Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
MM: Fachverband Metall- und Materialphysik
MM 10: Methods in Computational Materials Modelling (methodological aspects, numerics)
MM 10.5: Vortrag
Montag, 1. April 2019, 16:45–17:00, H45
Learning to use the force: DFTB repulsion with Gaussian Process Regression — Artur Engelmann, •Chiara Panosetti, Johannes T. Margraf, and Karsten Reuter — Chair for Theoretical Chemistry, Technische Universität München, Germany
Density-Functional Tight Binding (DFTB) is increasingly popular among computational modellists as it provides comparable accuracy to DFT at a fraction of the cost, enabling large scale simulations while retaining direct access to electronic structure properties. Yet, a bottleneck to this day remains the difficulty to parametrize the interactions for large subsets of atoms across the periodic table, let alone an extensive, universal parametrization. Especially challenging is the parametrization of the pairwise repulsion: an unescapable N2 effort, and a cumbersome one. Most schemes involve fitting the repulsion to some analytical potential by e.g. minimizing force residues. However, this presents a number of limitations, such as constraints on the reference geometries, and the necessity of projecting forces along the bonds. Further, any predefined functional form not only carries a certain degree of arbitrary bias in its very choice, but may also lack the flexibility to capture subtle features around equilibrium distances (where the “repulsive” potential may as well be attractive). We thereby propose to rather machine-learn the repulsive force, using Gaussian Process Regression similarly to the generation of GAP potentials [1]. We discuss a proof-of-principle application on carbon, showing how such an approach removes all the above limitations at once.
[1] A.P. Bartók et al., PRL 104, 136403 (2010)