Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Melting line of silicon modelled with a machine-learning potential

Yu. D. Fomin

TL;DR

The study assesses silicon's phase diagram using the SNAP machine-learning potential, employing a two-phase method to determine the melting line and an evolutionary crystal-structure search to map high-pressure phases. SNAP reproduces the qualitatively correct negative slope of the diamond-melting line but significantly underestimates melting temperatures and fails to capture high-pressure phases observed experimentally, with GAP performing somewhat better but still not in quantitative agreement. Comparisons to empirical models show SW matches experiment best among the tested options, while current ML potentials for silicon remain limited in accuracy for liquid-phase predictions. The authors advocate developing more accurate ML potentials for liquid silicon, ideally informed by DFT, to reliably describe both solid and liquid phases.

Abstract

In the present study we investigate the phase diagram of silicon within the framework of SNAP machine learning potential model. We show that the melting line of diamond phase of silicon is a linear function of pressure, which is in good agreement with experimental data. At the same time the melting temperature is strongly underestimated. Also, this model fails to predict the high pressure phases of silicon.

Melting line of silicon modelled with a machine-learning potential

TL;DR

The study assesses silicon's phase diagram using the SNAP machine-learning potential, employing a two-phase method to determine the melting line and an evolutionary crystal-structure search to map high-pressure phases. SNAP reproduces the qualitatively correct negative slope of the diamond-melting line but significantly underestimates melting temperatures and fails to capture high-pressure phases observed experimentally, with GAP performing somewhat better but still not in quantitative agreement. Comparisons to empirical models show SW matches experiment best among the tested options, while current ML potentials for silicon remain limited in accuracy for liquid-phase predictions. The authors advocate developing more accurate ML potentials for liquid silicon, ideally informed by DFT, to reliably describe both solid and liquid phases.

Abstract

In the present study we investigate the phase diagram of silicon within the framework of SNAP machine learning potential model. We show that the melting line of diamond phase of silicon is a linear function of pressure, which is in good agreement with experimental data. At the same time the melting temperature is strongly underestimated. Also, this model fails to predict the high pressure phases of silicon.

Paper Structure

This paper contains 6 sections, 2 figures.

Table of Contents

  1. Introduction
  2. System and Methods
  3. Results and discussion
  4. Melting line
  5. Crystal structure calculation
  6. Conclusions

Figures (2)

  • Figure S1: (a) Melting line of SNAP model of silicon. The symbols show raw data. The line is the linear fit of these data. (b) A comparison of the data for SNAP model with the literature ones. The curve 'exp' means the experimental curve from Ref. kubo. The symbols 'D SW', 'D EA', 'D KIMS' and 'D ZBL' refer to the results for SW, EA, KIMS and ZBL models reported by Dozhdikov and coauthors in Ref. dozhdikov. In the case of SNAP model we give only the linear fit of the data.
  • Figure S2: Relative difference between the energy of the diamond structure and $\beta$-Sn one for the SNAPP model of silicon as a function of pressure.