Efficient Prediction of Grain Boundary Energies from Atomistic Simulations via Sequential Design

Data based materials science is the new promise to accelerate materials design. Especially in computational materials science, data generation can easily be automatized. Usually, the focus is on processing and evaluating the data to derive rules or to discover new materials, while less attention is being paid on the strategy to generate the data. In this work, we show that by a sequential design of experiment scheme, the process of generating and learning from the data can be combined to discover the relevant sections of the parameter space. Our example is the energy of grain boundaries as a function of their geometric degrees of freedom, calculated via atomistic simulations. The sampling of this grain boundary energy space, or even subspaces of it, represents a challenge due to the presence of deep cusps of the energy, which are located at irregular intervals of the geometric parameters. Existing approaches to sample grain boundary energy subspaces therefore either need a huge amount of datapoints or a priori knowledge of the positions of these cusps. We combine statistical methods with atomistic simulations and a sequential sampling technique and compare this strategy to a regular sampling technique. We thereby demonstrate that this sequential design is able to sample a subspace with a minimal amount of points while finding unknown cusps automatically.