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

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

MM: Fachverband Metall- und Materialphysik

MM 59: Computational Materials Modelling - Potentials

MM 59.2: Vortrag

Donnerstag, 19. März 2020, 16:00–16:15, IFW A

Machine-learned density-functional tight-binding: Enabling high-quality electronic structure calculations on systems too large for DFT — •Simon Anniés, Chiara Panosetti, Christoph Scheurer, and Karsten Reuter — Technical University Munich

Density-Functional Theory (DFT) has been one of the go-to methods for electronic structure calculations for several decades. Nevertheless, inherent scaling properties and/or large prefactors for linear-scaling implementations still limit its usability for larger systems or extensive sampling.

Density-Functional Tight-Binding (DFTB), a semi-empirical approximation to DFT, is an alternative that - as opposed to forcefields - retains access to electronic structure properties while providing a speed-up of roughly three orders of magnitude. The trade-off is a two-part parametrization process. The numerical parameters of the electronic part are optimized by comparing with DFT bandstructures, the repulsion potential by forcematching against representative training sets from high-level electronic structure methods.

In our work, we apply a machine learning approach, making use of Gaussian Process Regression (GPR), in order to greatly increase the adaptability of the repulsion potential. By embracing a purely data-driven methodology, this overcomes the limitations of previously employed rigid functional forms. We demonstrate the superior transferability of this approach in the application to lithium-intercalated graphite as prevalent anode material of commercial lithium-ion batteries.