# ABSTRACT

# Introduction

G. Lucas and L. Pizzagalli pointed out
  that the use of available empirical potentials
  is the largest source of errors to calculate threshold displacement energies in 3C-SiC
  and called for the improvement of existing potentials.

# Method

## Training process

VASP KSPACING param

Spin-polarized calculations were considered
  to account for the possible spin polarization of various defect configurations.

DPGEN

## Interpolation for short range repulsion

use ZBL to fit the short range repulsion

# Results

## Near-equilibrium properties

Equation of state: 指的是体系能量随晶格常数的变化关系

use inelastic x-ray scattering (IXS) to measure the phonon dispersion curves

## Threshold displacement energies

## Defect production caused by a single PKA

The Wigner-Seitz cell analysis method was used to determine the defects number.

# Discussions

DP-ZBL is also about 100 times slower than the empirical potential.

Therefore, how to efficiently construct a complete training dataset for the target problem
  is the critical problem that the deep learning potentials
  and even all the machine learning potentials need to answer.

# Conclusions