1 Neural Network-Based Defect Detection On Tomography Images For Quality Control of Inertial Fusion Capsules

David Orozco, Brian Sammuli

Abstract:

Recent headlines of record-breaking fusion energy
yields using inertial confinement have led to a growing interest in
the metrology processes used to inspect the high-density carbon
(HDC) capsules that have achieved these yields. To ensure the
highest quality HDC shells, the tomographic reconstructions
are reviewed for defects to satisfy a certain standard of qual-
ity. Historically, a human expert would visually analyze HDC
tomography data searching for defects, which is very time-
consuming and labor-intensive. While classical computer vision
methods have also been employed, machine-learning methods
have demonstrated the ability to analyze vast amounts of data
at high speeds and more robustly identify patterns and anoma-
lies that may indicate the presence of previously undetectable
defects. Presented here is our novel process, ShellNet, that has
been developed to automate this defect detection task on HDC
shells using convolutional neural networks; the challenges and
limitations of this approach are also discussed. A machine-
learning model trained on thousands of expert-labeled defects and
tomography images, utilizing a custom labeling tool, is now being
used in production. Case studies are presented demonstrating the
effectiveness of machine learning in this application, as well as
outlines for future research in this area. Overall, machine learn-
ing has been shown to be an effective method in enhancing the
quality-based selection process of HDC shells, even exceeding the
recall performance of human labeled data. These improvements
promise to lead to overall improved reliability of HDC materials
in inertial fusion technology.