Fourth EAGE Digitalization Conference & Exhibition

The manual detection of the first break (FB) in seismic data relies on visual recognition of amplitude and waveform variations by experts. However, for vast datasets generated during seismic acquisitions, this manual process becomes highly time-consuming and subject to individual interpretation. In this study, we propose an automated FB picking method based on neural networks for detecting the initial arrival of synthetic data resembling the characteristics of the Middle Magdalena Valley in Colombia, a basin historically associated with hydrocarbon exploration. Our approach involves supervised training of a neural network (NN) comprising a 1D convolutional layer and two dense layers. The NN categorizes reprocessed trace samples into two groups: pre-FB and post-FB. Subsequently, through post-processing, we identify the most likely sample corresponding to the FB. Our method successfully detects the first arrival in 77.36%, 91.19%, and 95.86% of cases, allowing for a margin of error of 10 samples when signal-to-noise ratios (SNRs) are 0 dB, 6 dB, and 20 dB, respectively. This demonstrates its effectiveness, particularly for noisy signal conditions.

Integrating mixed reality (MR) technologies into the core interpretation process offers a promising solution to the challenges faced by traditional methods. The MR system, utilizing a headset with depth sensor technology, bridges the gap between digital data and physical core samples, allowing for the direct overlay of essential digital information onto physical core samples. This approach merges digital data with the physical world, significantly enhancing the accuracy and depth of core analysis. The depth sensor uses structured light range sensing, projecting a structured infrared light pattern onto the environment. As this infrared pattern interacts with objects in its path, the sensor captures the distortions in the pattern to accurately gauge depth and surface information, which is crucial for the device’s immersive mixed-reality experience. Users can view multiple logs overlaid on their corresponding core sample, with the system displaying relevant laboratory results directly on the physical core. The immersive MR experience encompasses wells, thin sections, and core sequence details, transforming the laboratory space into an informative display. Mixed reality technology creates a more interactive core analysis and interpretation environment, equipping geoscientists and engineers with dynamic tools for more precise examinations of geological cores.

Automated lineament detection based on the deep learning model, using gravity data and subsurface lineaments, was verified to avoid problems such as accuracy and consistency of the manual lineament extraction. For the training and detection, the model used both extracted lineaments and gravity data. The gravity data was transformed to shape index which later input to the model. The detection architecture is composed of “Lineament Detector” and “Pseudo-Hough Transformer” employing FPN and U-net, respectively. Pseudo-Hough Transformer, which is connected to Lineament Detector, has the role of forcing the model to highlight linear geometry. The automated detection firmly tracks the extracted lineaments in the validation area. The detection shows a desirable geometry that is linearly corrected by the Pseudo-Hough Transformer. Following evaluation reveals that the contribution of each lineament to the extraction accuracy is not simply a function of extraction difficulty of lineaments. The contribution tends to be lower for lineaments that can be predicted with relatively high accuracy even without including in the training data.

The GECO (Geological ECOsystem) project is aimed at implementing an ecosystem where any data sources, structured or unstructured, can be integrated and used to extract key insights and in the development of the geological model for the area of interest. The vision behind GECO is first and foremost to provide Geoscientists with a digital playground where it will be possible to structure relationships, to promote know-how exchange among the specialists, increasing at the same time the efficiency and the accuracy of the search results and data analytics, and reducing the realization time. Moreover, GECO aims to support the Geoscience community in the development and deployment of new digital workflows.

Sedimentological core description plays a crucial role in the understanding of subsurface data, and is essential for geological research. In the past vertical cores, extracted during drilling, were described in a graphical format. In recent years there has been a technological shift towards capturing rock properties directly in digital format, using a range of proprietary hardware and software. Digitization process started with data standardization whereby templates and agreed upon symbols were implemented. The next step was to create the appropriate data model to accommodate digitized core descriptions. In order to be incorporated in the database, a list of codes for each geological feature (texture, bedding contacts, sedimentary structures, etc.) was created. Furthermore, to expedite and automate the data extractions from standardized templates, prior to being uploaded to the database, a customized export script was created. We have also produced customized scripts to map and extract the digitized core descriptions directly to various basin modeling and visualization software packages used within the company through dedicated templates.

This study explores the transformation of Special Core Analysis (SCAL) data utilization as part of subsurface digitalization. The focus is on a cloud-native solution for leveraging SCAL data from the Norwegian Continental Shelf (NCS) and maximizing these resources to generate synthetic capillary pressure and relative permeability curves without the need for new core samples.

This innovative approach involves digitizing, structuring, and integrating SCAL data into a unified cloud platform, employing AI-based similarity ranking and SCAL trend analysis. The method integrates data collection, interpretation, and storage in the native cloud-based system, utilizing statistical techniques to identify analogous SCAL data. It enables the prediction of representative relative permeability and capillary pressure curves, as illustrated through field examples and uncertainty ranges based on the LET method.

Key conclusions highlight three main solutions: Core Data Management, SCAL Analog, and SCAL Trend Analysis. These solutions enhance reservoir characterization, facilitate data-driven decisions, and support full-field applications. This platform is adaptable, transparent, and open to further enhancements like SCAL upscaling and rock typing, presenting promising potential for broader applications in the oil and gas industry.

Oil and gas raster images are important source of data in understanding reservoir characteristics, drilling operations, and future well planning. Raster digitization i.e., converting traditional paper raster logs into digital formats can help in tapping the valuable information across diverse set of rasters for various applications such as well log interpretation and correlation. As the first step towards the raster digitization, the raster segmentation task automatically extracts the important raster regions such as plot segment, depth track, and log headers. While raster segmentation can effectively be performed using modern deep learning (DL)-based methods, such methods perform poorly on raster data which is not used during DL model training. In this work, a novel retraining framework is proposed to improve the performance of raster segmentation module on out-of-(training)-distribution data. The retraining of a DL model is a time-consuming process because of the data labeling requirement. The proposed framework overcomes the challenge of large-scale data labeling by demanding very few data labels from the user while simultaneously improving the model performance. Moreover, the novel retraining framework can also address the data residency and privacy issues with respect to different users.

Wintershall Dea has created a central Datahub solution with the help of different disciplines across the company like geologists and developers. We live in a changing world with lots of data, therefore we have created a solution which suits many different requirements with modern state of the art technologies to make use of all the existing data in an easy, user-friendly way.

Subsurface data is stored across multiple applications, databases, and in a wide range of flat file formats. Data may have different modalities (text, image, numerical), be unstructured or structured, or conform to different domain models. This variation makes data interoperability a major challenge for the industry. Finding, querying, and analyzing data by subsurface professionals is therefore a difficult and time-consuming task. We explore how recent advancements in artificial intelligence, specifically large language models, can be used to extract unstructured data from text into a domain model that can then easily be combined with application data. Once the domain model is population with multiple sources, language model -based agents can be used to ask any natural language question against the data. We find that language models are effective tools for data extraction when combined with domain knowledge and field descriptors in the prompt context. In our design, the domain model and field descriptors are highly customizable depending on the user’s data ecosystem. We also find that language model-based agents are effective at distilling user queries against the data once combined in a common data model.

Lithology is an important key factor in petroleum exploration and reservoir characterization. The traditional method for lithology classification in reservoir wells is usually performed manually which makes it expensive and labor-extensive. Also, it can be affected by user bias. Machine learning algorithms proved to be a suitable alternative for automatic lithology identification from well log and coring data. While the previous studies predominantly focused on predicting lithology from well logs using shallow networks, recent attention has shifted toward whole core images using deep learning approaches. In this study, a convolutional neural network is implemented to classify whole core images into three classes of limestone, sandstone, and shale. The integration of image augmentation techniques and data quality control noticeably enhanced prediction accuracy compared to prior published researches. By utilizing different architectures (ResNeXt-50, ResNet-50, and a CNN-based model) and fine-tunning hyperparameters, the ResNeXt-50 model achieved an impressive accuracy of 97.65% on unseen data for the classification of whole core images from 28 wells in southern Australia.