[SY-K5] Fast and scalable prediction of local energy at grain boundaries: Machine-learning based modeling of first-principles calculations
A GB is the interface between two grains or crystallites in a polycrystalline material, and the atomic configurations and chemical bonds near GB are distinct from those of the bulk crystal. Thus, the properties of materials with GBs can greatly differ from those of a single crystal. Since it is possible to improve various properties of materials significantly by GB engineering, it is of great importance to investigate physical and chemical properties of each atoms or local regions near GBs. By virtue of the development of efficient computational techniques of large-scale density functional theory (DFT) calculations and the rapid progress of parallel computers including supercomputers, we can deal with relatively large supercells for GB models. But, it needs much computational costs to perform DFT calculations with larger supercells, and it is impossible to cover all GB models.
We proposed a new scheme based on machine learning for the efficient screening in GB engineering. A set of results obtained from DFT calculations for a small number of GB systems is used as a training data set. In our scheme, by partitioning the total energy into atomic energies using a local-energy analysis scheme, we can increase the training data set significantly. We use atomic radial distribution functions and additional structural features as atom descriptors to predict atomic energies and GB energies simultaneously using the least absolute shrinkage and selection operator (LASSO). In the test study with fcc-Al [110] symmetric tilt GBs, we could achieve enough predictive accuracy to understand energy changes at and near GBs at a glance, even if we collected training data from only ten GB systems. The present scheme can emulate time-consuming DFT calculations for large GB systems with negligible computational costs, and thus enable the fast screening of possible alternative GB systems.
We proposed a new scheme based on machine learning for the efficient screening in GB engineering. A set of results obtained from DFT calculations for a small number of GB systems is used as a training data set. In our scheme, by partitioning the total energy into atomic energies using a local-energy analysis scheme, we can increase the training data set significantly. We use atomic radial distribution functions and additional structural features as atom descriptors to predict atomic energies and GB energies simultaneously using the least absolute shrinkage and selection operator (LASSO). In the test study with fcc-Al [110] symmetric tilt GBs, we could achieve enough predictive accuracy to understand energy changes at and near GBs at a glance, even if we collected training data from only ten GB systems. The present scheme can emulate time-consuming DFT calculations for large GB systems with negligible computational costs, and thus enable the fast screening of possible alternative GB systems.