Phonon dispersion filter: A physics-inspired feature selection for machine learning potentials

Abstract

How to improve the accuracy and precision of machine learning potential functions while reducing their computational cost has long been a subject of considerable interest. In this regard, a common approach is to reduce the number of descriptors through feature selection and dimensionality reduction, thereby improving computational efficiency. In our paper, we propose a descriptor selection method based on the material’s phonon spectrum, which is called a phonon dispersion filter (PDF) method. Compared to other mathematics-based machine learning feature selection methods, the PDF method is a more physics-based feature selection approach. Taking graphene and bulk silicon as examples, we provide a detailed introduction to the screening process of the PDF method and its underlying principles. Furthermore, we test the PDF method on two types of descriptors: Atom-centered symmetry functions descriptors and smooth overlap of atomic positions descriptors. Both demonstrate promising screening results.