Second EAGE Subsurface Intelligence Workshop

Many data formats for processed seismic data lack the availability of having a standard placeholder for indicating the processing domain for the given seismic file, and this can lead to having seismic files with missing or incorrect processing domains. A geoscientist may experience difficulties in locating required seismic stacks, in an area of interest, with unavailable indexed labels for processing domains.

To overcome this problem, we propose a data-driven approach to automate labeling and classifying stacked seismic data based on their processing domains (depth or time) by applying image processing techniques and train a deep learning model using convolutional neural networks (CNN). The proposed approach shows promising results on automatically labeling thousands of legacy and newly uploaded stacked seismic files in a corporate database with up to 90% accuracy level. This auto-labeling mechanism is based on a confidence level as a percentage and flagged as an auto-generated meta-data to differentiate between human and machine annotated data.

Earth Science Analytics presents an AI/ML driven workflow for the finding and exploitation of hydrocarbons that not only facilitates the disappearance of data silos (digitalization journey) but acts as a much needed exploration accelerator (data analytics and decision making journey). The proposed methodology empowers all geoscientists alike to easily leverage the potential of ML to find and exploit hydrocarbons, and when combined with established explorations methods and domain specialists forms a formidable tool. We will give relevant examples that lend credence to the application of ML to finding hydrocarbons close to infrastructure.

This paper present a hybrid Machine Learning and seismic structural attributes approach to extract detailed major and minor discontinuities from seismic data from a complex fault system that present a challenge using conventional interpretation techniques. Fault interpretation is image segmentation problem and we thus adopted U-net encoder-decoder architecture as a first step in this hybrid workflow, it is well suited for seismic discontinuities. Image segmentation can not only figure out whether a particular feature such as faults exist but can also create a mask showing where in the volume those features exist.

Advances in an integrated, physics-driven approach to reservoir characterization can be applied to enhance efficiencies in the development of unconventional reservoirs. The optimized production of unconventional reservoirs is dependent on lateral wellbore placement and completion practices. A high-resolution image-derived geomechanical model and Hayman factor-driven thin-bed analysis can provide enhanced characterization of unconventional reservoirs. Dipole sonic-derived mechanical property outputs do not adequately resolve thin beds and therefore do not provide a representative characterization of the reservoir for stimulation modelling and subsequent completion optimization. A novel workflow utilizing inversion modelling and regression analysis has been proposed to address this challenge of under-characterization of unconventional reservoirs. The workflow, applied to the Wolfcamp Formation in this example, integrates high-resolution borehole image data, and geo-mechanical logs. Ultimately, the workflow aims to support optimized lateral target selection through an enhanced understanding of the impact of the high-resolution rock properties on reservoir quality and drilling & completion performance. By integrating the results of this new workflow with other available downhole data and calibration to core, it is considered that the existing data resolution blind spot encountered with characterizing many unconventional reservoirs can be significantly reduced in a cost-effective manner.

Simultaneous sources aims at reducing the acquisition time of seismic surveys by either firing multiple sources at the same time, or closely after one another. This leads to entangled seismic data that has to be untangled in order to be meaningful for processing and imaging purposes. Untangling the seismic data amounts to solving an underdetermined linear inverse problem. The blending operator introduces coherent overlapping seismic events in the source gathers, but in the receiver gathers the entangled data will appear as noise. To obtain meaningful seismic data, we have to add regularization to filter this noise. In this work we propose the use of a self-supervised denoiser that filters the blending noise, that is cleverly coupled with the blending operator. Our algorithm thereby combines the physics of the blending process through the blending operator with a highly effective self-supervised denoiser. The denoiser requires no ground-truth labels or pre-training on synthetic data, which is a huge benefit over other machine learning algorithms.

Multi-channel Analysis of Surface Waves (MASW) is a seismic method employed to obtain useful information about the shear-wave velocities of the subsurface. A fundamental step in the methodology is the extraction of dispersion curves from dispersion spectra obtained after applying specific processing algorithms; to some extent, this extraction can be automated. However, it still requires extensive quality control, which can be time-demanding in large dataset scenarios. We present a novel approach that leverages deep learning to automatically identify a direct mapping between seismic shot gathers at the associated dispersion curves. Given a site of interest, a set of 1D velocity models is created using prior knowledge of the local geology; pairs of seismic shot gathers and Rayleigh-wave phase dispersion curves are then numerically modeled and used to train a simplified residual network. The proposed approach is shown to achieve satisfactory predictions of dispersion curves on a synthetic test dataset and is ultimately deployed on a field dataset. The predicted dispersion curves are finally inverted, and the resulting shear-wave velocity model is plausible and consistent with prior geological knowledge of the area.

Biostratigraphic analysis is one of the important tasks in determining the paleodepositional environment of subsurface rocks. This task is traditionally done by a biostratigrapher and involves recording microfossil observations from thin sections and interpreting these observations. Because the collected data can contain tens of microfossil species, an expert is needed to perform the correct interpretation. This work applies machine learning, specifically, ensemble decision trees, in predicting the paleodepositional environment from microfossil observations. The methodology is applied on data collected from forty-four wells. Results showed that data cleaning and data conditioning are essential steps for constructing a reliable model that can be used in understanding the depositional environment. Ensemble decision trees strike a good balance between obtaining good accuracy and allowing explainability. Finally, data imbalance is a problem in real-world datasets that must be addressed to produce better accuracy. Overall, machine learning allows for the integration of biostratigraphic data with other geologic observations, e.g., rock texture, to produce a consistent interpretation. A larger scale automated pipeline can be applied on the database to produce predictions as new data is inputted.

Seismic interpretation automation has been on the rise for the last few years: new emerging technologies, especially deep neural networks, have been successfully applied for detection numerous types of structural and geological patterns.

Unfortunately, the quality of such models directly depends on the quality and amount of training data. While historic and archive fields may provide enough material for prototypes and hypothesis testing, it is nowhere near the needed amount of data for truly production-ready algorithms.

To alleviate this problem, we’ve developed a synthetic generator, targeted specifically at model training. Despite being based on trivial assumptions and physical models, it uses a number of techniques to improve variation of created images, while keeping them look alike to real seismic surveys. With its help, we’ve been able to enhance the performance of our fault and horizon tracking models, as well as to tackle the seismic acoustic inversion task.

No summary provided.