Universal and interpretable classification of atomistic structural transitions via unsupervised graph learning

Abstract

Materials processing often occurs under extreme dynamic conditions leading to a multitude of unique structural environments. These structural environments generally occur at high temperatures and/or high pressures, often under non-equilibrium conditions, which results in drastic changes in the material's structure over time. Computational techniques, such as molecular dynamics simulations, can probe the atomic regime under these extreme conditions. However, characterizing the resulting diverse atomistic structures as a material undergoes extreme changes in its structure has proved challenging due to the inherently non-linear relationship between structures as large-scale changes occur. Here, we introduce SODAS++, a universal graph neural network framework, that can accurately and intuitively quantify the atomistic structural evolution corresponding to the transition between any two arbitrary phases. We showcase SODAS++ for both solid–solid and solid–liquid transitions for systems of increasing geometric and chemical complexity, such as colloidal systems, elemental Al, rutile and amorphous TiO2, and the non-stoichiometric ternary alloy Ag26Au5Cu19. We show that SODAS++ can accurately quantify all transitions in a physically interpretable manner, showcasing the power of unsupervised graph neural network encodings for capturing the complex and non-linear pathway, a material's structure takes as it evolves.

Local Structural Features Elucidate Crystallization of Complex Structures

“…In the Supporting Information, we provide benchmarks for the performance of the order parametertrained using Steinhardt’s and Spellings’ featurizationsagainst both the standard Steinhardt Q l parameter and CN to show the improved level of structural detail that can be captured with the here-presented method. The ability to train and apply our unsupervised models without the use of high-performance or GPU resources is a significant advantage over more sophisticated machine learning methods with fewer tuned features. ,− Furthermore, using features based on existing bond-orientational order metrics allows for greater interpretability of our resulting machine-learned models.…”