Electronic structures of crystalline and amorphous GeSe and GeSbTe compounds using machine learning empirical pseudopotentials

Abstract

The newly developed machine learning (ML) empirical pseudopotential (EP) method overcomes the poor transferability of the traditional EP method with the help of ML techniques while preserving its formal simplicity and computational efficiency. We apply the new method to binary and ternary systems such as GeSe and Ge-Sb-Te (GST) compounds, well-known materials for non-volatile phase-change memory and related technologies. Using a training set of {\it ab initio} electronic energy bands and rotation-covariant descriptors for various GeSe and GST compounds, we generate transferable EPs for Ge, Se, Sb, and Te. We demonstrate that the new ML model accurately reproduces the energy bands and wavefunctions of structures outside the training set, closely matching first-principles calculations. This accuracy is achieved with significantly lower computational costs due to the elimination of self-consistency iterations and the reduced size of the plane-wave basis set. Notably, the method maintains accuracy even for diverse local atomic environments, such as amorphous phases or larger systems not explicitly included in the training set.