Machine learning for detection of 3D features using sparse x-ray tomographic reconstruction.

Area of Science:

• Physics
• Engineering
• Computer Science

Background:

• One and two dimensional models inadequately explain inertial confinement fusion (ICF) experimental results.
• Three dimensional (3D) effects, arising from shell defects, fill tubes, and double shell target joints, are significant.
• X-ray radiography is crucial for capturing internal structures, but limited projections hinder 3D reconstruction.

Purpose of the Study:

• To develop and apply advanced machine learning techniques for 3D reconstruction of ICF implosions.
• To address the ill-posed inverse problem of reconstructing 3D objects from sparse x-ray radiographic views.
• To track key 3D features of ICF capsules, including ablator, inner shell, and hemisphere joints.

Main Methods:

• Utilized multiple convolutional neural networks (CNNs) for generating 3D representations from experimental ICF data.
• Employed deep supervision techniques to train neural networks for high-resolution 3D reconstructions.
• Leveraged prior information within neural networks to solve the ill-posed 3D reconstruction problem.

Main Results:

• Generated diverse 3D representations of ICF implosions using various CNNs.
• Successfully reconstructed and tracked 3D features like the ablator, inner shell, and joint structures.
• Demonstrated the capability of machine learning to handle sparse-view x-ray radiography.

Conclusions:

• Machine learning, particularly CNNs, is a powerful tool for 3D reconstruction in ICF research.
• The developed methods show promise for improving 3D analysis in x-ray radiography beyond ICF.
• Addressing 3D effects is critical for accurate modeling and understanding of ICF experiments.