently, the trained neural network was applied to 2D UHR MCS data that were acquired at the planned Hollandse Kust (west) Wind Farm Zone. The goal was to calculate continuous cone resistance values (i.e. qn*) from seismic data using a supervised neural network. Different network architectures and seismic attributes were tested and compared for this purpose. Here, the input data consisted of seismic attributes determined from 2D UHR MCS dataset, and the interpreted geological soil units (based on interpretation of these 2D data). The target training dataset consisted of an extracted subset of seafloor based CPTs. In general, predicted and measured qn* values showed good agreement, especially for the upper 20m below seafloor. Trend-type prediction applies to transitional and strongly layered (<1m scale) soil, as expected.
-
-
-
Wireline Spooling Automation through Computer Vision
Authors T. Su, M. Abuhaikal, C. Bogath and V. VarveropoulosSummaryIn the oil and gas industry, Wireline is used to lower logging toolstrings into a well for well intervention and reservoir evaluation. While the toolstring is retrieved from the well, the Wireline cable, typically under tension, is spooled on a drum. The cable could be thousands of feet long and can stack on the drum for as many as ∼100 layers, with each layer consisting of ∼100 wraps. Proper spooling is important because failure could lead to severe cable damage. Two computer-vision-based applications have been developed: the first one uses a convolutional neural network (CNN) and an encode-decoder network to detect spooling anomaly, the second one uses a similar approach to estimate the cable position in real time.
We evaluated 13 base networks for their convergence rate, accuracy and F1 score. We found that for the two different applications, the winner base network is different. For anomaly detection, the Inception-V3 base network performs the best, while for cable position prediction, the VGG-19 network outperforms others. We optimized the networks using TensorRT. To remove prediction flickering, we tested different filters and found an LSTM-based encoder-decoder network performs the best.
-
-
-
Automatic microseismic signals classification with Deep Learning using multi-input Convolutional Neural Networks
By S. RajeulSummaryMicroseismic monitoring is a key tool for many industry activities, where human-induced seismicity must be monitored for risk mitigation, production enhancement or operations continuity. Recorded signals are complex and composed of different classes: actual microearthquakes, anthropogenic noise, electromagnetic interferences, natural earthquakes, quarry blasts, etc. requiring time-consuming manual expert review. While manual processing can feel like a quality assurance, it is impacted by individual’s interpretation, fatigue and availability. Furthermore, profitability is reduced due to the blend of pertinent and non-essential data.
In this paper, we use Deep Learning with multi-input Convolutional Neural Networks (CNNs) to automate microseismic monitoring signals classification. Goals are threefold: improved quality thanks to consistent machine decisions, enhanced productivity by removing the need for manual sorting of false positives, and new technologies applications such as early warning of risks through 24/7 automatic processing. For that purpose, we train multi-input CNNs to identify images of our labelled data transformed into a time-frequency representation known as the scalogram. We demonstrate the efficiency of the method, capable of successfully reproducing expert classifications in real time, and providing a tool that reduces workload by 90 up to 100%.
-
-
-
Keynote 2: Emerging Opportunities with Data Analytics and Machine Learning in Subsurface Modeling
By M. PyrczSummaryInvited keynote talk. I think I only need a short abstract.
-
-
-
Finding new ideas in a codeless hackathon
By B. O’connorSummaryOGTC set out to deliver a hackathon event to support the ENGenious conference on the 22nd to the 24th of September 2020. Due to the Covid-19 pandemic and the requirement to work virtually, the decision was made to provide a virtual hackathon event to support the conference.
Code[less] was a unique, industry first, virtual, low-code hackathon event provided by OGTC and supported by Microsoft, Intelligent Plant, Offshore Renewable Energy Catapult (OREC), the Oil & Gas Authority (OGA), TAQA, Ithaca Energy, Scottish Enterprise and the Data Lab. OGTC set out to deliver an inclusive virtual hackathon using a codeless approach to raise awareness on the demand for automation and innovation in the Energy Sector and the push towards a Net Zero economy.
The event was open to all levels of people who were interested in developing skills and capabilities in using Microsoft’s Power Platform to solve industry challenges and understand innovative ways of working using real industrial datasets to boost productivity.
-
-
-
Fault detection from 3D seismic data using Artificial Intelligence
Authors J. Lowell and P. SzafianSummaryAccurate delineation and interpretation of the fault network in oil and gas reservoirs and their subsequent impact on reservoir performance is one of the most generalised components within modern basin and reservoir modelling workflows, partly because of the complexity of the task and the time constraints forever present in the G&G workflows. With the help of Artificial Intelligence, the speed and accuracy of fault delineation upon delivery of new seismic data enable regional and field assessment to benefit from the most up to date information and ensure significantly better-informed exploration and development decisions. A Foundation Network was developed to identify faults in a seismic cube. In this network the Artificial Intelligence is closely aligned with the interpreters’ way of working, allowing tightly coupled interaction as appropriate for the dataset and the individual interpreter’s workflow. Case studies show that the automated components of the AI-assisted workflow, combined with capturing the interpreter’s knowledge and experience, have demonstrated tremendous value in delineating the intricate details of a realistic subsurface, significantly reducing interpretation turnaround times while simultaneously increasing accuracy and comprehensiveness of the interpretation itself.
-
-
-
Bayesian history matching with machine learning objective function estimator
Authors N. Voskresenskiy, A. Kadyrova, T. Karimov, D. Agirre Kabrera, V. Bogan and A. KhlyupinSummaryIn this paper we propose Bayesian history matching workflow with the application of machine learning algorithms and discuss its implementation to synthetic and real hydrodynamic models. Since the traditional Bayesian approach is computationally intensive we suggest substituting simulator runs with a machine-learning algorithm to estimate the objective function value for generated models. For this purpose, we consider Kernel Ridge Regression trained on prior models obtained from the Latin Hypercube Sampling procedure. The workflow allows to reduce the uncertainty of model parameters and to indicate objective function sensitivity to model parameters.
-
-
-
SASI: a metric for Salt Body Reconstruction
Authors M. Araya-Polo and Y. SunSummaryWe propose a new metric called Spatial Aware Similitude Index (SASI) to compute quality of prediction and models that contain salt bodies. The metric is more informative than traditional general metrics such as SSIM or Jaccard.
-
-
-
TOC prediction using machine learning - a case study from offshore Norway
Authors D. Stoddart, L. Mosser, A. Hartwig and P. AursandSummaryThis work details how ML methodologies have been applied to predict TOC in source rock on a well and seismic scale.
-