Dr. Mariana Rossi
Center for Free-Electron Laser Science, Hamburg
will speak on the topic
Machine Learning for Electronic Structure Properties
Abstract:
A comprehensive ML framework at the nanoscale needs to target both nuclear and electronic subsystems. The electronic density is a fundamental quantity that allows accessing several material properties, in particular within the framework of density-functional theory (DFT). However, obtaining a ground-state electronic density of large materials is prohibitive due to the cost of converging the self-consistent procedure and simulating nuclear motion with quantum-mechanical accuracy adds an even larger overhead.
In this talk, we will discuss the results of recent performance and stability improvements to the symmetry-adapted learning of three-dimensional electron densities (SALTED) method, which allows one to directly predict the ground-state electronic density based only on atomic positions and composition. These improvements allow us to obtain accurate predictions of the density and derived properties (electronic states, band gaps, etc.) of a wide variety of systems. As a highlight, we will show results of band-structure prediction for a diverse set of twisted bilayer systems, when training on small unit cells and predicting on large unit cells with a low twisting angle. Finally, we couple SALTED predictions of twisted-bilayer systems to thermally displaced structures obtained from machine learning interatomic potentials (MLIPs) trained on smaller-scale DFT data.
Interested parties are cordially invited
Professor Dr. Kühn