Phd defense of Ahmed BEN SAAD

Ahmed BEN SAAD

Feeding CNNs with prior knowledge

Abstract

In the Oil & Gas industry, increasing production, recovery rates, and the lifespan of oil fields constitutes a major technological challenge. Borehole imaging tools, which provide high-resolution images of the borehole wall based on physical property contrasts, are essential for characterizing geological formations and ensuring borehole stability. These tools help detect sedimentary layers, faults, and fractures, which are critical for understanding the subsurface geological features. Despite the availability of various imaging tools, traditional interpretation methods require significant expertise and are time-consuming, necessitating advancements to meet modern demands for efficiency and accuracy.

This thesis aims to integrate domain knowledge into machine learning methods to enhance borehole image analysis. We propose a novel data-driven object detection framework that leverages synthetic data generation through neural style transfer to mimic real borehole image textures. This synthetic data, combined with real data, improves model training, enabling the detection and parameterization of dips as a regression problem rather than a segmentation task. This approach significantly enhances detection accuracy and computational efficiency, particularly for sinusoidal and closed dips.

First, we look onto a way to expand the available datasets by inpainting images obtained from pad-based tools using Generative Adversarial Networks (GANs), enhancing texture inpainting techniques to utilize incomplete datasets.

Next, we explore semi-supervised learning through Contrastive Learning, enabling neural networks to learn robust image representations from large, unlabeled datasets, subsequently fine-tuned with smaller labeled datasets.

After that, we focus on feeding convolutional neural networks (CNNs) with prior knowledge in the form of exogenous data for breakout segmentation as a first type of prior knowledge. By incorporating additional data sources, such as ultrasonic or thermal measurements, the model's performance is enhanced, leveraging the diverse information to improve detection accuracy and robustness.

Than we look onto the second way of prior knowledge integration into CNNs by modifying the loss function for object detection. A weighting term based on the object's area size is integrated into the loss function, which helps the object detector learn features for detecting objects of all sizes. This approach utilizes the rich details and textures of large objects to improve detection performance across all object sizes, resulting in significant improvements in mean Average Precision (mAP) and recall rates.

Finally, we shift our focus on modifying the object detection framework and using suited parameterization for dip picking. The proposed framework formulates dip detection as an object detection problem, where the parameters of the dips are directly regressed by the model. By using synthetic data generation through neural style transfer to mimic real borehole image textures, the framework enhances model training and detection accuracy for both sinusoidal and closed dips, demonstrating effective generalization from synthetic to real data. This chapter highlights the benefits of framework modification and appropriate parameterization in enhancing the detection of complex geological features.

Jury

• Rodrigo Verschae
• Pascal Monasse
• Qiang Qiu
• Hugues Talbot
• Alasdair Newson

Supervision

• Gabriele Facciolo
• Enric Meinhardt
• Axel Davy