Improved Mesoscale Grain Boundary Energy Models for Silicon Carbide from Atomistic-Informed Regression

Author(s):

Alexander S. Hauck, Matthew Guziewski, and Efrain Hernandez-Rivera

Abstract:

Grain boundaries have significant influence over material properties such as strength, toughness, fatigue, creep resistance, and thermal conductivity. This is magnified within extreme environments as temperature and pressure control microstructure or radiation impacts local chemistry. Modeling techniques of grain boundary structures have been well explored but are often limited to atomistic techniques such as molecular dynamics. To fully understand material behavior, information from microscale techniques should be passed to macroscale representations of the grain boundaries. This enables their use in higher-scale modeling such as kinetic Monte Carlo to support processing studies such as grain growth with increased accuracy compared to the currently used Read–Shockley model. Here, decision-tree-based regression models are trained using a dataset of 332 [1 0 0], [1 1 0], and [1 1 1] tilt and twist grain boundaries generated in β-SiC. Models are generated and tuned to achieve mean absolute percentage errors below 10.0% with reasonably high coefficients of determination. These models are transformed from Python to ONNX, which enables their use in the open-source C++ kinetic Monte Carlo software SPPARKS.