Machine Learning-based 3D Reconstruction of ICF Capsules via Self-Emission Images

In Inertial Confinement Fusion (ICF) experiments, pusher shots are composed of shell/ablator and filled with a mix of D-T and typically a noble gas for diagnostic purposes. Asymmetries during capsule compression can inaccurately determine parameters like neutron yield with 1D and 2D models. Even with experimental techniques, the lack of available multiple views fails to encompass all the 3-d effects of the implosion, necessitating theoretical assumptions. A key quantity for judging the accuracy of these models is amount of compression of the capsule, which can be gained from self-emission images. We propose a synthetic data for 3D reconstruction through Machine Learning techniques for the purpose of determining the convergence ratio for comparison with the simulation codes. Previous work done by Wolfe et al. used convolutional neural networks (CNN) to produce 3D reconstructions of ICF models using simulated back-lighter images [1]. We adapted this work, we generated plasma profiles with Python, processed by Prism’s SPECT-3D [2] for accurate simulated self-emission images. This approach resulted in a training set consisting of thousands of profile-image pairs. Using our dataset, we will train a CNN to generate 3D reconstructions from self-emission images to improve feature distinction in imploding ICF capsules.

References:

[1] Wolfe, Bradley T. et al. “Machine Learning for Detection of 3D Features Using Sparse X-Ray Tomographic Reconstruction.” Review of scientific instruments 94.2 (2023): 023504. Print.

[2] . MacFarlane, J.J. et al. “SPECT3D – A Multi-Dimensional Collisional-Radiative Code for Generating Diagnostic Signatures Based on Hydrodynamics and PIC Simulation Output.” High energy density physics 3.1 (2007): 181–190. Print.