It is possible to construct a grain boundary diagrams as they relate to temperature and bulk composition that are akin to phase diagrams. However, as useful as these "complexion diagrams" are a grain boundary has five macroscopic, or crystallographic, degrees of freedom, the solutions of which a US team describes in this journal are "mission impossible". Constructing a complexation diagram as a function of five degrees of freedom cannot be done either experimentally or through modeling.

Now, Jian Luo and colleagues at the University of California San Diego, La Jolla, have found a way to surmount this fivefold obstacle by turning to deep learning to train a genetic algorithm to construct the diagram for them. They begin with isobaric semi-grand-canonical ensemble hybrid Monte Carlo and molecular dynamics (hybrid MC/MD) simulations and push the outputs to genetic algorithm and deep neural network models to approach the impossible. [Luo, J. et al. Mater Today (2020); DOI: 10.1016/j.mattod.2020.03.004]

Grain boundaries are technically defects, the interfaces between different regions of crystallinity in a polycrystalline substance. They are coincident with deviations from bulk properties of thermal and electrical conductivity and other characteristics. They also act as areas within the substance where corrosion or a phase change might begin leading to weakening within the material.

Fundamentally, the team shows that deep neural network operates approximately 100 million times faster than any atomistic simulation and so can construct property diagrams much faster for millions of distinctly different grain boundaries. This could represent a paradigm shift in studying the performance properties of polycrystalline materials. Given that most engineered and natural materials are polycrystalline this could a useful turning point in such research.

The team adds that their approach offers excellent prediction accuracies not only for symmetric-tilt and twist grain boundaries but also for asymmetric-tilt and mixed tilt-twist grain boundaries. They point out that the latter are more complex and not nearly so well understood as the former, but they are ubiquitous and often represent an insurmountable barrier to calculating the properties of real polycrystals. The data-driven prediction of grain boundary properties as function of temperature, bulk composition, and five crystallographic degrees of freedom opens up an entirely new way to work with material properties, the research suggests.

Click here to read the article in the journal.