Dresden 2020 – wissenschaftliches Programm

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

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

CPP: Fachverband Chemische Physik und Polymerphysik

CPP 99: Frontiers in Electronic-Structure Theory - Focus on Electron-Phonon Interactions V (joint session O/CPP/DS/HL)

CPP 99.4: Vortrag

Donnerstag, 19. März 2020, 16:00–16:15, GER 38

Error Estimation of Energy-per-Atom of Semiconductor Compounds Using Statistical Learning — •Daniel T. Speckhard1,2, Sven Lubeck2, Christian Carbogno1, Luca Ghiringhelli1, Claudia Draxl1,2, and Matthias Scheffler11Fritz-Haber-Institut der Max-Planck-Gesellschaft, Berlin, Germany — 2Humboldt-Universität zu Berlin, Institut für Physik and IRIS Adlershof, Berlin, Germany

Material databases such as NOMAD give researchers the ability to work with millions of material simulation results [1]. However, it is typically unclear to which extent calculations performed with different numerical settings and computer codes can be trusted and related to each other. This project presents statistical learning strategies to model errors in energies for two all-electron DFT codes, FHI-aims and exciting, for different basis-set sizes and k-point densities. Specifically, we use mutual information scores to select features that are able to capture the energy-per-atom errors. With respect to several metrics, random forest regression on the selected features shows the most promising results. This work lays the foundation for estimating errors in DFT data in NOMAD and helps to save computing resources by a priori predicting the DFT simulation settings required to achieve a desired level of precision. This also enables us to estimate the basis-set and k-point converged results of not fully converged calculations.

[1] C. Draxl and M. Scheffler, J. Phys. Mat., 2 036001 (2019). https://nomad-coe.eu

100% | Mobil-Ansicht | English Version | Kontakt/Impressum/Datenschutz
DPG-Physik > DPG-Verhandlungen > 2020 > Dresden