Generative Models for Atomistic Grain Boundaries with Score-based Denoiser Model

The structure of materials interfaces, such as grain boundaries or epitaxial matching, is known to play an important role in macroscopic properties of materials. Conventionally, the computational sampling of interface structures depends on sampling atomistic structures (e.g., by Monte Carlo) and minimising the energy of the system (e.g., using DFT), which are expensive to perform. Recently, generative machine learning has shown excellent ability to perform “inpainting”, or interpolate between parts of an input given a specific context. In this talk, we show how a score-based generative model can be used to sample the structure of grain boundaries while using a single data point as training data. We will show multiple representative examples of interface structures of both 2D and 3D sampled by the denoiser model and make a comparison between them and the structure diagrams from previously reported experimental data. Finally, we explain the relationship between the generative approach and the energy calculations using information entropy as a surrogate metric for the scoring function. In the future, this approach can accelerate the sampling of complex interfaces and structures.