Parts | Days | Selection | Search | Updates | Downloads | Help

O: Fachverband Oberflächenphysik

O 46: New Methods and Developments 3: Theory

O 46.4: Talk

Wednesday, September 7, 2022, 15:45–16:00, H6

A Revised Fourth-Generation Neural Network Potential for the Accurate Representation of Multiple Charge States — •Alexander Knoll, Tsz Wai Ko, and Jörg Behler — Georg-August-Universität Göttingen, Institut für Physikalische Chemie, Theoretische Chemie, Göttingen, Germany

Machine learning potentials (MLPs) have become a mature tool for large-scale atomistic simulations in chemistry and materials science. Recently, a fourth-generation high-dimensional neural network potential (4G-HDNNP) has been introduced, in which the atomic charges are determined in a charge equilibration step enabling the description of long-range charge transfer. The quality of the charge prediction depends on environment-dependent electronegativities expressed by atomic neural networks, which poses a challenge for structures with differing total charges but nearly-identical nuclear positions. Here, we propose a generalized method applicable to these situations, and for a series of model systems we demonstrate that this extension leads to additional flexibility of the atomic electronegativities, ultimately resulting in more accurate atomic charges, energies, and forces.

100% | Screen Layout | Deutsche Version | Contact/Imprint/Privacy
DPG-Physik > DPG-Verhandlungen > 2022 > Regensburg