Neural network for 3D inertial confinement fusion shell reconstruction from single radiographs

Area of Science:

• Plasma Physics
• Fusion Energy Research
• Advanced Imaging Techniques

Background:

• X-ray radiography is crucial for diagnosing inertial confinement fusion (ICF) implosion dynamics.
• Traditional 3D reconstruction methods require multiple views, which are often limited in ICF experiments due to space and cost constraints.
• Convolutional neural networks (CNNs) show promise for 3D reconstruction from limited 2D data.

Purpose of the Study:

• To develop and validate a CNN for reconstructing 3D ICF spherical shells from single X-ray radiographs.
• To investigate the impact of illumination models and preprocessing techniques on reconstruction accuracy.
• To address the challenge of limited 3D supervision data in ICF experiments.

Main Methods:

• A CNN model was designed to reconstruct 3D ICF shells from single 2D radiographs.
• Synthetic radiographs generated from simulations were used for training the CNN to overcome the lack of 3D experimental supervision.
• The sensitivity of the reconstruction to different illumination models and pseudo-flatfielding preprocessing was analyzed.

Main Results:

• The proposed CNN successfully reconstructs 3D ICF spherical shells from single radiographs.
• Training with synthetic data enables accurate reconstruction of experimental data with similar characteristics.
• The CNN effectively reconstructs shells exhibiting low mode asymmetries.
• Reconstruction quality is influenced by illumination models and preprocessing techniques.

Conclusions:

• CNNs offer a viable solution for 3D ICF shell reconstruction from limited radiographic data.
• The use of simulated data for training is a practical approach to address supervision limitations.
• The developed method has the potential to enhance the analysis of ICF implosion dynamics, particularly for asymmetric shells.