Time: 11:30 a.m.
Place: DPG Spring Meeting, Dresden, Germany
Holger Dette, Mathematik III, Ruhr-Universität Bochum, Bochum, Germany
High-throughput numerical simulations as well as experiments allow a systematic variation of individual parameters, such as e.g.composition, and the coverage of a broad range of these parameters. Nevertheless, the majority of properties that are available today are so-called intrinsic properties like stability, stiffness, or band gaps. Extrinsic properties, like interface distribution functions, energies or mobilities, are less frequently available. As a matter of fact, because they depend on several variables at the same time, such properties form multidimensional databases of their own for one particular material. For such cases we want to promote an efficient sampling of the variable space that is based on design of experiment principles. As a suitable example we pick grain boundary energies, which depend on five geometric degrees of freedom. We introduce two methods to improve the current state of the art. Based on an existing energy model the location and number of the energy minima along which the hierarchical sampling takes place is predicted from existing data points without any a-priori knowledge, using a predictor function. Furthermore we show that in many cases, it is more efficient to use the above mentioned sequential sampling, rather than sampling all observations homogeneously in one batch.