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 58.8: Poster
Wednesday, September 7, 2022, 18:00–20:00, P4
Electronic properties of Density Functional Tight Binding by Machine Learning — •Guozheng Fan1, Adam McSloy1, Balint Aradi1, Chi-Yung Yam2, and Thomas Frauenheim1, 2 — 1Bremen Center for Computational Materials Science (BCCMS), University of Bremen, Bremen, Germany — 2Beijing Computational Science Research Center (CSRC), Beijing, China
We have introduced a machine learning workflow, which could optimize electronic properties in density functional tight binding method. With this workflow, we can train and predict electronic properties in a cheap, accurate and transferable way. The implementation features of batch calculations greatly improve the calculation efficiency, especially for high throughput calculations. This workflow could optimize electronic properties by train basis functions or train a spline model to generate two center integrals for off-diagonal and onsite for diagonal Hamiltonian and overlap. The results show that compared with previous Slater-Koster parameters, the dipole moments, charges, the ratios of the on-site populations and the atomic numbers in charge population analysis method can be improved by both tuning basis function parameters or optimizing integrals in spline model directly. The training on basis functions could prevent the two center integrals go randomly and keep Hamiltonian and overlap in reasonable range. Besides, the multiple electronic properties could be improved simultaneously.