The road to accuracy: machine-learning-accelerated silicon ab initio simulations

TL;DR

This paper discusses hybrid approaches combining ab initio methods and machine learning to significantly accelerate silicon simulations while maintaining quantum-level accuracy, demonstrated through defect searches and deep learning potentials.

Contribution

It introduces new machine learning models that enhance the speed of silicon ab initio simulations without sacrificing accuracy, enabling more practical nanoscale material studies.

Findings

Successful identification of low energy defect complexes in silicon.

Development of deep learning potentials for molecular dynamics.

Reproduction of silicon's thermodynamic properties with high fidelity.

Abstract

Ab initio simulations are capable of providing detailed information of material behavior at the nanoscale. Simulating experimentally relevant situations is, however, often computationally intense. Using hybrid approaches between ab initio methods such as density functional theory (DFT) and machine learning, new models can be constructed which retain quantum accuracy while being computationally faster by several orders of magnitude. Two examples are discussed in this paper. The first is the computational search for low energy substitutional defect complexes in silicon. The second is the construction of deep learning potentials for ab initio-level molecular dynamics simulations. The latter is applied to reproduce the 0 K equation of state and high temperature thermal expansion of Si using the same model.