Predicting phase behavior of grain boundaries with evolutionary search and machine learning

TL;DR

This paper introduces a novel computational approach combining evolutionary algorithms and machine learning to predict and analyze the complex phase behavior of grain boundaries, revealing rich polymorphism and structural transitions.

Contribution

It develops a new computational tool for efficient grain boundary structure prediction and automatic phase identification, advancing understanding of interface phase behavior.

Findings

Uncovered rich polymorphism in grain boundary structures in Cu

Identified new ground and metastable states with different atomic densities

Demonstrated multiple phases and structural transitions across misorientation ranges

Abstract

The study of grain boundary phase transitions is an emerging field until recently dominated by experiments. The major bottleneck in exploration of this phenomenon with atomistic modeling has been the lack of a robust computational tool that can predict interface structure. Here we develop a new computational tool based on evolutionary algorithms that performs efficient grand-canonical grain boundary structure search and we design a clustering analysis that automatically identifies different grain boundary phases. Its application to a model system of symmetric tilt boundaries in Cu uncovers an unexpected rich polymorphism in the grain boundary structures. We find new ground and metastable states by exploring structures with different atomic densities. Our results demonstrate that the grain boundaries within the entire misorientation range have multiple phases and exhibit structural transitions, suggesting that phase behavior of interfaces is likely a general phenomenon.