Machine-Learning Informed Representations for Grain Boundary Structures

Abstract

The atomic structure of grain boundaries plays a defining but poorly understood role in the properties they exhibit. Due to the complex nature of these structures, machine learning is a natural tool for extracting meaningful relationships and new physical insight. We apply a new structural representation, called the scattering transform, that uses wavelet-based convolutional neural networks to characterize the complete three-dimensional atomic structure of a grain boundary. The machine learning to predict GB energy, mobility, and shear coupling using the scattering transform representation is compared and contrasted with learning using a smooth overlap of atomic positions (SOAP) based representation. While predictions using the scattering transform are not as good as those of SOAP, other factors suggest that the scattering transform may yet play an important role in GB structure learning. These factors include the ability of the scattering transform to learn well on larger datasets, in a process similar to deep learning, as well as their ability to provide physically interpretable information about what aspects of the GB structure contribute to the learning through an inverse scattering transform.