Regensburg 2022 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
O: Fachverband Oberflächenphysik
O 58: Poster Wednesday: New Methods and Developments, Frontiers of Electronic Structure Theory
O 58.6: Poster
Wednesday, September 7, 2022, 18:00–20:00, P4
Fortnet, a software package for training Behler-Parrinello neural networks — •Tammo van der Heide1, Jolla Kullgren2, Peter Broqvist2, Vladimir Bačić3, Thomas Frauenheim4,5,1, and Bálint Aradi1 — 1BCCMS, University of Bremen, Bremen, Germany — 2Dept. of Chemistry - Ångström Laboratory, Uppsala University, Uppsala, Sweden — 3Dept. of Physics and Earth Sciences, Jacobs University Bremen, Bremen, Germany — 4Beijing CSRC, 100193 Beijing, P. R. China — 5Shenzhen JL CSAR Institute, Shenzhen 518110, P. R. China
A new, open source, parallel, stand-alone software package (Fortnet) has been developed, which implements Behler-Parrinello neural networks. It covers the entire workflow from feature generation to the evaluation of generated potentials, coupled with higher-level analysis such as the analytic calculation of atomic forces. The functionality is demonstrated by driving the training for the fitted correction functions of the density functional tight binding (DFTB) method, which are commonly used to compensate the inaccuracies resulting from the DFTB approximations to the Kohn-Sham Hamiltonian. Their usual two-body form limits the transferability of parameterizations between very different structural environments. After investigating various approaches, we have found the combination of DFTB with a near-sighted artificial neural network, acting on-top of baseline correction functions, the most promising one. It allows to introduce many-body corrections on top of two-body parametrizations, while excellent transferability to deviating chemical environments could be demonstrated.