Regensburg 2022 – wissenschaftliches Programm

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

O: Fachverband Oberflächenphysik

O 67: Frontiers of Electronic Structure Theory: Focus on Artificial Intelligence Applied to Real Materials 3

O 67.5: Vortrag

Donnerstag, 8. September 2022, 11:30–11:45, S054

Predicting condensed-phase electron densities using machine learning — •Alan Lewis¹, Andrea Grisafi², Michele Ceriotti², and Mariana Rossi¹ — ¹MPI for Structure and Dynamics of Materials, Hamburg, Germany — ²École Polytéchnique Fédèrale de Lausanne, Lausanne, Switzerland

The electron density is a fundamental quantity for understanding physical phenomena in materials, and is central to electronic structure theories such as density-functional theory. We present the SALTED machine learning method and demonstrate its ability to learn and predict the electronic densities of a range of materials from simple liquids and metals to hybrid organic-inorganic perovskites. This extends the framework presented in ACS Cent. Sci. 5, 57, 2019 to work with periodic boundary conditions and uses a resolution of the identity on a numeric atom-centered orbital basis to expand the all-electron periodic density. A Gaussian process regression model that makes use of local symmetry-adapted representations of the atomic structure is employed, making our method both data-efficient and highly transferable.[1] We also compare various methods of dealing with the non-orthogonality of the basis, accounting for correlations between pairs of off-centered density components, finding that the best compromise between accuracy and computational efficiency comes from approximating the density expansion coefficients by directly minimizing the loss function. The total energies derived from the densities obtained in this way present errors with respect to DFT of just 0.1 meV/atom.

[1] Lewis, Grisafi, Ceriotti, Rossi, JCTC 17, 11, 7203 (2021)