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

MM: Fachverband Metall- und Materialphysik

MM 66: Topical Session (Symposium MM): Big Data in Materials Science - Managing and exploiting the raw material of the 21st century

MM 66.3: Vortrag

Donnerstag, 15. März 2018, 18:15–18:30, H 0107

Generation of ab initio datasets with predefined precision using uncertainty quantification — •Jan Janßen, Tilmann Hickel, and Jörg Neugebauer — Max-Planck-Institut für Eisenforschung GmbH, Düsseldorf, Germany

A major challenge in multiscale materials simulation is the ab initio prediction of phase stabilities in multi-phase materials. To extend the ab initio accuracy to larger length and time scales the fitting of machine learning potentials seems promising, but this approach is intrinsically limited to the accuracy of the input data. Therefore it is essential to quantify the different sources of uncertainty in ab initio calculation, including the systematical error of convergence, the statistical or numerical error and the model error for derived quantities. Already the determination of the equilibrium lattice constant and bulk modulus requires a careful analysis of the fitting of energy-volume curves, going beyond the consideration of standard convergence parameters like cutoff and k-points. In order to handle this delicate interplay of uncertainties, we introduce the concept of uncertainty phase diagrams. Based on the uncertainty phase diagrams we model the convergence gradients of the contributing errors, to automate the convergence process not only for the error in energy. The modelling of uncertainties in relation to the corresponding ab initio calculation is enabled by our recently developed Python based workbench pyiron. Our investigations revealed that commonly used rules of thumb for fitting ground state materials properties become invalid for high precision calculations.