Deep Convolutional Neural Network for Tomographic Reconstruction of Strain

The neutron transmission Bragg edge method has gained interest over the past decade as a technique for providing a fast, high spatial-resolution and high sample penetration mapping of strain. Since its advent, there has been a strong desire to extend the transmission Bragg edge method to provide a full 3D tomographic reconstruction of the full strain tensor. However, it has been shown that traditional tomographic reconstruction algorithms are, in general, unable to provide a unique solution as the mathematical problem of inversion of Bragg edge data is ill-posed. The major complication of strain tomography compared to traditional scalar tomography is the directional dependence of the strain tensor. Deep Convolutional Neural Networks (DCNNs) have shown success in scalar tomographic reconstruction, in particular for cases in which scalar tomography is ill-posed (e.g. when sample exposure or rotation is limited). We show that DCNNs can be extended to perform tomographic reconstruction of strain tensors based on neutron transmission Bragg edge data. We show three training strategies and a DCNN architecture with success in reconstructing 2D strain fields. Additionally, we explore the advancements necessary to make a fully generalizable model for tomographic reconstruction of strain.