Comparative Study of Neural Networks for 3D Reconstruction of 2D X-Ray Radiograph

X-ray radiography acts as an important diagnostic for tracking the evolution of inertial confinement fusion targets and the growth of shell asymmetries. Traditionally, the analysis of the radiographs use information such as 2D contours to determine the modes of asymmetry. This method of analysis fails to account for 3D shell asymmetry which is present in the experiments. To account for various 3D features, we describe neural networks to reconstruct 3D models from experimental radiographs. We use these 3D models in order to determine 3D modes of asymmetry using various representations such as wavelets, Fourier modes and spherical harmonics. The evolution of 3D features of the ICF shells such as the joint between shell hemispheres has been characterized. We use multiple neural networks in order to compare the robustness and uncertainties of the results.