Parts | Days | Selection | Search | Updates | Downloads | Help
O: Fachverband Oberflächenphysik
O 82: Electronic Structure Theory I
O 82.4: Talk
Thursday, March 21, 2024, 11:15–11:30, MA 043
Conventional definitions of the absolute energy reference can lead to suboptimal machine learning performance for the electronic density of states — •Wei Bin How, Sanggyu Chong, Federico Grasselli, Kevin Kazuki Huguenin-Dumittan, and Michele Ceriotti — Ecole Polytechnique Fédérale de Lausanne, Lausanne, Switzerland
The electronic density of states (eDOS) provides a simple and clear picture of the distribution of energy states for electrons, granting key insights regarding the behaviour of electrons in the material. Currently, the most popular way to obtain the eDOS is through DFT calculations, but the cubic scaling behaviour of DFT has motivated a push to apply machine learning to obtain the eDOS at a much lower cost. However, when constructing the dataset for the machine learning problem, one has to choose an energy reference for the dataset. Current conventions define the energy reference of the eDOS of bulk systems using either the average Hartree potential in the cell or the Fermi level of the entire system. As the absolute value of these quantities are typically not well defined for bulk systems, eDOSes of different bulks have to be interpreted individually. However, machine learning methods typically treat these energy references as absolute which may hinder prediction performance. In this talk, we explore two different ways to provide an optimal absolute energy reference for machine learning and showcase the significant improvement in performance over conventional definitions.
Keywords: Machine Learning; Electronic Density of States; Alignment; DOS; Optimize