Latest In

News

Can We Characterize Atomic Environment Disorders?

Understanding how progressing local structural environments influence strength and reliability requires estimating the threshold of atomic environment disorder within materials.

Alexander McCaslin
Mar 27, 20221 Shares495 Views
Understanding how progressing local structural environments influence strength and reliability requires estimating the threshold of atomic environmentdisorder within materials. The team of researchers from Lawrence Livermore National Laboratory in Canada and the United States of America, led by James Chapman and Tim Hsu, used graph neural networks to define a physically interpretable metric for local disorder. This metric quantifies the diversity of local atomic configurations as a continuous spectrum between solid and liquid phases. This novel methodology was applied to three prototypical examples with varying degrees of disorder: (1) solid-liquid interfaces, (2) polycrystalline microstructures, and (3) grain boundaries. They demonstrated how, using elemental aluminum as an example, a paradigm can track the spatio-temporal evolution of interfaces by incorporating a mathematically defined description of the spatial boundary between order and disorder. They also demonstrated how to extract physics-preserved gradients from continuous disorder fields, which can be used to understand and predict the performance and failure of materials. Overall, the framework is easy to use and can be used to figure out how complex local atomic structure and coarse-grained material phenomena are linked.

Characterization

Understanding the relationship between atomic-level environments and long-range structural features is critical for understanding how atomic-level perturbations change the properties of a material. Quantifying phenomena like interface nucleation and growth in polycrystalline configurations requires an accurate description of how local atomic geometries evolve over time, resulting in differences in the material's long-range structure. To do that, the researchers came up with a way to describe how much disorder there is in an atomic geometry. By combining the explicit physics-preserving nature of graph representations with the learning power of machine learning, the researchers established a pipeline in which graph neural networks can learn the subtle differences in local structure between different material phases. This capability gives us a continuous metric for scoring local atomic environments based on their location in the abstract phase space. When dynamic atomistic simulations is run, this metric helps to learn about the long-term structural features of a material by automatically predicting how many grains will form.
The researchers proposed applying their methodology to a wide range of complex materials characterization problems, especially those that rely on the subtle and intricate relationship between atomic perturbations and long-range structure. Because of the explicit encoding of structural information via graph representations, our methodology can uniquely identify even subtle differences in local atomic geometries, making it ideal for predicting material phase differences. Rather than attempting to unsupervisedly classify each atomic environment, we characterise these local structures as a continuous value between two distinct material phases, allowing for an intuitive and physics-aware method of identifying even subtle structural differences both locally and over longer length scales. In theory, more than two channels can be used, allowing for the identification of more complex features within the material.

Mapping Atoms To A Continuous Field

The researchers presented the technique as a critical tool for creating more physically stable and realistic simulation models that rely on continuous field representations, such as those used in phase field and continuum models. The ability to predict properties such as stress points at interface regions by mapping the discrete atomic structure to a scalar field representation using the Stochastic Orderness Degree for Atomic Structures metric shows that our computational framework can be used to switch between representation schemes reliably. Rather than estimating what the field representation should be, the approach may give an explicit atoms-to-field mapping, allowing for more precise and physics-informed phase field models. This mapping also provides access to an accurate and interpretable approach to understanding the gradients of these fields, which is critical in modelling how continuous representation models change over time. Because the foundations of phase field and continuum models can be developed from explicit atomistic descriptions, they have an unparalleled degree of precision.

Conclusion

In their work, the researchers described the slow melting of homogenous bulk structures. The researchers next demonstrated how the combination of Stochastic Orderness Degree for Atomic Structures may be utilized to identify and describe the time-dependent evolution of different microstructural environments during grain coarsening in systems with more than 1.5 million atoms. This opens up the possibility of mapping the atomic configurations to a continuous scalar field so that we can figure out where the stress will be in the areas where grains meet.
Jump to
Latest Articles
Popular Articles