Abstract Scope |
Deep learning methods often fail to transfer to settings outside the training domain. A common tactic is to leverage the information contained in similar or simulated data, and transfer this knowledge to a target task (e.g. experimental predictions of material property). We consider the task of predicting 3D descriptors of the strain and stress fields of a microstructure from 2D micrographs using a dataset of 36 synthetic 3D equiaxed polycrystalline microstructures simulated using dream.3D under uniaxial tensile deformation. This dataset is divided into six FCC textures, and we attempt to generalize our predictions from one set of microstructures and textures to another set without training labels. Compared to our baselines, our models reduce the mean squared error of our property predictions by up to 50% on textures outside of the training set, and by up to 70% on microstructures of the same texture, but not in the training set. |