Dresden 2020 – wissenschaftliches Programm

Die DPG-Frühjahrstagung in Dresden musste abgesagt werden! Lesen Sie mehr ...

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

MM: Fachverband Metall- und Materialphysik

MM 30: Poster Session II

MM 30.38: Poster

Dienstag, 17. März 2020, 18:15–20:00, P4

Machine-learning Driven Global Optimization of Atomic Surface Structures — •Sami Kaappa and Karsten Wedel Jacobsen — CAMD, Department of Physics, Technical University of Denmark, 2800 Kongens Lyngby, Denmark

Efficient global optimization of atomic structures is a long-pursued objective in material sciences since the required number of ab initio calculations is usually computationally infeasible. In this work, a machine-learning guided approach is utilized to model the potential energy hypersurface (PES) as a function of atomic coordinates, and the surrogate model is used to intelligently sample the search space for the global minimum in order to reduce the number of expensive DFT calculations. In the method, the translational and rotational symmetries as well as symmetries with respect to interchanging positions of alike atoms are naturally inherited by a global fingerprint, and both energy and force information of DFT calculations are used in the Gaussian process machinery to model the PES. We will present both the performance of the method in comparison to previously reported, similar procedures, and predicted global minima for certain atomic structures where the optimal atomic configuration is not trivial. Although DFT calculations are carried out here, we note that higher-level theories can be used as well to probe energies and forces of single structures, to be offered as training data for the model.