Predicting Grain Textures from Microscale Strain Fields of Polycrystalline Metals using Fully Convolutional Networks

Determining the microscale strain fields from the heterogeneous grain texture of polycrystalline metals is a well-established forward problem in computational mechanics. In contrast, establishing the microstructural characteristics of the metal from the measured microscale strain fields is a challenging inverse problem, which has significant implications in grain texture engineering and microstructural design and optimization. Here, we augment a fully convolutional network (FCN) with modern architectural elements, including skip connections, residual functions, and attention mechanisms, to predict the heterogeneous Euler angles representing the undeformed grain texture (output) from localized 2D total strain patterns (input) of polycrystalline metals. We train the FCN on limited data generated by crystal plasticity finite element method (CPFEM) for representative volume elements (RVEs) with similar grain shape, texture, and size distributions at a single macroscopic strain state, and show that the FCN is capable of accurate predictions far beyond its training domain – across a wide range of grain shapes, textures and size distributions, and across different macroscopic strain states. Results further show that the FCN can be adapted to new material response (stress-strain) curves with orders-of-magnitude reduction in training data through transfer learning.