Predicting polarizabilities of silicon clusters using local chemical environments

TL;DR

This paper develops machine learning models to efficiently predict polarizabilities of large silicon clusters, overcoming computational challenges of first-principles methods, and demonstrates accurate predictions for clusters too large for direct calculations.

Contribution

The study introduces a machine learning approach using SOAP descriptors to predict polarizabilities of large silicon clusters, enabling efficient and accurate estimations beyond first-principles capabilities.

Findings

ML models accurately fit RPA polarizability data

Models predict polarizabilities for large clusters beyond computational limits

Predicted bulk limit agrees with previous studies

Abstract

Calculating polarizabilities of large clusters with first-principles techniques is challenging because of the unfavorable scaling of computational cost with cluster size. To address this challenge, we demonstrate that polarizabilities of large hydrogenated silicon clusters containing thousands of atoms can be efficiently calculated with machine learning methods. Specifically, we construct machine learning models based on the smooth overlap of atomic positions (SOAP) descriptor and train the models using a database of calculated random-phase approximation polarizabilities for clusters containing up to 110 silicon atoms. We first demonstrate the ability of the machine learning models to fit the data and then assess their ability to predict cluster polarizabilities using k-fold cross validation. Finally, we study the machine learning predictions for clusters that are too large for explicit first-principles calculations and find that they accurately describe the dependence of the polarizabilities on the ratio of hydrogen to silicon atoms and also predict a bulk limit that is in good agreement with previous studies.