PythonImage Log AnalysisPyTorch2024

Leveraging Transformers for End-To-End Bed and Fractures Identification, Classification, Localization and Parameter Estimation using Image Logs for Subsurface Geological Analysis

Event Year

2024

BibTeX / Citation

@inproceedings{Seeba_2025, series={ICHEP2024}, title={End-to-end tau reconstruction and identification using transformers}, url={http://dx.doi.org/10.22323/1.476.1036}, DOI={10.22323/1.476.1036}, booktitle={Proceedings of 42nd International Conference on High Energy Physics — PoS(ICHEP2024)}, publisher={Sissa Medialab}, author={Seeba, Norman and Tani, Laurits and Vanaveski, Hardi and Pata, Joosep and Lange, Torben}, year={2025}, month=Jan, pages={1036}, collection={ICHEP2024} }

Visit Official Page

Manuscript Preview

Open Fullscreen

Executive Summary

Accurate identification, classification, localization, and parameter estimation of subsurface geological features, such as beddings and fractures, are critical for understanding reservoir characteristics. Traditionally, such analyses rely on manual interpretation of borehole image logs, which is labor-intensive and time-consuming. Recent advancements have introduced semi-automated, mask-based segmentation approaches with the intention of assisting manual picking, but these methodologies require manual mask creation and post-processing, limiting their effectiveness in achieving a truly automated workflow. To address these limitations, we propose an end-to-end methodology using a transformer-based model, GeoBFDT, to automatically identify, classify, localize and estimate the parameters of geological features in borehole image logs. The GeoBFDT model eliminates the need for manual mask creation and post-processing by directly identifying whether a feature exists, classifying it as a bedding or fracture, localizing the feature, and estimating its parameters (e.g., dip, azimuth). This streamlined workflow enables efficient and accurate subsurface feature analysis. Several custom evaluation metrics were developed to rigorously evaluate the model's performance. The adaptability of GeoBFDT was validated using data from multiple well orientations (vertical and horizontal) and various imaging tools (FMI, CMI, and ThruBit). While end-to- end methodologies like GeoBFDT are inherently challenging to optimize due to their complexity, the benefits of eliminating manual intervention and enabling fully automated workflows make them highly promising. This study underscores the potential of such approaches for advancing geological feature analysis, particularly in industry, where data is abundant. We envision that the proposed methodology could pave the way for enhanced subsurface feature detection in a truly end-to-end manner, surpassing semi-automated approaches that still rely on manual interpretation.

Related Works

Automated end-to-end fracture identification, classification, localization, and parameter estimation for enabling rapid risk management and CO₂ storage optimization in CCUS applications

The success of Carbon Capture, Utilization, and Storage (CCUS) projects heavily depends on understanding subsurface fluid flow behaviour particularly through fracture networks. Fractures play a dual role in such operations: they can enhance reservoir injectivity and storage capacity by providing pathways for CO₂ injection, but they also pose risks by potentially compromising caprock integrity, increasing the risk of structural storage failure thereby enabling CO₂ leakage. Accurate fracture detection and characterization is essential for optimizing injection strategies, ensuring effective containment, and mitigating environmental risks. Fractures influence critical processes such as trapping mechanisms and pressure distribution within the reservoir. Furthermore, understanding their orientation and density is vital for designing safe and efficient CO₂ injection operations. These factors highlight the importance of robust, non-bias, automated, and scalable fracture detection methods. Traditional fracture identification methods rely heavily on manual interpretation, which is time-intensive, subjective, and challenging to scale for large fields with several wells. This study proposes a scalable automated methodology employing advanced deep-learning techniques to detect fractures from borehole imaging tools such as FMI, CMI, and ThruBit logs. The proposed approach uses detection transformers which eliminates the need for manual mask creation and post-processing steps by adopting an end-to-end framework, which not only identifies the presence of fractures but also estimates their orientation and density. Custom evaluation metrics were developed to measure the model's performance (in comparison with expert’s fracture analysis) in handling diverse geological and well conditions, including vertical and horizontal well orientations. The automated workflow facilitates speedy assessment of fracture networks which in turn can offer speedy actionable insights for CO₂ injection optimization, caprock stability assessment, and risk management. The model demonstrated an interpretation speed of less than one minute per 2 meters, with an ~80% F1 score (6 cm depth error margin), ~91% accuracy in dip picking (3° error margin), and ~93% accuracy in dip estimation (15° dip margin). By utilizing the proposed automated fracture detection model based on transformers, CCUS project planning and designing can be accelerated. Furthermore, integrating MLOps into the workflow ensures the scalability, maintainability, and adaptability of these models for practical deployment. While this methodology is tailored to CCUS, its versatility extends to a much wider range of applications, including geothermal energy, mining, and other subsurface characterization domains.

Automated Workflow for Detecting Fractures and Bedding Planes

In geoscience, manually picking fractures and bedding from Formation Micro-Imager (FMI) logs is labour-intensive, requiring substantial time and effort. This study addresses the critical need for an automated solution to streamline this picking and identification process, emphasizing the significance of expediting reservoir characterization and decision-making in the petroleum industry. The model presented here aims to provide a quick turnaround of results delivery while significantly reducing the manual workload by automating the identification of fractures and beddings. The study leverages a comprehensive dataset comprising FMI logs from 14 vertical wells in Oman. The methodology adopts an advanced Fracture Detection Model based on the Detection Transformer architecture, customized to the unique requirements of the study. Precision, Recall, and F1-Score are key evaluation metrics computed through a tailored confusion matrix and depth thresholding for nuanced predictions. Our results are validated through different depth thresholds for fractures and beds. Sensitivity tests reveals that 8cm threshold generates a recall ~85% in comparison to 4cm threshold of ~75%. Visual analyses, performance metrics, and comparative plots underscore the model’s proficiency in accurately identifying subsurface features. This Fracture Detection Model, validated through testing on several wells, stands out as an efficient and accurate automated tool for reservoir characterization.