Sharon Finlay
Technology & Innovation Advisor
North Oil Company
With nearly three decades of comprehensive experience in petrophysics, Sharon has developed deep expertise across all key elements of Petrophysics across the development cycle, appraisal, brownfield projects, core analysis, reservoir optimization and EOR. More recently, she has focused on advancing technology and innovation, driving the integration of cutting-edge tools and methodologies to enhance reservoir characterization, asset performance and asset development., and enjoys leveraging deep expertise to simplify complex challenges.
Participates in
TECHNICAL PROGRAMME | Energy Technologies
The Energy Transition: The Role of Digitalisation, AI, and Cybersecurity
Forum 23 | Technical Programme Hall 4
29
April
14:30
16:00
UTC+3
Introduction:
Fractures appear as sinusoidal curves on borehole images (BHI’s) in horizontal wells and provide key insights into reservoir performance, they can indicate highly productive zones or highlight potential concerns for waterflood shortcuts. Traditionally interpretations are manual and is inherently time-consuming, laborious, and subject to interpreter bias and variability. Automating this process could result in significant cost savings allowing geological expertise to be focused elsewhere. However, computer vision algorithms face substantial challenges due to complexities in data quality, particularly in a Logging While Drilling (LWD) environment. LWD Images often experience distortions, noise, missing data patches, and artifacts, exacerbated by downhole shocks and vibrations that tend to increase with drilling depth. These imperfections can obscure true fractures or create patterns that mimic the sinusoidal shape of fractures, leading to high rates of false positives in automated detection. We aimed to develop an AI-assisted interpretation method to address many of these challenges.
Method & Applications:
Recognizing the impact of data quality on fracture interpretation reliability, we address the subjective and meticulous task of image quality labeling. We developed a deep learning-based workflow using an Efficient Net CNN, fine-tuned with borehole images classified as 'good', 'fair', and 'poor'. The resulting model achieves an F1 score of 80%, providing consistent and rapid identification of reliable data sections for interpretation whether manual or automated.
This work also presents a dual approach to improve the accuracy and efficiency of fracture analysis from borehole images. First, we define a set of image quality criteria and design corresponding scoring measures. Emphasis is placed on image quality features, as artifacts and missing data can significantly impact detection either by obscuring actual fractures or introducing false ones. These quality features are leveraged to minimize false positives while preserving true fracture detections, initially identified using a pre-existing sinusoid detection algorithm. This filtering process identifies key discriminative features that help distinguish genuine natural fractures from false detections. We managed to reduce false detections by 98% while keeping 94% of true fractures.
In fields developed with ERD wells Borehole Images are a critical element to understand reservoir performance. Human interpretations can be highly subjective, AI-Assisted methods significantly improves the repeatability and consistency required in field wide reservoir modeling and understanding reservoir performance. In addition, AI-Assisted methods significantly reduce interpretation times from days to less than an hour.
Conclusions:
In conclusion, AI-Assisted methods have been used to significantly improve the timelines and robustness of Borehole Image interpretation.
Fractures appear as sinusoidal curves on borehole images (BHI’s) in horizontal wells and provide key insights into reservoir performance, they can indicate highly productive zones or highlight potential concerns for waterflood shortcuts. Traditionally interpretations are manual and is inherently time-consuming, laborious, and subject to interpreter bias and variability. Automating this process could result in significant cost savings allowing geological expertise to be focused elsewhere. However, computer vision algorithms face substantial challenges due to complexities in data quality, particularly in a Logging While Drilling (LWD) environment. LWD Images often experience distortions, noise, missing data patches, and artifacts, exacerbated by downhole shocks and vibrations that tend to increase with drilling depth. These imperfections can obscure true fractures or create patterns that mimic the sinusoidal shape of fractures, leading to high rates of false positives in automated detection. We aimed to develop an AI-assisted interpretation method to address many of these challenges.
Method & Applications:
Recognizing the impact of data quality on fracture interpretation reliability, we address the subjective and meticulous task of image quality labeling. We developed a deep learning-based workflow using an Efficient Net CNN, fine-tuned with borehole images classified as 'good', 'fair', and 'poor'. The resulting model achieves an F1 score of 80%, providing consistent and rapid identification of reliable data sections for interpretation whether manual or automated.
This work also presents a dual approach to improve the accuracy and efficiency of fracture analysis from borehole images. First, we define a set of image quality criteria and design corresponding scoring measures. Emphasis is placed on image quality features, as artifacts and missing data can significantly impact detection either by obscuring actual fractures or introducing false ones. These quality features are leveraged to minimize false positives while preserving true fracture detections, initially identified using a pre-existing sinusoid detection algorithm. This filtering process identifies key discriminative features that help distinguish genuine natural fractures from false detections. We managed to reduce false detections by 98% while keeping 94% of true fractures.
In fields developed with ERD wells Borehole Images are a critical element to understand reservoir performance. Human interpretations can be highly subjective, AI-Assisted methods significantly improves the repeatability and consistency required in field wide reservoir modeling and understanding reservoir performance. In addition, AI-Assisted methods significantly reduce interpretation times from days to less than an hour.
Conclusions:
In conclusion, AI-Assisted methods have been used to significantly improve the timelines and robustness of Borehole Image interpretation.
TECHNICAL PROGRAMME | Primary Energy Supply
Advances in Geoscience
Forum 05 | Digital Poster Plaza 1
30
April
10:00
12:00
UTC+3
Introduction:
Fractures appear as sinusoidal curves on borehole images (BHI’s) in horizontal wells and provide key insights into reservoir performance, they can indicate highly productive zones or highlight potential concerns for waterflood shortcuts. Traditionally interpretations are manual and is inherently time-consuming, laborious, and subject to interpreter bias and variability. Automating this process could result in significant cost savings allowing geological expertise to be focused elsewhere. However, computer vision algorithms face substantial challenges due to complexities in data quality, particularly in a Logging While Drilling (LWD) environment. LWD Images often experience distortions, noise, missing data patches, and artifacts, exacerbated by downhole shocks and vibrations that tend to increase with drilling depth. These imperfections can obscure true fractures or create patterns that mimic the sinusoidal shape of fractures, leading to high rates of false positives in automated detection. We aimed to develop an AI-assisted interpretation method to address many of these challenges.
Method & Applications:
Recognizing the impact of data quality on fracture interpretation reliability, we address the subjective and meticulous task of image quality labeling. We developed a deep learning-based workflow using an Efficient Net CNN, fine-tuned with borehole images classified as 'good', 'fair', and 'poor'. The resulting model achieves an F1 score of 80%, providing consistent and rapid identification of reliable data sections for interpretation whether manual or automated.
This work also presents a dual approach to improve the accuracy and efficiency of fracture analysis from borehole images. First, we define a set of image quality criteria and design corresponding scoring measures. Emphasis is placed on image quality features, as artifacts and missing data can significantly impact detection either by obscuring actual fractures or introducing false ones. These quality features are leveraged to minimize false positives while preserving true fracture detections, initially identified using a pre-existing sinusoid detection algorithm. This filtering process identifies key discriminative features that help distinguish genuine natural fractures from false detections. We managed to reduce false detections by 98% while keeping 94% of true fractures.
In fields developed with ERD wells Borehole Images are a critical element to understand reservoir performance. Human interpretations can be highly subjective, AI-Assisted methods significantly improves the repeatability and consistency required in field wide reservoir modeling and understanding reservoir performance. In addition, AI-Assisted methods significantly reduce interpretation times from days to less than an hour.
Conclusions:
In conclusion, AI-Assisted methods have been used to significantly improve the timelines and robustness of Borehole Image interpretation.
Fractures appear as sinusoidal curves on borehole images (BHI’s) in horizontal wells and provide key insights into reservoir performance, they can indicate highly productive zones or highlight potential concerns for waterflood shortcuts. Traditionally interpretations are manual and is inherently time-consuming, laborious, and subject to interpreter bias and variability. Automating this process could result in significant cost savings allowing geological expertise to be focused elsewhere. However, computer vision algorithms face substantial challenges due to complexities in data quality, particularly in a Logging While Drilling (LWD) environment. LWD Images often experience distortions, noise, missing data patches, and artifacts, exacerbated by downhole shocks and vibrations that tend to increase with drilling depth. These imperfections can obscure true fractures or create patterns that mimic the sinusoidal shape of fractures, leading to high rates of false positives in automated detection. We aimed to develop an AI-assisted interpretation method to address many of these challenges.
Method & Applications:
Recognizing the impact of data quality on fracture interpretation reliability, we address the subjective and meticulous task of image quality labeling. We developed a deep learning-based workflow using an Efficient Net CNN, fine-tuned with borehole images classified as 'good', 'fair', and 'poor'. The resulting model achieves an F1 score of 80%, providing consistent and rapid identification of reliable data sections for interpretation whether manual or automated.
This work also presents a dual approach to improve the accuracy and efficiency of fracture analysis from borehole images. First, we define a set of image quality criteria and design corresponding scoring measures. Emphasis is placed on image quality features, as artifacts and missing data can significantly impact detection either by obscuring actual fractures or introducing false ones. These quality features are leveraged to minimize false positives while preserving true fracture detections, initially identified using a pre-existing sinusoid detection algorithm. This filtering process identifies key discriminative features that help distinguish genuine natural fractures from false detections. We managed to reduce false detections by 98% while keeping 94% of true fractures.
In fields developed with ERD wells Borehole Images are a critical element to understand reservoir performance. Human interpretations can be highly subjective, AI-Assisted methods significantly improves the repeatability and consistency required in field wide reservoir modeling and understanding reservoir performance. In addition, AI-Assisted methods significantly reduce interpretation times from days to less than an hour.
Conclusions:
In conclusion, AI-Assisted methods have been used to significantly improve the timelines and robustness of Borehole Image interpretation.
